What are the characteristics of anomaly detection?  What are the detection problems and methods? W


  1. What are the characteristics of anomaly detection? 
  2. What are the detection problems and methods?
  3. What are the statistical approaches when there is an anomaly found?
  4. Compare and contrast proximity and clustering based approaches.


Don't use plagiarized sources. Get Your Custom Essay on
What are the characteristics of anomaly detection?  What are the detection problems and methods? W
Just from $10/Page
Order Essay


Michigan State Universit
University of Minnesota
University of Minnesota
University of Minnesota

330 Hudson Street, NY NY 10013

Director, Portfolio Management: Engineering, Computer Science & Global
Editions: Julian Partridge

Specialist, Higher Ed Portfolio Management: Matt Goldstein

Portfolio Management Assistant: Meghan Jacoby

Managing Content Producer: Scott Disanno

Content Producer: Carole Snyder

Web Developer: Steve Wright

Rights and Permissions Manager: Ben Ferrini

Manufacturing Buyer, Higher Ed, Lake Side Communications Inc (LSC):
Maura Zaldivar-Garcia

Inventory Manager: Ann Lam

Product Marketing Manager: Yvonne Vannatta

Field Marketing Manager: Demetrius Hall

Marketing Assistant: Jon Bryant

Cover Designer: Joyce Wells, jWellsDesign

Full-Service Project Management: Chandrasekar Subramanian, SPi Global

Copyright ©2019 Pearson Education, Inc. All rights reserved. Manufactured in
the United States of America. This publication is protected by Copyright, and
permission should be obtained from the publisher prior to any prohibited
reproduction, storage in a retrieval system, or transmission in any form or by
any means, electronic, mechanical, photocopying, recording, or likewise. For
information regarding permissions, request forms and the appropriate
contacts within the Pearson Education Global Rights & Permissions
department, please visit www.pearsonhighed.com/permissions/.

Many of the designations by manufacturers and sellers to distinguish their
products are claimed as trademarks. Where those designations appear in this
book, and the publisher was aware of a trademark claim, the designations
have been printed in initial caps or all caps.

Library of Congress Cataloging-in-Publication Data on File

Names: Tan, Pang-Ning, author. | Steinbach, Michael, author. | Karpatne,
Anuj, author. | Kumar, Vipin, 1956- author.

Title: Introduction to Data Mining / Pang-Ning Tan, Michigan State University,
Michael Steinbach, University of Minnesota, Anuj Karpatne, University of
Minnesota, Vipin Kumar, University of Minnesota.

Description: Second edition. | New York, NY : Pearson Education, [2019] |
Includes bibliographical references and index.

Identifiers: LCCN 2017048641 | ISBN 9780133128901 | ISBN 0133128903

Subjects: LCSH: Data mining.

Classification: LCC QA76.9.D343 T35 2019 | DDC 006.3/12–dc23 LC record
available at https://lccn.loc.gov/2017048641

1 18

ISBN-10: 0133128903

ISBN-13: 9780133128901

To our families …

Preface to the Second Edition
Since the first edition, roughly 12 years ago, much has changed in the field of
data analysis. The volume and variety of data being collected continues to
increase, as has the rate (velocity) at which it is being collected and used to
make decisions. Indeed, the term, Big Data, has been used to refer to the
massive and diverse data sets now available. In addition, the term data
science has been coined to describe an emerging area that applies tools and
techniques from various fields, such as data mining, machine learning,
statistics, and many others, to extract actionable insights from data, often big

The growth in data has created numerous opportunities for all areas of data
analysis. The most dramatic developments have been in the area of predictive
modeling, across a wide range of application domains. For instance, recent
advances in neural networks, known as deep learning, have shown
impressive results in a number of challenging areas, such as image
classification, speech recognition, as well as text categorization and
understanding. While not as dramatic, other areas, e.g., clustering,
association analysis, and anomaly detection have also continued to advance.
This new edition is in response to those advances.


As with the first edition, the second edition of the book provides a
comprehensive introduction to data mining and is designed to be accessible
and useful to students, instructors, researchers, and professionals. Areas

covered include data preprocessing, predictive modeling, association
analysis, cluster analysis, anomaly detection, and avoiding false discoveries.
The goal is to present fundamental concepts and algorithms for each topic,
thus providing the reader with the necessary background for the application of
data mining to real problems. As before, classification, association analysis
and cluster analysis, are each covered in a pair of chapters. The introductory
chapter covers basic concepts, representative algorithms, and evaluation
techniques, while the more following chapter discusses advanced concepts
and algorithms. As before, our objective is to provide the reader with a sound
understanding of the foundations of data mining, while still covering many
important advanced topics. Because of this approach, the book is useful both
as a learning tool and as a reference.

To help readers better understand the concepts that have been presented, we
provide an extensive set of examples, figures, and exercises. The solutions to
the original exercises, which are already circulating on the web, will be made
public. The exercises are mostly unchanged from the last edition, with the
exception of new exercises in the chapter on avoiding false discoveries. New
exercises for the other chapters and their solutions will be available to
instructors via the web. Bibliographic notes are included at the end of each
chapter for readers who are interested in more advanced topics, historically
important papers, and recent trends. These have also been significantly
updated. The book also contains a comprehensive subject and author index.

What is New in the Second Edition?

Some of the most significant improvements in the text have been in the two
chapters on classification. The introductory chapter uses the decision tree
classifier for illustration, but the discussion on many topics—those that apply

across all classification approaches—has been greatly expanded and
clarified, including topics such as overfitting, underfitting, the impact of training
size, model complexity, model selection, and common pitfalls in model
evaluation. Almost every section of the advanced classification chapter has
been significantly updated. The material on Bayesian networks, support vector
machines, and artificial neural networks has been significantly expanded. We
have added a separate section on deep networks to address the current
developments in this area. The discussion of evaluation, which occurs in the
section on imbalanced classes, has also been updated and improved.

The changes in association analysis are more localized. We have completely
reworked the section on the evaluation of association patterns (introductory
chapter), as well as the sections on sequence and graph mining (advanced
chapter). Changes to cluster analysis are also localized. The introductory
chapter added the K-means initialization technique and an updated the
discussion of cluster evaluation. The advanced clustering chapter adds a new
section on spectral graph clustering. Anomaly detection has been greatly
revised and expanded. Existing approaches—statistical, nearest
neighbor/density-based, and clustering based—have been retained and
updated, while new approaches have been added: reconstruction-based, one-
class classification, and information-theoretic. The reconstruction-based
approach is illustrated using autoencoder networks that are part of the deep
learning paradigm. The data chapter has been updated to include discussions
of mutual information and kernel-based techniques.

The last chapter, which discusses how to avoid false discoveries and produce
valid results, is completely new, and is novel among other contemporary
textbooks on data mining. It supplements the discussions in the other
chapters with a discussion of the statistical concepts (statistical significance,
p-values, false discovery rate, permutation testing, etc.) relevant to avoiding
spurious results, and then illustrates these concepts in the context of data

mining techniques. This chapter addresses the increasing concern over the
validity and reproducibility of results obtained from data analysis. The addition
of this last chapter is a recognition of the importance of this topic and an
acknowledgment that a deeper understanding of this area is needed for those
analyzing data.

The data exploration chapter has been deleted, as have the appendices, from
the print edition of the book, but will remain available on the web. A new
appendix provides a brief discussion of scalability in the context of big data.

To the Instructor

As a textbook, this book is suitable for a wide range of students at the
advanced undergraduate or graduate level. Since students come to this
subject with diverse backgrounds that may not include extensive knowledge of
statistics or databases, our book requires minimal prerequisites. No database
knowledge is needed, and we assume only a modest background in statistics
or mathematics, although such a background will make for easier going in
some sections. As before, the book, and more specifically, the chapters
covering major data mining topics, are designed to be as self-contained as
possible. Thus, the order in which topics can be covered is quite flexible. The
core material is covered in chapters 2 (data), 3 (classification), 5 (association
analysis), 7 (clustering), and 9 (anomaly detection). We recommend at least a
cursory coverage of Chapter 10 (Avoiding False Discoveries) to instill in
students some caution when interpreting the results of their data analysis.
Although the introductory data chapter (2) should be covered first, the basic
classification (3), association analysis (5), and clustering chapters (7), can be
covered in any order. Because of the relationship of anomaly detection (9) to
classification (3) and clustering (7), these chapters should precede Chapter 9.

Various topics can be selected from the advanced classification, association
analysis, and clustering chapters (4, 6, and 8, respectively) to fit the schedule
and interests of the instructor and students. We also advise that the lectures
be augmented by projects or practical exercises in data mining. Although they
are time consuming, such hands-on assignments greatly enhance the value of
the course.

Support Materials

Support materials available to all readers of this book are available at

PowerPoint lecture slides
Suggestions for student projects
Data mining resources, such as algorithms and data sets
Online tutorials that give step-by-step examples for selected data mining
techniques described in the book using actual data sets and data analysis

Additional support materials, including solutions to exercises, are available
only to instructors adopting this textbook for classroom use. The book’s
resources will be mirrored at www.pearsonhighered.com/cs-resources.
Comments and suggestions, as well as reports of errors, can be sent to the
authors through [email protected]


Many people contributed to the first and second editions of the book. We
begin by acknowledging our families to whom this book is dedicated. Without
their patience and support, this project would have been impossible.

We would like to thank the current and former students of our data mining
groups at the University of Minnesota and Michigan State for their
contributions. Eui-Hong (Sam) Han and Mahesh Joshi helped with the initial
data mining classes. Some of the exercises and presentation slides that they
created can be found in the book and its accompanying slides. Students in our
data mining groups who provided comments on drafts of the book or who
contributed in other ways include Shyam Boriah, Haibin Cheng, Varun
Chandola, Eric Eilertson, Levent Ertöz, Jing Gao, Rohit Gupta, Sridhar Iyer,
Jung-Eun Lee, Benjamin Mayer, Aysel Ozgur, Uygar Oztekin, Gaurav Pandey,
Kashif Riaz, Jerry Scripps, Gyorgy Simon, Hui Xiong, Jieping Ye, and
Pusheng Zhang. We would also like to thank the students of our data mining
classes at the University of Minnesota and Michigan State University who
worked with early drafts of the book and provided invaluable feedback. We
specifically note the helpful suggestions of Bernardo Craemer, Arifin Ruslim,
Jamshid Vayghan, and Yu Wei.

Joydeep Ghosh (University of Texas) and Sanjay Ranka (University of Florida)
class tested early versions of the book. We also received many useful
suggestions directly from the following UT students: Pankaj Adhikari, Rajiv
Bhatia, Frederic Bosche, Arindam Chakraborty, Meghana Deodhar, Chris
Everson, David Gardner, Saad Godil, Todd Hay, Clint Jones, Ajay Joshi,
Joonsoo Lee, Yue Luo, Anuj Nanavati, Tyler Olsen, Sunyoung Park, Aashish
Phansalkar, Geoff Prewett, Michael Ryoo, Daryl Shannon, and Mei Yang.

Ronald Kostoff (ONR) read an early version of the clustering chapter and
offered numerous suggestions. George Karypis provided invaluable LATEX
assistance in creating an author index. Irene Moulitsas also provided

assistance with LATEX and reviewed some of the appendices. Musetta
Steinbach was very helpful in finding errors in the figures.

We would like to acknowledge our colleagues at the University of Minnesota
and Michigan State who have helped create a positive environment for data
mining research. They include Arindam Banerjee, Dan Boley, Joyce Chai, Anil
Jain, Ravi Janardan, Rong Jin, George Karypis, Claudia Neuhauser, Haesun
Park, William F. Punch, György Simon, Shashi Shekhar, and Jaideep
Srivastava. The collaborators on our many data mining projects, who also
have our gratitude, include Ramesh Agrawal, Maneesh Bhargava, Steve
Cannon, Alok Choudhary, Imme Ebert-Uphoff, Auroop Ganguly, Piet C. de
Groen, Fran Hill, Yongdae Kim, Steve Klooster, Kerry Long, Nihar Mahapatra,
Rama Nemani, Nikunj Oza, Chris Potter, Lisiane Pruinelli, Nagiza Samatova,
Jonathan Shapiro, Kevin Silverstein, Brian Van Ness, Bonnie Westra, Nevin
Young, and Zhi-Li Zhang.

The departments of Computer Science and Engineering at the University of
Minnesota and Michigan State University provided computing resources and a
supportive environment for this project. ARDA, ARL, ARO, DOE, NASA,
NOAA, and NSF provided research support for Pang-Ning Tan, Michael Stein-
bach, Anuj Karpatne, and Vipin Kumar. In particular, Kamal Abdali, Mitra
Basu, Dick Brackney, Jagdish Chandra, Joe Coughlan, Michael Coyle,
Stephen Davis, Frederica Darema, Richard Hirsch, Chandrika Kamath,
Tsengdar Lee, Raju Namburu, N. Radhakrishnan, James Sidoran, Sylvia
Spengler, Bhavani Thuraisingham, Walt Tiernin, Maria Zemankova, Aidong
Zhang, and Xiaodong Zhang have been supportive of our research in data
mining and high-performance computing.

It was a pleasure working with the helpful staff at Pearson Education. In
particular, we would like to thank Matt Goldstein, Kathy Smith, Carole Snyder,

and Joyce Wells. We would also like to thank George Nichols, who helped
with the art work and Paul Anagnostopoulos, who provided LATEX support.

We are grateful to the following Pearson reviewers: Leman Akoglu (Carnegie
Mellon University), Chien-Chung Chan (University of Akron), Zhengxin Chen
(University of Nebraska at Omaha), Chris Clifton (Purdue University), Joy-
deep Ghosh (University of Texas, Austin), Nazli Goharian (Illinois Institute of
Technology), J. Michael Hardin (University of Alabama), Jingrui He (Arizona
State University), James Hearne (Western Washington University), Hillol
Kargupta (University of Maryland, Baltimore County and Agnik, LLC), Eamonn
Keogh (University of California-Riverside), Bing Liu (University of Illinois at
Chicago), Mariofanna Milanova (University of Arkansas at Little Rock),
Srinivasan Parthasarathy (Ohio State University), Zbigniew W. Ras (University
of North Carolina at Charlotte), Xintao Wu (University of North Carolina at
Charlotte), and Mohammed J. Zaki (Rensselaer Polytechnic Institute).

Over the years since the first edition, we have also received numerous
comments from readers and students who have pointed out typos and various
other issues. We are unable to mention these individuals by name, but their
input is much appreciated and has been taken into account for the second

Preface to the Second Edition v

1 Introduction 1
1.1 What Is Data Mining? 4

1.2 Motivating Challenges 5

1.3 The Origins of Data Mining 7

1.4 Data Mining Tasks 9

1.5 Scope and Organization of the Book 13

1.6 Bibliographic Notes 15

1.7 Exercises 21

2 Data 23
2.1 Types of Data 26

2.1.1 Attributes and Measurement 27

2.1.2 Types of Data Sets 34

2.2 Data Quality 42
2.2.1 Measurement and Data Collection Issues 42

2.2.2 Issues Related to Applications 49

2.3 Data Preprocessing 50
2.3.1 Aggregation 51

2.3.2 Sampling 52

2.3.3 Dimensionality Reduction 56

2.3.4 Feature Subset Selection 58

2.3.5 Feature Creation 61

2.3.6 Discretization and Binarization 63

2.3.7 Variable Transformation 69

2.4 Measures of Similarity and Dissimilarity 71
2.4.1 Basics 72

2.4.2 Similarity and Dissimilarity between Simple Attributes 74

2.4.3 Dissimilarities between Data Objects 76

2.4.4 Similarities between Data Objects 78

2.4.5 Examples of Proximity Measures 79

2.4.6 Mutual Information 88

2.4.7 Kernel Functions* 90

2.4.8 Bregman Divergence* 94

2.4.9 Issues in Proximity Calculation 96

2.4.10 Selecting the Right Proximity Measure 98

2.5 Bibliographic Notes 100

2.6 Exercises 105

3 Classification: Basic Concepts and Techniques 113

3.1 Basic Concepts 114

3.2 General Framework for Classification 117

3.3 Decision Tree Classifier 119
3.3.1 A Basic Algorithm to Build a Decision Tree 121

3.3.2 Methods for Expressing Attribute Test Conditions 124

3.3.3 Measures for Selecting an Attribute Test Condition 127

3.3.4 Algorithm for Decision Tree Induction 136

3.3.5 Example Application: Web Robot Detection 138

3.3.6 Characteristics of Decision Tree Classifiers 140

3.4 Model Overfitting 147
3.4.1 Reasons for Model Overfitting 149

3.5 Model Selection 156
3.5.1 Using a Validation Set 156

3.5.2 Incorporating Model Complexity 157

3.5.3 Estimating Statistical Bounds 162

3.5.4 Model Selection for Decision Trees 162

3.6 Model Evaluation 164
3.6.1 Holdout Method 165

3.6.2 Cross-Validation 165

3.7 Presence of Hyper-parameters 168
3.7.1 Hyper-parameter Selection 168

3.7.2 Nested Cross-Validation 170

3.8 Pitfalls of Model Selection and Evaluation 172
3.8.1 Overlap between Training and Test Sets 172

3.8.2 Use of Validation Error as Generalization Error 172

3.9 Model Comparison 173
3.9.1 Estimating the Confidence Interval for Accuracy 174

3.9.2 Comparing the Performance of Two Models 175

3.10 Bibliographic Notes 176

3.11 Exercises 185

4 Classification: Alternative Techniques 193
4.1 Types of Classifiers 193

4.2 Rule-Based Classifier 195
4.2.1 How a Rule-Based Classifier Works 197

4.2.2 Properties of a Rule Set 198

4.2.3 Direct Methods for Rule Extraction 199

4.2.4 Indirect Methods for Rule Extraction 204

4.2.5 Characteristics of Rule-Based Classifiers 206

4.3 Nearest Neighbor Classifiers 208
4.3.1 Algorithm 209

4.3.2 Characteristics of Nearest Neighbor Classifiers 210


4.4 Naïve Bayes Classifier 212
4.4.1 Basics of Probability Theory 213

4.4.2 Naïve Bayes Assumption 218

4.5 Bayesian Networks 227
4.5.1 Graphical Representation 227

4.5.2 Inference and Learning 233

4.5.3 Characteristics of Bayesian Networks 242

4.6 Logistic Regression 243
4.6.1 Logistic Regression as a Generalized Linear Model 244

4.6.2 Learning Model Parameters 245

4.6.3 Characteristics of Logistic Regression 248

4.7 Artificial Neural Network (ANN) 249
4.7.1 Perceptron 250

4.7.2 Multi-layer Neural Network 254

4.7.3 Characteristics of ANN 261

4.8 Deep Learning 262
4.8.1 Using Synergistic Loss Functions 263

4.8.2 Using Responsive Activation Functions 266

4.8.3 Regularization 268

4.8.4 Initialization of Model Parameters 271

4.8.5 Characteristics of Deep Learning 275

4.9 Support Vector Machine (SVM) 276
4.9.1 Margin of a Separating Hyperplane 276

4.9.2 Linear SVM 278

4.9.3 Soft-margin SVM 284

4.9.4 Nonlinear SVM 290

4.9.5 Characteristics of SVM 294

4.10 Ensemble Methods 296
4.10.1 Rationale for Ensemble Method 297

4.10.2 Methods for Constructing an Ensemble Classifier 297

4.10.3 Bias-Variance Decomposition 300

4.10.4 Bagging 302

4.10.5 Boosting 305

4.10.6 Random Forests 310

4.10.7 Empirical Comparison among Ensemble Methods 312

4.11 Class Imbalance Problem 313
4.11.1 Building Classifiers with Class Imbalance 314

4.11.2 Evaluating Performance with Class Imbalance 318

4.11.3 Finding an Optimal Score Threshold 322

4.11.4 Aggregate Evaluation of Performance 323

4.12 Multiclass Problem 330

4.13 Bibliographic Notes 333

4.14 Exercises 345

5 Association Analysis: Basic Concepts and Algorithms 357
5.1 Preliminaries 358

5.2 Frequent Itemset Generation 362
5.2.1 The Apriori Principle 363

5.2.2 Frequent Itemset Generation in the Apriori Algorithm 364

5.2.3 Candidate Generation and Pruning 368

5.2.4 Support Counting 373

5.2.5 Computational Complexity 377

5.3 Rule Generation 380
5.3.1 Confidence-Based Pruning 380

5.3.2 Rule Generation in Apriori Algorithm 381

5.3.3 An Example: Congressional Voting Records 382

5.4 Compact Representation of Frequent Itemsets 384
5.4.1 Maximal Frequent Itemsets 384

5.4.2 Closed Itemsets 386

5.5 Alternative Methods for Generating Frequent Itemsets* 389

5.6 FP-Growth Algorithm* 393
5.6.1 FP-Tree Representation 394

5.6.2 Frequent Itemset Generation in FP-Growth Algorithm 397

5.7 Evaluation of Association Patterns 401

5.7.1 Objective Measures of Interestingness 402

5.7.2 Measures beyond Pairs of Binary Variables 414

5.7.3 Simpson’s Paradox 416

5.8 Effect of Skewed Support Distribution 418

5.9 Bibliographic Notes 424

5.10 Exercises 438

6 Association Analysis: Advanced Concepts 451
6.1 Handling Categorical Attributes 451

6.2 Handling Continuous Attributes 454
6.2.1 Discretization-Based Methods 454

6.2.2 Statistics-Based Methods 458

6.2.3 Non-discretization Methods 460

6.3 Handling a Concept Hierarchy 462

6.4 Sequential Patterns 464
6.4.1 Preliminaries 465

6.4.2 Sequential Pattern Discovery 468

6.4.3 Timing Constraints 473

6.4.4 Alternative Counting Schemes 477

6.5 Subgraph Patterns 479
6.5.1 Preliminaries 480

6.5.2 Frequent Subgraph Mining 483

6.5.3 Candidate Generation 487

6.5.4 Candidate Pruning 493

6.5.5 Support Counting 493

6.6 Infrequent Patterns 493
6.6.1 Negative Patterns 494

6.6.2 Negatively Correlated Patterns 495

6.6.3 Comparisons among Infrequent Patterns, Negative Patterns,
and Negatively Correlated Patterns 496

6.6.4 Techniques for Mining Interesting Infrequent Patterns 498

6.6.5 Techniques Based on Mining Negative Patterns 499

6.6.6 Techniques Based on Support Expectation 501

6.7 Bibliographic Notes 505

6.8 Exercises 510

7 Cluster Analysis: Basic Concepts and Algorithms 525
7.1 Overview 528

7.1.1 What Is Cluster Analysis? 528

7.1.2 Different Types of Clusterings 529

7.1.3 Different Types of Clusters 531

7.2 K-means 534
7.2.1 The Basic K-means Algorithm 535

7.2.2 K-means: Additional Issues 544

7.2.3 Bisecting K-means 547

7.2.4 K-means and Different Types of Clusters 548

7.2.5 Strengths and Weaknesses 549

7.2.6 K-means as an Optimization Problem 549

7.3 Agglomerative Hierarchical Clustering 554
7.3.1 Basic Agglomerative Hierarchical Clustering Algorithm

7.3.2 Specific Techniques 557

7.3.3 The Lance-Williams Formula for Cluster Proximity 562

7.3.4 Key Issues in Hierarchical Clustering 563

7.3.5 Outliers 564

7.3.6 Strengths and Weaknesses 565

7.4 DBSCAN 565
7.4.1 Traditional Density: Center-Based Approach 565

7.4.2 The DBSCAN Algorithm 567

7.4.3 Strengths and Weaknesses 569

7.5 Cluster Evaluation 571
7.5.1 Overview 571

7.5.2 Unsupervised Cluster Evaluation Using Cohesion and
Separation 574

7.5.3 Unsupervised Cluster Evaluation Using the Proximity Matrix

7.5.4 Unsupervised Evaluation of Hierarchical Clustering 585

7.5.5 Determining the Correct Number of Clusters 587

7.5.6 Clustering Tendency 588

7.5.7 Supervised Measures of Cluster Validity 589

7.5.8 Assessing the Significance of Cluster Validity Measures

7.5.9 Choosing a Cluster Validity Measure 596

7.6 Bibliographic Notes 597

7.7 Exercises 603

8 Cluster Analysis: Additional Issues and Algorithms 613
8.1 Characteristics of Data, Clusters, and Clustering Algorithms

8.1.1 Example: Comparing K-means and DBSCAN 614

8.1.2 Data Characteristics 615

8.1.3 Cluster Characteristics 617

8.1.4 General Characteristics of Clustering Algorithms 619

8.2 Prototype-Based Clustering 621
8.2.1 Fuzzy Clustering 621

8.2.2 Clustering Using Mixture Models 627

8.2.3 Self-Organizing Maps (SOM) 637

8.3 Density-Based Clustering 644
8.3.1 Grid-Based Clustering 644

8.3.2 Subspace Clustering 648

8.3.3 DENCLUE: A Kernel-Based Scheme for Density-Based
Clustering 652

8.4 Graph-Based Clustering 656
8.4.1 Sparsification 657

8.4.2 Minimum Spanning Tree (MST) Clustering 658

8.4.3 OPOSSUM: Optimal Partitioning of Sparse Similarities Using

8.4.4 Chameleon: Hierarchical Clustering with Dynamic Modeling

8.4.5 Spectral Clustering 666

8.4.6 Shared Nearest Neighbor Similarity 673

8.4.7 The Jarvis-Patrick Clustering Algorithm 676

8.4.8 SNN Density 678

8.4.9 SNN Density-Based Clustering 679

8.5 Scalable Clustering Algorithms 681
8.5.1 Scalability: General Issues and Approaches 681

8.5.2 BIRCH 684

8.5.3 CURE 686

8.6 Which Clustering Algorithm? 690

8.7 Bibliographic Notes 693

8.8 Exercises 699

9 Anomaly Detection 703
9.1 Characteristics of Anomaly Detection Problems 705

9.1.1 A Definition of an Anomaly 705

9.1.2 Nature of Data 706

9.1.3 How Anomaly Detection is Used 707

9.2 Characteristics of Anomaly Detection Methods 708

9.3 Statistical Approaches 710
9.3.1 Using Parametric Models 710

9.3.2 Using Non-parametric Models 714

9.3.3 Modeling Normal and Anomalous Classes 715

9.3.4 Assessing Statistical Significance 717

9.3.5 Strengths and Weaknesses 718

9.4 Proximity-based Approaches 719
9.4.1 Distance-based Anomaly Score 719

9.4.2 Density-based Anomaly Score 720

9.4.3 Relative Density-based Anomaly Score 722

9.4.4 Strengths and Weaknesses 723

9.5 Clustering-based Approaches 724
9.5.1 Finding Anomalous Clusters 724

9.5.2 Finding Anomalous Instances 725

9.5.3 Strengths and Weaknesses 728

9.6 Reconstruction-based Approaches 728
9.6.1 Strengths and Weaknesses 731

9.7 One-class Classification 732
9.7.1 Use of Kernels 733

9.7.2 The Origin Trick 734

9.7.3 Strengths and Weaknesses 738

9.8 Information Theoretic Approaches 738
9.8.1 Strengths and Weaknesses 740

9.9 Evaluation of Anomaly Detection 740

9.10 Bibliographic Notes 742

9.11 Exercises 749

10 Avoiding False Discoveries 755
10.1 Preliminaries: Statistical Testing 756

10.1.1 Significance Testing 756

10.1.2 Hypothesis Testing 761

10.1.3 Multiple Hypothesis Testing 767

10.1.4 Pitfalls in Statistical Testing 776

10.2 Modeling Null and Alternative Distributions 778
10.2.1 Generating Synthetic Data Sets 781

10.2.2 Randomizing Class Labels 782

10.2.3 Resampling Instances 782

10.2.4 Modeling the Distribution of the Test Statistic 783

10.3 Statistical Testing for Classification 783
10.3.1 Evaluating Classification Performance 783

10.3.2 Binary Classification as Multiple Hypothesis Testing 785

10.3.3 Multiple Hypothesis Testing in Model Selection 786

10.4 Statistical Testing for Association Analysis 787
10.4.1 Using Statistical Models 788

10.4.2 Using Randomization Methods 794

10.5 Statistical Testing for Cluster Analysis 795
10.5.1 Generating a Null Distribution for Internal Indices 796

10.5.2 Generating a Null Distribution for External Indices 798

10.5.3 Enrichment 798

10.6 Statistical Testing for Anomaly Detection 800

10.7 Bibliographic Notes 803

10.8 Exercises 808

Author Index 816

Subject Index 829

Copyright Permissions 839

1 Introduction

Rapid advances in data collection and storage
technology, coupled with the ease with which data can
be generated and disseminated, have triggered the
explosive growth of data, leading to the current age of
big data. Deriving actionable insights from these large
data sets is increasingly important in decision making
across almost all areas of society, including business
and industry; science and engineering; medicine and
biotechnology; and government and individuals.
However, the amount of data (volume), its complexity
(variety), and the rate at which it is being collected and
processed (velocity) have simply become too great for
humans to analyze unaided. Thus, there is a great
need for automated tools for extracting useful
information from the big data despite the challenges
posed by its enormity and diversity.

Data mining blends traditional data analysis methods
with sophisticated algorithms for processing this
abundance of data. In this introductory chapter, we
present an overview of data mining and outline the key
topics to be covered in this book. We start with a

description of some applications that require more
advanced techniques for data analysis.

Business and Industry Point-of-sale data collection (bar code scanners,
radio frequency identification (RFID), and smart card technology) have
allowed retailers to collect up-to-the-minute data about customer purchases at
the checkout counters of their stores. Retailers can utilize this information,
along with other business-critical data, such as web server logs from e-
commerce websites and customer service records from call centers, to help
them better understand the needs of their customers and make more informed
business decisions.

Data mining techniques can be used to support a wide range of business
intelligence applications, such as customer profiling, targeted marketing,
workflow management, store layout, fraud detection, and automated buying
and selling. An example of the last application is high-speed stock trading,
where decisions on buying and selling have to be made in less than a second
using data about financial transactions. Data mining can also help retailers
answer important business questions, such as “Who are the most profitable
customers?” “What products can be cross-sold or up-sold?” and “What is the
revenue outlook of the company for next year?” These questions have
inspired the development of such data mining techniques as association
analysis (Chapters 5 and 6 ).

As the Internet continues to revolutionize the way we interact and make
decisions in our everyday lives, we are generating massive amounts of data
about our online experiences, e.g., web browsing, messaging, and posting on
social networking websites. This has opened several opportunities for
business applications that use web data. For example, in the e-commerce
sector, data about our online viewing or shopping preferences can be used to

provide personalized recommendations of products. Data mining also plays a
prominent role in supporting several other Internet-based services, such as
filtering spam messages, answering search queries, and suggesting social
updates and connections. The large corpus of text, images, and videos
available on the Internet has enabled a number of advancements in data
mining methods, including deep learning, which is discussed in Chapter 4 .
These developments have led to great advances in a number of applications,
such as object recognition, natural language translation, and autonomous

Another domain that has undergone a rapid big data transformation is the use
of mobile sensors and devices, such as smart phones and wearable
computing devices. With better sensor technologies, it has become possible
to collect a variety of information about our physical world using low-cost
sensors embedded on everyday objects that are connected to each other,
termed the Internet of Things (IOT). This deep integration of physical sensors
in digital systems is beginning to generate large amounts of diverse and
distributed data about our environment, which can be used for designing
convenient, safe, and energy-efficient home systems, as well as for urban
planning of smart cities.

Medicine, Science, and Engineering Researchers in medicine, science, and
engineering are rapidly accumulating data that is key to significant new
discoveries. For example, as an important step toward improving our
understanding of the Earth’s climate system, NASA has deployed a series of
Earth-orbiting satellites that continuously generate global observations of the
land surface, oceans, and atmosphere. However, because of the size and
spatio-temporal nature of the data, traditional methods are often not suitable
for analyzing these data sets. Techniques developed in data mining can aid
Earth scientists in answering questions such as the following: “What is the
relationship between the frequency and intensity of ecosystem disturbances

such as droughts and hurricanes to global warming?” “How is land surface
precipitation and temperature affected by ocean surface temperature?” and
“How well can we predict the beginning and end of the growing season for a

As another example, researchers in molecular biology hope to use the large
amounts of genomic data to better understand the structure and function of
genes. In the past, traditional methods in molecular biology allowed scientists
to study only a few genes at a time in a given experiment. Recent
breakthroughs in microarray technology have enabled scientists to compare
the behavior of thousands of genes under various situations. Such
comparisons can help determine the function of each gene, and perhaps
isolate the genes responsible for certain diseases. However, the noisy, high-
dimensional nature of data requires new data analysis methods. In addition to
analyzing gene expression data, data mining can also be used to address
other important biological challenges such as protein structure prediction,
multiple sequence alignment, the modeling of biochemical pathways, and

Another example is the use of data mining techniques to analyze electronic
health record (EHR) data, which has become increasingly available. Not very
long ago, studies of patients required manually examining the physical
records of individual patients and extracting very specific pieces of information
pertinent to the particular question being investigated. EHRs allow for a faster
and broader exploration of such data. However, there are significant
challenges since the observations on any one patient typically occur during
their visits to a doctor or hospital and only a small number of details about the
health of the patient are measured during any particular visit.

Currently, EHR analysis focuses on simple types of data, e.g., a patient’s
blood pressure or the diagnosis code of a disease. However, large amounts of

more complex types of medical data are also being collected, such as
electrocardiograms (ECGs) and neuroimages from magnetic resonance
imaging (MRI) or functional Magnetic Resonance Imaging (fMRI). Although
challenging to analyze, this data also provides vital information about patients.
Integrating and analyzing such data, with traditional EHR and genomic data is
one of the capabilities needed to enable precision medicine, which aims to
provide more personalized patient care.

1.1 What Is Data Mining?
Data mining is the process of automatically discovering useful information in
large data repositories. Data mining techniques are deployed to scour large
data sets in order to find novel and useful patterns that might otherwise
remain unknown. They also provide the capability to predict the outcome of a
future observation, such as the amount a customer will spend at an online or a
brick-and-mortar store.

Not all information discovery tasks are considered to be data mining.
Examples include queries, e.g., looking up individual records in a database or
finding web pages that contain a particular set of keywords. This is because
such tasks can be accomplished through simple interactions with a database
management system or an information retrieval system. These systems rely
on traditional computer science techniques, which include sophisticated
indexing structures and query processing algorithms, for efficiently organizing
and retrieving information from large data repositories. Nonetheless, data
mining techniques have been used to enhance the performance of such
systems by improving the quality of the search results based on their
relevance to the input queries.

Data Mining and Knowledge Discovery in
Data mining is an integral part of knowledge discovery in databases (KDD),
which is the overall process of converting raw data into useful information, as
shown in Figure 1.1 . This process consists of a series of steps, from data
preprocessing to postprocessing of data mining results.

Figure 1.1.
The process of knowledge discovery in databases (KDD).

The input data can be stored in a variety of formats (flat files, spreadsheets, or
relational tables) and may reside in a centralized data repository or be
distributed across multiple sites. The purpose of preprocessing is to
transform the raw input data into an appropriate format for subsequent
analysis. The steps involved in data preprocessing include fusing data from
multiple sources, cleaning data to remove noise and duplicate observations,
and selecting records and features that are relevant to the data mining task at
hand. Because of the many ways data can be collected and stored, data
preprocessing is perhaps the most laborious and time-consuming step in the
overall knowledge discovery process.

“Closing the loop” is a phrase often used to refer to the process of integrating
data mining results into decision support systems. For example, in business
applications, the insights offered by data mining results can be integrated with
campaign management tools so that effective marketing promotions can be
conducted and tested. Such integration requires a postprocessing step to
ensure that only valid and useful results are incorporated into the decision
support system. An example of postprocessing is visualization, which allows
analysts to explore the data and the data mining results from a variety of
viewpoints. Hypothesis testing methods can also be applied during

postprocessing to eliminate spurious data mining results. (See Chapter
10 .)

1.2 Motivating Challenges
As mentioned earlier, traditional data analysis techniques have often
encountered practical difficulties in meeting the challenges posed by big data
applications. The following are some of the specific challenges that motivated
the development of data mining.


Because of advances in data generation and collection, data sets with sizes of
terabytes, petabytes, or even exabytes are becoming common. If data mining
algorithms are to handle these massive data sets, they must be scalable.
Many data mining algorithms employ special search strategies to handle
exponential search problems. Scalability may also require the implementation
of novel data structures to access individual records in an efficient manner.
For instance, out-of-core algorithms may be necessary when processing data
sets that cannot fit into main memory. Scalability can also be improved by
using sampling or developing parallel and distributed algorithms. A general
overview of techniques for scaling up data mining algorithms is given in
Appendix F.

High Dimensionality

It is now common to encounter data sets with hundreds or thousands of
attributes instead of the handful common a few decades ago. In
bioinformatics, progress in microarray technology has produced gene
expression data involving thousands of features. Data sets with temporal or
spatial components also tend to have high dimensionality. For example,

consider a data set that contains measurements of temperature at various
locations. If the temperature measurements are taken repeatedly for an
extended period, the number of dimensions (features) increases in proportion
to the number of measurements taken. Traditional data analysis techniques
that were developed for low-dimensional data often do not work well for such
high-dimensional data due to issues such as curse of dimensionality (to be
discussed in Chapter 2 ). Also, for some data analysis algorithms, the
computational complexity increases rapidly as the dimensionality (the number
of features) increases.

Heterogeneous and Complex Data

Traditional data analysis methods often deal with data sets containing
attributes of the same type, either continuous or categorical. As the role of
data mining in business, science, medicine, and other fields has grown, so
has the need for techniques that can handle heterogeneous attributes. Recent
years have also seen the emergence of more complex data objects.
Examples of such non-traditional types of data include web and social media
data containing text, hyperlinks, images, audio, and videos; DNA data with
sequential and three-dimensional structure; and climate data that consists of
measurements (temperature, pressure, etc.) at various times and locations on
the Earth’s surface. Techniques developed for mining such complex objects
should take into consideration relationships in the data, such as temporal and
spatial autocorrelation, graph connectivity, and parent-child relationships
between the elements in semi-structured text and XML documents.

Data Ownership and Distribution

Sometimes, the data needed for an analysis is not stored in one location or
owned by one organization. Instead, the data is geographically distributed
among resources belonging to multiple entities. This requires the development

of distributed data mining techniques. The key challenges faced by distributed
data mining algorithms include the following: (1) how to reduce the amount of
communication needed to perform the distributed computation, (2) how to
effectively consolidate the data mining results obtained from multiple sources,
and (3) how to address data security and privacy issues.

Non-traditional Analysis

The traditional statistical approach is based on a hypothesize-and-test
paradigm. In other words, a hypothesis is proposed, an experiment is
designed to gather the data, and then the data is analyzed with respect to the
hypothesis. Unfortunately, this process is extremely labor-intensive. Current
data analysis tasks often require the generation and evaluation of thousands
of hypotheses, and consequently, the development of some data mining
techniques has been motivated by the desire to automate the process of
hypothesis generation and evaluation. Furthermore, the data sets analyzed in
data mining are typically not the result of a carefully designed experiment and
often represent opportunistic samples of the data, rather than random

1.3 The Origins of Data Mining
While data mining has traditionally been viewed as an intermediate process
within the KDD framework, as shown in Figure 1.1 , it has emerged over the
years as an academic field within computer science, focusing on all aspects of
KDD, including data preprocessing, mining, and postprocessing. Its origin can
be traced back to the late 1980s, following a series of workshops organized
on the topic of knowledge discovery in databases. The workshops brought
together researchers from different disciplines to discuss the challenges and
opportunities in applying computational techniques to extract actionable
knowledge from large databases. The workshops quickly grew into hugely
popular conferences that were attended by researchers and practitioners from
both the academia and industry. The success of these conferences, along
with the interest shown by businesses and industry in recruiting new hires with
data mining background, have fueled the tremendous growth of this field.

The field was initially built upon the methodology and algorithms that
researchers had previously used. In particular, data mining researchers draw
upon ideas, such as (1) sampling, estimation, and hypothesis testing from
statistics and (2) search algorithms, modeling techniques, and learning
theories from artificial intelligence, pattern recognition, and machine learning.
Data mining has also been quick to adopt ideas from other areas, including
optimization, evolutionary computing, information theory, signal processing,
visualization, and information retrieval, and extending them to solve the
challenges of mining big data.

A number of other areas also play key supporting roles. In particular, database
systems are needed to provide support for efficient storage, indexing, and
query processing. Techniques from high performance (parallel) computing are

often important in addressing the massive size of some data sets. Distributed
techniques can also help address the issue of size and are essential when the
data cannot be gathered in one location. Figure 1.2 shows the relationship
of data mining to other areas.

Figure 1.2.
Data mining as a confluence of many disciplines.

Data Science and Data-Driven Discovery
Data science is an interdisciplinary field that studies and applies tools and
techniques for deriving useful insights from data. Although data science is
regarded as an emerging field with a distinct identity of its own, the tools and
techniques often come from many different areas of data analysis, such as
data mining, statistics, AI, machine learning, pattern recognition, database
technology, and distributed and parallel computing. (See Figure 1.2 .)

The emergence of data science as a new field is a recognition that, often,
none of the existing areas of data analysis provides a complete set of tools for
the data analysis tasks that are often encountered in emerging applications.

Instead, a broad range of computational, mathematical, and statistical skills is
often required. To illustrate the challenges that arise in analyzing such data,
consider the following example. Social media and the Web present new
opportunities for social scientists to observe and quantitatively measure
human behavior on a large scale. To conduct such a study, social scientists
work with analysts who possess skills in areas such as web mining, natural
language processing (NLP), network analysis, data mining, and statistics.
Compared to more traditional research in social science, which is often based
on surveys, this analysis requires a broader range of skills and tools, and
involves far larger amounts of data. Thus, data science is, by necessity, a
highly interdisciplinary field that builds on the continuing work of many fields.

The data-driven approach of data science emphasizes the direct discovery of
patterns and relationships from data, especially in large quantities of data,
often without the need for extensive domain knowledge. A notable example of
the success of this approach is represented by advances in neural networks,
i.e., deep learning, which have been particularly successful in areas which
have long proved challenging, e.g., recognizing objects in photos or videos
and words in speech, as well as in other application areas. However, note that
this is just one example of the success of data-driven approaches, and
dramatic improvements have also occurred in many other areas of data
analysis. Many of these developments are topics described later in this book.

Some cautions on potential limitations of a purely data-driven approach are
given in the Bibliographic Notes.

1.4 Data Mining Tasks
Data mining tasks are generally divided into two major categories:

Predictive tasks The objective of these tasks is to predict the value of a
particular attribute based on the values of other attributes. The attribute to be
predicted is commonly known as the target or dependent variable, while the
attributes used for making the prediction are known as the explanatory or
independent variables.

Descriptive tasks Here, the objective is to derive patterns (correlations,
trends, clusters, trajectories, and anomalies) that summarize the underlying
relationships in data. Descriptive data mining tasks are often exploratory in
nature and frequently require postprocessing techniques to validate and
explain the results.

Figure 1.3 illustrates four of the core data mining tasks that are described
in the remainder of this book.

Figure 1.3.
Four of the core data mining tasks.

Predictive modeling refers to the task of building a model for the target
variable as a function of the explanatory variables. There are two types of
predictive modeling tasks: classification, which is used for discrete target
variables, and regression, which is used for continuous target variables. For
example, predicting whether a web user will make a purchase at an online
bookstore is a classification task because the target variable is binary-valued.
On the other hand, forecasting the future price of a stock is a regression task
because price is a continuous-valued attribute. The goal of both tasks is to
learn a model that minimizes the error between the predicted and true values
of the target variable. Predictive modeling can be used to identify customers
who will respond to a marketing campaign, predict disturbances in the Earth’s

ecosystem, or judge whether a patient has a particular disease based on the
results of medical tests.

Example 1.1 (Predicting the Type of a Flower).
Consider the task of predicting a species of flower based on the
characteristics of the flower. In particular, consider classifying an Iris flower
as one of the following three Iris species: Setosa, Versicolour, or Virginica.
To perform this task, we need a data set containing the characteristics of
various flowers of these three species. A data set with this type of
information is the well-known Iris data set from the UCI Machine Learning
Repository at http://www.ics.uci.edu/~mlearn. In addition to the species
of a flower, this data set contains four other attributes: sepal width, sepal
length, petal length, and petal width. Figure 1.4 shows a plot of petal
width versus petal length for the 150 flowers in the Iris data set. Petal width
is broken into the categories low, medium, and high, which correspond to
the intervals [0, 0.75), [0.75, 1.75), , respectively. Also, petal
length is broken into categories low, medium,and high, which correspond
to the intervals [0, 2.5), [2.5, 5), , respectively. Based on these
categories of petal width and length, the following rules can be derived:

Petal width low and petal length low implies Setosa.

Petal width medium and petal length medium implies Versicolour.

Petal width high and petal length high implies Virginica.

While these rules do not classify all the flowers, they do a good (but not
perfect) job of classifying most of the flowers. Note that flowers from the
Setosa species are well separated from the Versicolour and Virginica
species with respect to petal width and length, but the latter two species
overlap somewhat with respect to these attributes.

[1.75, ∞)

[5, ∞)

Figure 1.4.
Petal width versus petal length for 150 Iris flowers.

Association analysis is used to discover patterns that describe strongly
associated features in the data. The discovered patterns are typically
represented in the form of implication rules or feature subsets. Because of the
exponential size of its search space, the goal of association analysis is to
extract the most interesting patterns in an efficient manner. Useful applications
of association analysis include finding groups of genes that have related
functionality, identifying web pages that are accessed together, or
understanding the relationships between different elements of Earth’s climate

Example 1.2 (Market Basket Analysis).

The transactions shown in Table 1.1 illustrate point-of-sale data
collected at the checkout counters of a grocery store. Association analysis
can be applied to find items that are frequently bought together by
customers. For example, we may discover the rule ,
which suggests that customers who buy diapers also tend to buy milk. This
type of rule can be used to identify potential cross-selling opportunities
among related items.

Table 1.1. Market basket data.

Transaction ID Items

1 {Bread, Butter, Diapers, Milk}

2 {Coffee, Sugar, Cookies, Salmon}

3 {Bread, Butter, Coffee, Diapers, Milk, Eggs}

4 {Bread, Butter, Salmon, Chicken}

5 {Eggs, Bread, Butter}

6 {Salmon, Diapers, Milk}

7 {Bread, Tea, Sugar, Eggs}

8 {Coffee, Sugar, Chicken, Eggs}

9 {Bread, Diapers, Milk, Salt}

10 {Tea, Eggs, Cookies, Diapers, Milk}

Cluster analysis seeks to find groups of closely related observations so that
observations that belong to the same cluster are more similar to each other
than observations that belong to other clusters. Clustering has been used to


group sets of related customers, find areas of the ocean that have a
significant impact on the Earth’s climate, and compress data.

Example 1.3 (Document Clustering).
The collection of news articles shown in Table 1.2 can be grouped
based on their respective topics. Each article is represented as a set of
word-frequency pairs (w : c), where w is a word and c is the number of
times the word appears in the article. There are two natural clusters in the
data set. The first cluster consists of the first four articles, which
correspond to news about the economy, while the second cluster contains
the last four articles, which correspond to news about health care. A good
clustering algorithm should be able to identify these two clusters based on
the similarity between words that appear in the articles.

Table 1.2. Collection of news articles.

Article Word-frequency pairs

1 dollar: 1, industry: 4, country: 2, loan: 3, deal: 2, government: 2

2 machinery: 2, labor: 3, market: 4, industry: 2, work: 3, country: 1

3 job: 5, inflation: 3, rise: 2, jobless: 2, market: 3, country: 2, index: 3

4 domestic: 3, forecast: 2, gain: 1, market: 2, sale: 3, price: 2

5 patient: 4, symptom: 2, drug: 3, health: 2, clinic: 2, doctor: 2

6 pharmaceutical: 2, company: 3, drug: 2, vaccine: 1, flu: 3

7 death: 2, cancer: 4, drug: 3, public: 4, health: 3, director: 2

8 medical: 2, cost: 3, increase: 2, patient: 2, health: 3, care: 1

Anomaly detection is the task of identifying observations whose
characteristics are significantly different from the rest of the data. Such
observations are known as anomalies or outliers. The goal of an anomaly
detection algorithm is to discover the real anomalies and avoid falsely labeling
normal objects as anomalous. In other words, a good anomaly detector must
have a high detection rate and a low false alarm rate. Applications of anomaly
detection include the detection of fraud, network intrusions, unusual patterns
of disease, and ecosystem disturbances, such as droughts, floods, fires,
hurricanes, etc.

Example 1.4 (Credit Card Fraud Detection).
A credit card company records the transactions made by every credit card
holder, along with personal information such as credit limit, age, annual
income, and address. Since the number of fraudulent cases is relatively
small compared to the number of legitimate transactions, anomaly
detection techniques can be applied to build a profile of legitimate
transactions for the users. When a new transaction arrives, it is compared
against the profile of the user. If the characteristics of the transaction are
very different from the previously created profile, then the transaction is
flagged as potentially fraudulent.

1.5 Scope and Organization of the
This book introduces the major principles and techniques used in data mining
from an algorithmic perspective. A study of these principles and techniques is
essential for developing a better understanding of how data mining technology
can be applied to various kinds of data. This book also serves as a starting
point for readers who are interested in doing research in this field.

We begin the technical discussion of this book with a chapter on data
(Chapter 2 ), which discusses the basic types of data, data quality,
preprocessing techniques, and measures of similarity and dissimilarity.
Although this material can be covered quickly, it provides an essential
foundation for data analysis. Chapters 3 and 4 cover classification.
Chapter 3 provides a foundation by discussing decision tree classifiers and
several issues that are important to all classification: overfitting, underfitting,
model selection, and performance evaluation. Using this foundation, Chapter
4 describes a number of other important classification techniques: rule-
based systems, nearest neighbor classifiers, Bayesian classifiers, artificial
neural networks, including deep learning, support vector machines, and
ensemble classifiers, which are collections of classifiers. The multiclass and
imbalanced class problems are also discussed. These topics can be covered

Association analysis is explored in Chapters 5 and 6 . Chapter 5
describes the basics of association analysis: frequent itemsets, association
rules, and some of the algorithms used to generate them. Specific types of
frequent itemsets—maximal, closed, and hyperclique—that are important for

data mining are also discussed, and the chapter concludes with a discussion
of evaluation measures for association analysis. Chapter 6 considers a
variety of more advanced topics, including how association analysis can be
applied to categorical and continuous data or to data that has a concept
hierarchy. (A concept hierarchy is a hierarchical categorization of objects, e.g.,
store items .) This chapter also
describes how association analysis can be extended to find sequential
patterns (patterns involving order), patterns in graphs, and negative
relationships (if one item is present, then the other is not).

Cluster analysis is discussed in Chapters 7 and 8 . Chapter 7 first
describes the different types of clusters, and then presents three specific
clustering techniques: K-means, agglomerative hierarchical clustering, and
DBSCAN. This is followed by a discussion of techniques for validating the
results of a clustering algorithm. Additional clustering concepts and
techniques are explored in Chapter 8 , including fuzzy and probabilistic
clustering, Self-Organizing Maps (SOM), graph-based clustering, spectral
clustering, and density-based clustering. There is also a discussion of
scalability issues and factors to consider when selecting a clustering

Chapter 9 , is on anomaly detection. After some basic definitions, several
different types of anomaly detection are considered: statistical, distance-
based, density-based, clustering-based, reconstruction-based, one-class
classification, and information theoretic. The last chapter, Chapter 10 ,
supplements the discussions in the other Chapters with a discussion of the
statistical concepts important for avoiding spurious results, and then
discusses those concepts in the context of data mining techniques studied in
the previous chapters. These techniques include statistical hypothesis testing,
p-values, the false discovery rate, and permutation testing. Appendices A
through F give a brief review of important topics that are used in portions of

store items→clothing→shoes→sneakers

the book: linear algebra, dimensionality reduction, statistics, regression,
optimization, and scaling up data mining techniques for big data.

The subject of data mining, while relatively young compared to statistics or
machine learning, is already too large to cover in a single book. Selected
references to topics that are only briefly covered, such as data quality, are
provided in the Bibliographic Notes section of the appropriate chapter.
References to topics not covered in this book, such as mining streaming data
and privacy-preserving data mining are provided in the Bibliographic Notes of
this chapter.

1.6 Bibliographic Notes
The topic of data mining has inspired many textbooks. Introductory textbooks
include those by Dunham [16], Han et al. [29], Hand et al. [31], Roiger and
Geatz [50], Zaki and Meira [61], and Aggarwal [2]. Data mining books with a
stronger emphasis on business applications include the works by Berry and
Linoff [5], Pyle [47], and Parr Rud [45]. Books with an emphasis on statistical
learning include those by Cherkassky and Mulier [11], and Hastie et al. [32].
Similar books with an emphasis on machine learning or pattern recognition
are those by Duda et al. [15], Kantardzic [34], Mitchell [43], Webb [57], and
Witten and Frank [58]. There are also some more specialized books:
Chakrabarti [9] (web mining), Fayyad et al. [20] (collection of early articles on
data mining), Fayyad et al. [18] (visualization), Grossman et al. [25] (science
and engineering), Kargupta and Chan [35] (distributed data mining), Wang et
al. [56] (bioinformatics), and Zaki and Ho [60] (parallel data mining).

There are several conferences related to data mining. Some of the main
conferences dedicated to this field include the ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (KDD), the IEEE
International Conference on Data Mining (ICDM), the SIAM International
Conference on Data Mining (SDM), the European Conference on Principles
and Practice of Knowledge Discovery in Databases (PKDD), and the Pacific-
Asia Conference on Knowledge Discovery and Data Mining (PAKDD). Data
mining papers can also be found in other major conferences such as the
Conference and Workshop on Neural Information Processing Systems
(NIPS),the International Conference on Machine Learning (ICML), the ACM
SIGMOD/PODS conference, the International Conference on Very Large Data
Bases (VLDB), the Conference on Information and Knowledge Management
(CIKM), the International Conference on Data Engineering (ICDE), the

National Conference on Artificial Intelligence (AAAI), the IEEE International
Conference on Big Data, and the IEEE International Conference on Data
Science and Advanced Analytics (DSAA).

Journal publications on data mining include IEEE Transactions on Knowledge
and Data Engineering, Data Mining and Knowledge Discovery, Knowledge
and Information Systems, ACM Transactions on Knowledge Discovery from
Data, Statistical Analysis and Data Mining, and Information Systems. There
are various open-source data mining software available, including Weka [27]
and Scikit-learn [46]. More recently, data mining software such as Apache
Mahout and Apache Spark have been developed for large-scale problems on
the distributed computing platform.

There have been a number of general articles on data mining that define the
field or its relationship to other fields, particularly statistics. Fayyad et al. [19]
describe data mining and how it fits into the total knowledge discovery
process. Chen et al. [10] give a database perspective on data mining.
Ramakrishnan and Grama [48] provide a general discussion of data mining
and present several viewpoints. Hand [30] describes how data mining differs
from statistics, as does Friedman [21]. Lambert [40] explores the use of
statistics for large data sets and provides some comments on the respective
roles of data mining and statistics. Glymour et al. [23] consider the lessons
that statistics may have for data mining. Smyth et al. [53] describe how the
evolution of data mining is being driven by new types of data and applications,
such as those involving streams, graphs, and text. Han et al. [28] consider
emerging applications in data mining and Smyth [52] describes some
research challenges in data mining. Wu et al. [59] discuss how developments
in data mining research can be turned into practical tools. Data mining
standards are the subject of a paper by Grossman et al. [24]. Bradley [7]
discusses how data mining algorithms can be scaled to large data sets.

The emergence of new data mining applications has produced new
challenges that need to be addressed. For instance, concerns about privacy
breaches as a result of data mining have escalated in recent years,
particularly in application domains such as web commerce and health care.
As a result, there is growing interest in developing data mining algorithms that
maintain user privacy. Developing techniques for mining encrypted or
randomized data is known as privacy-preserving data mining. Some
general references in this area include papers by Agrawal and Srikant [3],
Clifton et al. [12] and Kargupta et al. [36]. Vassilios et al. [55] provide a survey.
Another area of concern is the bias in predictive models that may be used for
some applications, e.g., screening job applicants or deciding prison parole
[39]. Assessing whether such applications are producing biased results is
made more difficult by the fact that the predictive models used for such
applications are often black box models, i.e., models that are not interpretable
in any straightforward way.

Data science, its constituent fields, and more generally, the new paradigm of
knowledge discovery they represent [33], have great potential, some of which
has been realized. However, it is important to emphasize that data science
works mostly with observational data, i.e., data that was collected by various
organizations as part of their normal operation. The consequence of this is
that sampling biases are common and the determination of causal factors
becomes more problematic. For this and a number of other reasons, it is often
hard to interpret the predictive models built from this data [42, 49]. Thus,
theory, experimentation and computational simulations will continue to be the
methods of choice in many areas, especially those related to science.

More importantly, a purely data-driven approach often ignores the existing
knowledge in a particular field. Such models may perform poorly, for example,
predicting impossible outcomes or failing to generalize to new situations.
However, if the model does work well, e.g., has high predictive accuracy, then

this approach may be sufficient for practical purposes in some fields. But in
many areas, such as medicine and science, gaining insight into the underlying
domain is often the goal. Some recent work attempts to address these issues
in order to create theory-guided data science, which takes pre-existing domain
knowledge into account [17, 37].

Recent years have witnessed a growing number of applications that rapidly
generate continuous streams of data. Examples of stream data include
network traffic, multimedia streams, and stock prices. Several issues must be
considered when mining data streams, such as the limited amount of memory
available, the need for online analysis, and the change of the data over time.
Data mining for stream data has become an important area in data mining.
Some selected publications are Domingos and Hulten [14] (classification),
Giannella et al. [22] (association analysis), Guha et al. [26] (clustering), Kifer
et al. [38] (change detection), Papadimitriou et al. [44] (time series), and Law
et al. [41] (dimensionality reduction).

Another area of interest is recommender and collaborative filtering systems [1,
6, 8, 13, 54], which suggest movies, television shows, books, products, etc.
that a person might like. In many cases, this problem, or at least a component
of it, is treated as a prediction problem and thus, data mining techniques can
be applied [4, 51].

[1] G. Adomavicius and A. Tuzhilin. Toward the next generation of

recommender systems: A survey of the state-of-the-art and possible
extensions. IEEE transactions on knowledge and data engineering,
17(6):734–749, 2005.

[2] C. Aggarwal. Data mining: The Textbook. Springer, 2009.

[3] R. Agrawal and R. Srikant. Privacy-preserving data mining. In Proc. of
2000 ACMSIGMOD Intl. Conf. on Management of Data, pages 439–450,
Dallas, Texas, 2000. ACM Press.

[4] X. Amatriain and J. M. Pujol. Data mining methods for recommender
systems. In Recommender Systems Handbook, pages 227–262. Springer,

[5] M. J. A. Berry and G. Linoff. Data Mining Techniques: For Marketing,
Sales, and Customer Relationship Management. Wiley Computer
Publishing, 2nd edition, 2004.

[6] J. Bobadilla, F. Ortega, A. Hernando, and A. Gutiérrez. Recommender
systems survey. Knowledge-based systems, 46:109–132, 2013.

[7] P. S. Bradley, J. Gehrke, R. Ramakrishnan, and R. Srikant. Scaling mining
algorithms to large databases. Communications of the ACM, 45(8):38–43,

[8] R. Burke. Hybrid recommender systems: Survey and experiments. User
modeling and user-adapted interaction, 12(4):331–370, 2002.

[9] S. Chakrabarti. Mining the Web: Discovering Knowledge from Hypertext
Data. Morgan Kaufmann, San Francisco, CA, 2003.

[10] M.-S. Chen, J. Han, and P. S. Yu. Data Mining: An Overview from a
Database Perspective. IEEE Transactions on Knowledge and Data
Engineering, 8(6):866–883, 1996.

[11] V. Cherkassky and F. Mulier. Learning from Data: Concepts, Theory, and
Methods. Wiley-IEEE Press, 2nd edition, 1998.

[12] C. Clifton, M. Kantarcioglu, and J. Vaidya. Defining privacy for data
mining. In National Science Foundation Workshop on Next Generation
Data Mining, pages 126– 133, Baltimore, MD, November 2002.

[13] C. Desrosiers and G. Karypis. A comprehensive survey of neighborhood-
based recommendation methods. Recommender systems handbook,
pages 107–144, 2011.

[14] P. Domingos and G. Hulten. Mining high-speed data streams. In Proc. of
the 6th Intl. Conf. on Knowledge Discovery and Data Mining, pages 71–80,

Boston, Massachusetts, 2000. ACM Press.

[15] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley
… Sons, Inc., New York, 2nd edition, 2001.

[16] M. H. Dunham. Data Mining: Introductory and Advanced Topics. Prentice
Hall, 2006.

[17] J. H. Faghmous, A. Banerjee, S. Shekhar, M. Steinbach, V. Kumar, A. R.
Ganguly, and N. Samatova. Theory-guided data science for climate
change. Computer, 47(11):74–78, 2014.

[18] U. M. Fayyad, G. G. Grinstein, and A. Wierse, editors. Information
Visualization in Data Mining and Knowledge Discovery. Morgan Kaufmann
Publishers, San Francisco, CA, September 2001.

[19] U. M. Fayyad, G. Piatetsky-Shapiro, and P. Smyth. From Data Mining to
Knowledge Discovery: An Overview. In Advances in Knowledge Discovery
and Data Mining, pages 1–34. AAAI Press, 1996.

[20] U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy,
editors. Advances in Knowledge Discovery and Data Mining. AAAI/MIT
Press, 1996.

[21] J. H. Friedman. Data Mining and Statistics: What’s the Connection?
Unpublished. www-stat.stanford.edu/~jhf/ftp/dm-stat.ps, 1997.

[22] C. Giannella, J. Han, J. Pei, X. Yan, and P. S. Yu. Mining Frequent
Patterns in Data Streams at Multiple Time Granularities. In H. Kargupta, A.
Joshi, K. Sivakumar, and Y. Yesha, editors, Next Generation Data Mining,
pages 191–212. AAAI/MIT, 2003.

[23] C. Glymour, D. Madigan, D. Pregibon, and P. Smyth. Statistical Themes
and Lessons for Data Mining. Data Mining and Knowledge Discovery,
1(1):11–28, 1997.

[24] R. L. Grossman, M. F. Hornick, and G. Meyer. Data mining standards
initiatives. Communications of the ACM, 45(8):59–61, 2002.

[25] R. L. Grossman, C. Kamath, P. Kegelmeyer, V. Kumar, and R. Namburu,
editors. Data Mining for Scientific and Engineering Applications. Kluwer
Academic Publishers, 2001.

[26] S. Guha, A. Meyerson, N. Mishra, R. Motwani, and L. O’Callaghan.
Clustering Data Streams: Theory and Practice. IEEE Transactions on
Knowledge and Data Engineering, 15(3):515–528, May/June 2003.

[27] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H.
Witten. The WEKA Data Mining Software: An Update. SIGKDD
Explorations, 11(1), 2009.

[28] J. Han, R. B. Altman, V. Kumar, H. Mannila, and D. Pregibon. Emerging
scientific applications in data mining. Communications of the ACM,
45(8):54–58, 2002.

[29] J. Han, M. Kamber, and J. Pei. Data Mining: Concepts and Techniques.
Morgan Kaufmann Publishers, San Francisco, 3rd edition, 2011.

[30] D. J. Hand. Data Mining: Statistics and More? The American Statistician,
52(2): 112–118, 1998.

[31] D. J. Hand, H. Mannila, and P. Smyth. Principles of Data Mining. MIT
Press, 2001.

[32] T. Hastie, R. Tibshirani, and J. H. Friedman. The Elements of Statistical
Learning: Data Mining, Inference, Prediction. Springer, 2nd edition, 2009.

[33] T. Hey, S. Tansley, K. M. Tolle, et al. The fourth paradigm: data-intensive
scientific discovery, volume 1. Microsoft research Redmond, WA, 2009.

[34] M. Kantardzic. Data Mining: Concepts, Models, Methods, and Algorithms.
Wiley-IEEE Press, Piscataway, NJ, 2003.

[35] H. Kargupta and P. K. Chan, editors. Advances in Distributed and Parallel
Knowledge Discovery. AAAI Press, September 2002.

[36] H. Kargupta, S. Datta, Q. Wang, and K. Sivakumar. On the Privacy
Preserving Properties of Random Data Perturbation Techniques. In Proc.
of the 2003 IEEE Intl. Conf. on Data Mining, pages 99–106, Melbourne,
Florida, December 2003. IEEE Computer Society.

[37] A. Karpatne, G. Atluri, J. Faghmous, M. Steinbach, A. Banerjee, A.
Ganguly, S. Shekhar, N. Samatova, and V. Kumar. Theory-guided Data
Science: A New Paradigm for Scientific Discovery from Data. IEEE
Transactions on Knowledge and Data Engineering, 2017.

[38] D. Kifer, S. Ben-David, and J. Gehrke. Detecting Change in Data
Streams. In Proc. of the 30th VLDB Conf., pages 180–191, Toronto,
Canada, 2004. Morgan Kaufmann.

[39] J. Kleinberg, J. Ludwig, and S. Mullainathan. A Guide to Solving Social
Problems with Machine Learning. Harvard Business Review, December

[40] D. Lambert. What Use is Statistics for Massive Data? In ACM SIGMOD
Workshop on Research Issues in Data Mining and Knowledge Discovery,
pages 54–62, 2000.

[41] M. H. C. Law, N. Zhang, and A. K. Jain. Nonlinear Manifold Learning for
Data Streams. In Proc. of the SIAM Intl. Conf. on Data Mining, Lake Buena
Vista, Florida, April 2004. SIAM.

[42] Z. C. Lipton. The mythos of model interpretability. arXiv preprint
arXiv:1606.03490, 2016.

[43] T. Mitchell. Machine Learning. McGraw-Hill, Boston, MA, 1997.

[44] S. Papadimitriou, A. Brockwell, and C. Faloutsos. Adaptive, unsupervised
stream mining. VLDB Journal, 13(3):222–239, 2004.

[45] O. Parr Rud. Data Mining Cookbook: Modeling Data for Marketing, Risk
and Customer Relationship Management. John Wiley … Sons, New York,
NY, 2001.

[46] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A.
Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-
learn: Machine Learning in Python. Journal of Machine Learning
Research, 12:2825–2830, 2011.

[47] D. Pyle. Business Modeling and Data Mining. Morgan Kaufmann, San
Francisco, CA, 2003.

[48] N. Ramakrishnan and A. Grama. Data Mining: From Serendipity to
Science—Guest Editors’ Introduction. IEEE Computer, 32(8):34–37, 1999.

[49] M. T. Ribeiro, S. Singh, and C. Guestrin. Why should i trust you?:
Explaining the predictions of any classifier. In Proceedings of the 22nd
ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, pages 1135–1144. ACM, 2016.

[50] R. Roiger and M. Geatz. Data Mining: A Tutorial Based Primer. Addison-
Wesley, 2002.

[51] J. Schafer. The Application of Data-Mining to Recommender Systems.
Encyclopedia of data warehousing and mining, 1:44–48, 2009.

[52] P. Smyth. Breaking out of the Black-Box: Research Challenges in Data
Mining. In Proc. of the 2001 ACM SIGMOD Workshop on Research Issues
in Data Mining and Knowledge Discovery, 2001.

[53] P. Smyth, D. Pregibon, and C. Faloutsos. Data-driven evolution of data
mining algorithms. Communications of the ACM, 45(8):33–37, 2002.

[54] X. Su and T. M. Khoshgoftaar. A survey of collaborative filtering
techniques. Advances in artificial intelligence, 2009:4, 2009.

[55] V. S. Verykios, E. Bertino, I. N. Fovino, L. P. Provenza, Y. Saygin, and Y.
Theodoridis. State-of-the-art in privacy preserving data mining. SIGMOD
Record, 33(1):50–57, 2004.

[56] J. T. L. Wang, M. J. Zaki, H. Toivonen, and D. E. Shasha, editors. Data
Mining in Bioinformatics. Springer, September 2004.

[57] A. R. Webb. Statistical Pattern Recognition. John Wiley … Sons, 2nd
edition, 2002.

[58] I. H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools
and Techniques. Morgan Kaufmann, 3rd edition, 2011.

[59] X. Wu, P. S. Yu, and G. Piatetsky-Shapiro. Data Mining: How Research
Meets Practical Development? Knowledge and Information Systems,
5(2):248–261, 2003.

[60] M. J. Zaki and C.-T. Ho, editors. Large-Scale Parallel Data Mining.
Springer, September 2002.

[61] M. J. Zaki and W. Meira Jr. Data Mining and Analysis: Fundamental
Concepts and Algorithms. Cambridge University Press, New York, 2014.

1.7 Exercises
1. Discuss whether or not each of the following activities is a data mining task.

a. Dividing the customers of a company according to their gender.

b. Dividing the customers of a company according to their profitability.

c. Computing the total sales of a company.

d. Sorting a student database based on student identification numbers.

e. Predicting the outcomes of tossing a (fair) pair of dice.

f. Predicting the future stock price of a company using historical records.

g. Monitoring the heart rate of a patient for abnormalities.

h. Monitoring seismic waves for earthquake activities.

i. Extracting the frequencies of a sound wave.

2. Suppose that you are employed as a data mining consultant for an Internet
search engine company. Describe how data mining can help the company by
giving specific examples of how techniques, such as clustering, classification,
association rule mining, and anomaly detection can be applied.

3. For each of the following data sets, explain whether or not data privacy is
an important issue.

a. Census data collected from 1900–1950.

b. IP addresses and visit times of web users who visit your website.

c. Images from Earth-orbiting satellites.

d. Names and addresses of people from the telephone book.

e. Names and email addresses collected from the Web.

2 Data

This chapter discusses several data-related issues that
are important for successful data mining:

The Type of Data Data sets differ in a number of ways. For example, the
attributes used to describe data objects can be of different types—quantitative
or qualitative—and data sets often have special characteristics; e.g., some
data sets contain time series or objects with explicit relationships to one
another. Not surprisingly, the type of data determines which tools and
techniques can be used to analyze the data. Indeed, new research in data
mining is often driven by the need to accommodate new application areas and
their new types of data.

The Quality of the Data Data is often far from perfect. While most data
mining techniques can tolerate some level of imperfection in the data, a focus
on understanding and improving data quality typically improves the quality of
the resulting analysis. Data quality issues that often need to be addressed
include the presence of noise and outliers; missing, inconsistent, or duplicate
data; and data that is biased or, in some other way, unrepresentative of the
phenomenon or population that the data is supposed to describe.

Preprocessing Steps to Make the Data More Suitable for Data Mining
Often, the raw data must be processed in order to make it suitable for

analysis. While one objective may be to improve data quality, other goals
focus on modifying the data so that it better fits a specified data mining
technique or tool. For example, a continuous attribute, e.g., length, sometimes
needs to be transformed into an attribute with discrete categories, e.g., short,
medium, or long, in order to apply a particular technique. As another example,
the number of attributes in a data set is often reduced because many
techniques are more effective when the data has a relatively small number of

Analyzing Data in Terms of Its Relationships One approach to data
analysis is to find relationships among the data objects and then perform the
remaining analysis using these relationships rather than the data objects
themselves. For instance, we can compute the similarity or distance between
pairs of objects and then perform the analysis—clustering, classification, or
anomaly detection—based on these similarities or distances. There are many
such similarity or distance measures, and the proper choice depends on the
type of data and the particular application.

Example 2.1 (An Illustration of Data-Related
To further illustrate the importance of these issues, consider the following
hypothetical situation. You receive an email from a medical researcher
concerning a project that you are eager to work on.


I’ve attached the data file that I mentioned in my previous email. Each line contains the

information for a single patient and consists of five fields. We want to predict the last field using

the other fields. I don’t have time to provide any more information about the data since I’m going

out of town for a couple of days, but hopefully that won’t slow you down too much. And if you

don’t mind, could we meet when I get back to discuss your preliminary results? I might invite a

few other members of my team.

Thanks and see you in a couple of days.

Despite some misgivings, you proceed to analyze the data. The first few rows
of the file are as follows:

012 232 33.5 0 10.7

020 121 16.9 2 210.1

027 165 24.0 0 427.6

A brief look at the data reveals nothing strange. You put your doubts aside
and start the analysis. There are only 1000 lines, a smaller data file than you
had hoped for, but two days later, you feel that you have made some
progress. You arrive for the meeting, and while waiting for others to arrive, you
strike up a conversation with a statistician who is working on the project.
When she learns that you have also been analyzing the data from the project,
she asks if you would mind giving her a brief overview of your results.

Statistician: So, you got the data for all the patients?

Data Miner: Yes. I haven’t had much time for analysis, but I do have a
few interesting results.

Statistician: Amazing. There were so many data issues with this set of
patients that I couldn’t do much.

Data Miner: Oh? I didn’t hear about any possible problems.

Statistician: Well, first there is field 5, the variable we want to predict.

It’s common knowledge among people who analyze this type of data
that results are better if you work with the log of the values, but I didn’t
discover this until later. Was it mentioned to you?

Data Miner: No.

Statistician: But surely you heard about what happened to field 4? It’s
supposed to be measured on a scale from 1 to 10, with 0 indicating a
missing value, but because of a data entry error, all 10’s were changed
into 0’s. Unfortunately, since some of the patients have missing values
for this field, it’s impossible to say whether a 0 in this field is a real 0 or
a 10. Quite a few of the records have that problem.

Data Miner: Interesting. Were there any other problems?

Statistician: Yes, fields 2 and 3 are basically the same, but I assume
that you probably noticed that.

Data Miner: Yes, but these fields were only weak predictors of field 5.

Statistician: Anyway, given all those problems, I’m surprised you were
able to accomplish anything.

Data Miner: True, but my results are really quite good. Field 1 is a very
strong predictor of field 5. I’m surprised that this wasn’t noticed before.

Statistician: What? Field 1 is just an identification number.

Data Miner: Nonetheless, my results speak for themselves.

Statistician: Oh, no! I just remembered. We assigned ID numbers after
we sorted the records based on field 5. There is a strong connection,
but it’s meaningless. Sorry.

Although this scenario represents an extreme situation, it emphasizes the
importance of “knowing your data.” To that end, this chapter will address each

of the four issues mentioned above, outlining some of the basic challenges
and standard approaches.

2.1 Types of Data
A data set can often be viewed as a collection of data objects. Other names
for a data object are record, point, vector, pattern, event, case, sample,
instance, observation, or entity. In turn, data objects are described by a
number of attributes that capture the characteristics of an object, such as the
mass of a physical object or the time at which an event occurred. Other
names for an attribute are variable, characteristic, field, feature, or dimension.

Example 2.2 (Student Information).
Often, a data set is a file, in which the objects are records (or rows) in the
file and each field (or column) corresponds to an attribute. For example,
Table 2.1 shows a data set that consists of student information. Each
row corresponds to a student and each column is an attribute that
describes some aspect of a student, such as grade point average (GPA) or
identification number (ID).

Table 2.1. A sample data set containing student information.

Student ID Year Grade Point Average (GPA) …

1034262 Senior 3.24 …

1052663 Freshman 3.51 …

1082246 Sophomore 3.62 …

Although record-based data sets are common, either in flat files or relational
database systems, there are other important types of data sets and systems
for storing data. In Section 2.1.2 , we will discuss some of the types of data
sets that are commonly encountered in data mining. However, we first
consider attributes.

2.1.1 Attributes and Measurement

In this section, we consider the types of attributes used to describe data
objects. We first define an attribute, then consider what we mean by the type
of an attribute, and finally describe the types of attributes that are commonly

What Is an Attribute?
We start with a more detailed definition of an attribute.

Definition 2.1.
An attribute is a property or characteristic of an object that can
vary, either from one object to another or from one time to

For example, eye color varies from person to person, while the temperature of
an object varies over time. Note that eye color is a symbolic attribute with a

small number of possible values {brown, black, blue, green, hazel, etc.} , while
temperature is a numerical attribute with a potentially unlimited number of

At the most basic level, attributes are not about numbers or symbols.
However, to discuss and more precisely analyze the characteristics of objects,
we assign numbers or symbols to them. To do this in a well-defined way, we
need a measurement scale.

Definition 2.2.
A measurement scale is a rule (function) that associates a
numerical or symbolic value with an attribute of an object.

Formally, the process of measurement is the application of a measurement
scale to associate a value with a particular attribute of a specific object. While
this may seem a bit abstract, we engage in the process of measurement all
the time. For instance, we step on a bathroom scale to determine our weight,
we classify someone as male or female, or we count the number of chairs in a
room to see if there will be enough to seat all the people coming to a meeting.
In all these cases, the “physical value” of an attribute of an object is mapped
to a numerical or symbolic value.

With this background, we can discuss the type of an attribute, a concept that
is important in determining if a particular data analysis technique is consistent
with a specific type of attribute.

The Type of an Attribute
It is common to refer to the type of an attribute as the type of a measurement
scale. It should be apparent from the previous discussion that an attribute can
be described using different measurement scales and that the properties of an
attribute need not be the same as the properties of the values used to
measure it. In other words, the values used to represent an attribute can have
properties that are not properties of the attribute itself, and vice versa. This is
illustrated with two examples.

Example 2.3 (Employee Age and ID Number).
Two attributes that might be associated with an employee are ID and age
(in years). Both of these attributes can be represented as integers.
However, while it is reasonable to talk about the average age of an
employee, it makes no sense to talk about the average employee ID.
Indeed, the only aspect of employees that we want to capture with the ID
attribute is that they are distinct. Consequently, the only valid operation for
employee IDs is to test whether they are equal. There is no hint of this
limitation, however, when integers are used to represent the employee ID
attribute. For the age attribute, the properties of the integers used to
represent age are very much the properties of the attribute. Even so, the
correspondence is not complete because, for example, ages have a
maximum, while integers do not.

Example 2.4 (Length of Line Segments).
Consider Figure 2.1 , which shows some objects—line segments—and
how the length attribute of these objects can be mapped to numbers in two
different ways. Each successive line segment, going from the top to the
bottom, is formed by appending the topmost line segment to itself. Thus,

the second line segment from the top is formed by appending the topmost
line segment to itself twice, the third line segment from the top is formed by
appending the topmost line segment to itself three times, and so forth. In a
very real (physical) sense, all the line segments are multiples of the first.
This fact is captured by the measurements on the right side of the figure,
but not by those on the left side. More specifically, the measurement scale
on the left side captures only the ordering of the length attribute, while the
scale on the right side captures both the ordering and additivity properties.
Thus, an attribute can be measured in a way that does not capture all the
properties of the attribute.

Figure 2.1.
The measurement of the length of line segments on two different scales of

Knowing the type of an attribute is important because it tells us which
properties of the measured values are consistent with the underlying

properties of the attribute, and therefore, it allows us to avoid foolish actions,
such as computing the average employee ID.

The Different Types of Attributes
A useful (and simple) way to specify the type of an attribute is to identify the
properties of numbers that correspond to underlying properties of the attribute.
For example, an attribute such as length has many of the properties of
numbers. It makes sense to compare and order objects by length, as well as
to talk about the differences and ratios of length. The following properties
(operations) of numbers are typically used to describe attributes.

1. Distinctness and
2. Order and
3. Addition and
4. Multiplication and /

Given these properties, we can define four types of attributes: nominal ,
ordinal, interval , and ratio. Table 2.2 gives the definitions of these types,
along with information about the statistical operations that are valid for each
type. Each attribute type possesses all of the properties and operations of the
attribute types above it. Consequently, any property or operation that is valid
for nominal, ordinal, and interval attributes is also valid for ratio attributes. In
other words, the definition of the attribute types is cumulative. However, this
does not mean that the statistical operations appropriate for one attribute type
are appropriate for the attribute types above it.

Table 2.2. Different attribute types.

Attribute Type Description Examples Operations

Categorical Nominal The values of a nominal attribute zip codes, mode,

= ≠
<, ≤, >, ≥

+ −

(Qualitative) are just different names; i.e.,
nominal values provide only
enough information to distinguish
one object from another.

employee ID
numbers, eye
color, gender


Ordinal The values of an ordinal attribute
provide enough information to
order objects.

hardness of
minerals, {good,
better, best},
grades, street

run tests,
sign tests


Interval For interval attributes, the
differences between values are
meaningful, i.e., a unit of
measurement exists.

calendar dates,
temperature in
Celsius or

t and F

Ratio For ratio variables, both
differences and ratios are

temperature in
Kelvin, monetary
quantities, counts,
age, mass,
length, electrical


Nominal and ordinal attributes are collectively referred to as categorical or
qualitative attributes. As the name suggests, qualitative attributes, such as
employee ID, lack most of the properties of numbers. Even if they are
represented by numbers, i.e., integers, they should be treated more like
symbols. The remaining two types of attributes, interval and ratio, are
collectively referred to as quantitative or numeric attributes. Quantitative
attributes are represented by numbers and have most of the properties of

(=, ≠) χ2

(<, >)

(+, −)

(×, /)

numbers. Note that quantitative attributes can be integer-valued or

The types of attributes can also be described in terms of transformations that
do not change the meaning of an attribute. Indeed, S. Smith Stevens, the
psychologist who originally defined the types of attributes shown in Table
2.2 , defined them in terms of these permissible transformations. For
example, the meaning of a length attribute is unchanged if it is measured in
meters instead of feet.

The statistical operations that make sense for a particular type of attribute are
those that will yield the same results when the attribute is transformed by
using a transformation that preserves the attribute’s meaning. To illustrate, the
average length of a set of objects is different when measured in meters rather
than in feet, but both averages represent the same length. Table 2.3 shows
the meaning-preserving transformations for the four attribute types of Table
2.2 .

Table 2.3. Transformations that define attribute levels.

Attribute Type Transformation Comment


Nominal Any one-to-one mapping,
e.g., a permutation of values

If all employee ID numbers are
reassigned, it will not make any

Ordinal An order-preserving change
of values, i.e.,

where f is a monotonic

An attribute encompassing the notion
of good, better, best can be
represented equally well by the values
{1, 2, 3} or by {0.5, 1, 10}.


a and b constants.

The Fahrenheit and Celsius
temperature scales differ in the



location of their zero value and the
size of a degree (unit).

Ratio Length can be measured in meters or

Example 2.5 (Temperature Scales).
Temperature provides a good illustration of some of the concepts that have
been described. First, temperature can be either an interval or a ratio
attribute, depending on its measurement scale. When measured on the
Kelvin scale, a temperature of 2 is, in a physically meaningful way, twice

that of a temperature of 1 . This is not true when temperature is measured

on either the Celsius or Fahrenheit scales, because, physically, a
temperature of 1 Fahrenheit (Celsius) is not much different than a

temperature of 2 Fahrenheit (Celsius). The problem is that the zero points

of the Fahrenheit and Celsius scales are, in a physical sense, arbitrary,
and therefore, the ratio of two Celsius or Fahrenheit temperatures is not
physically meaningful.

Describing Attributes by the Number of Values
An independent way of distinguishing between attributes is by the number of
values they can take.

Discrete A discrete attribute has a finite or countably infinite set of values.
Such attributes can be categorical, such as zip codes or ID numbers, or
numeric, such as counts. Discrete attributes are often represented using
integer variables. Binary attributes are a special case of discrete attributes
and assume only two values, e.g., true/false, yes/no, male/female, or 0/1.


Binary attributes are often represented as Boolean variables, or as integer
variables that only take the values 0 or 1.

Continuous A continuous attribute is one whose values are real numbers.
Examples include attributes such as temperature, height, or weight.
Continuous attributes are typically represented as floating-point variables.
Practically, real values can be measured and represented only with limited

In theory, any of the measurement scale types—nominal, ordinal, interval, and
ratio—could be combined with any of the types based on the number of
attribute values—binary, discrete, and continuous. However, some
combinations occur only infrequently or do not make much sense. For
instance, it is difficult to think of a realistic data set that contains a continuous
binary attribute. Typically, nominal and ordinal attributes are binary or discrete,
while interval and ratio attributes are continuous. However, count attributes ,
which are discrete, are also ratio attributes.

Asymmetric Attributes
For asymmetric attributes, only presence—a non-zero attribute value—is
regarded as important. Consider a data set in which each object is a student
and each attribute records whether a student took a particular course at a
university. For a specific student, an attribute has a value of 1 if the student
took the course associated with that attribute and a value of 0 otherwise.
Because students take only a small fraction of all available courses, most of
the values in such a data set would be 0. Therefore, it is more meaningful and
more efficient to focus on the non-zero values. To illustrate, if students are
compared on the basis of the courses they don’t take, then most students
would seem very similar, at least if the number of courses is large. Binary
attributes where only non-zero values are important are called asymmetric

binary attributes. This type of attribute is particularly important for
association analysis, which is discussed in Chapter 5 . It is also possible to
have discrete or continuous asymmetric features. For instance, if the number
of credits associated with each course is recorded, then the resulting data set
will consist of asymmetric discrete or continuous attributes.

General Comments on Levels of Measurement
As described in the rest of this chapter, there are many diverse types of data.
The previous discussion of measurement scales, while useful, is not complete
and has some limitations. We provide the following comments and guidance.

Distinctness, order, and meaningful intervals and ratios are only four
properties of data—many others are possible. For instance, some data
is inherently cyclical, e.g., position on the surface of the Earth or time. As
another example, consider set valued attributes, where each attribute
value is a set of elements, e.g., the set of movies seen in the last year.
Define one set of elements (movies) to be greater (larger) than a second
set if the second set is a subset of the first. However, such a relationship
defines only a partial order that does not match any of the attribute types
just defined.
The numbers or symbols used to capture attribute values may not
capture all the properties of the attributes or may suggest properties
that are not there. An illustration of this for integers was presented in
Example 2.3 , i.e., averages of IDs and out of range ages.
Data is often transformed for the purpose of analysis—see Section
2.3.7 . This often changes the distribution of the observed variable to a
distribution that is easier to analyze, e.g., a Gaussian (normal) distribution.
Often, such transformations only preserve the order of the original values,
and other properties are lost. Nonetheless, if the desired outcome is a

statistical test of differences or a predictive model, such a transformation is
The final evaluation of any data analysis, including operations on
attributes, is whether the results make sense from a domain point of

In summary, it can be challenging to determine which operations can be
performed on a particular attribute or a collection of attributes without
compromising the integrity of the analysis. Fortunately, established practice
often serves as a reliable guide. Occasionally, however, standard practices
are erroneous or have limitations.

2.1.2 Types of Data Sets

There are many types of data sets, and as the field of data mining develops
and matures, a greater variety of data sets become available for analysis. In
this section, we describe some of the most common types. For convenience,
we have grouped the types of data sets into three groups: record data, graph-
based data, and ordered data. These categories do not cover all possibilities
and other groupings are certainly possible.

General Characteristics of Data Sets
Before providing details of specific kinds of data sets, we discuss three
characteristics that apply to many data sets and have a significant impact on
the data mining techniques that are used: dimensionality, distribution, and


The dimensionality of a data set is the number of attributes that the objects in
the data set possess. Analyzing data with a small number of dimensions tends
to be qualitatively different from analyzing moderate or high-dimensional data.
Indeed, the difficulties associated with the analysis of high-dimensional data
are sometimes referred to as the curse of dimensionality. Because of this,
an important motivation in preprocessing the data is dimensionality
reduction. These issues are discussed in more depth later in this chapter and
in Appendix B.


The distribution of a data set is the frequency of occurrence of various values
or sets of values for the attributes comprising data objects. Equivalently, the
distribution of a data set can be considered as a description of the
concentration of objects in various regions of the data space. Statisticians
have enumerated many types of distributions, e.g., Gaussian (normal), and
described their properties. (See Appendix C.) Although statistical approaches
for describing distributions can yield powerful analysis techniques, many data
sets have distributions that are not well captured by standard statistical

As a result, many data mining algorithms do not assume a particular statistical
distribution for the data they analyze. However, some general aspects of
distributions often have a strong impact. For example, suppose a categorical
attribute is used as a class variable, where one of the categories occurs 95%
of the time, while the other categories together occur only 5% of the time. This
skewness in the distribution can make classification difficult as discussed in
Section 4.11. (Skewness has other impacts on data analysis that are not
discussed here.)

A special case of skewed data is sparsity. For sparse binary, count or
continuous data, most attributes of an object have values of 0. In many cases,
fewer than 1% of the values are non-zero. In practical terms, sparsity is an
advantage because usually only the non-zero values need to be stored and
manipulated. This results in significant savings with respect to computation
time and storage. Indeed, some data mining algorithms, such as the
association rule mining algorithms described in Chapter 5 , work well only
for sparse data. Finally, note that often the attributes in sparse data sets are
asymmetric attributes.


It is frequently possible to obtain data at different levels of resolution, and
often the properties of the data are different at different resolutions. For
instance, the surface of the Earth seems very uneven at a resolution of a few
meters, but is relatively smooth at a resolution of tens of kilometers. The
patterns in the data also depend on the level of resolution. If the resolution is
too fine, a pattern may not be visible or may be buried in noise; if the
resolution is too coarse, the pattern can disappear. For example, variations in
atmospheric pressure on a scale of hours reflect the movement of storms and
other weather systems. On a scale of months, such phenomena are not

Record Data
Much data mining work assumes that the data set is a collection of records
(data objects), each of which consists of a fixed set of data fields (attributes).
See Figure 2.2(a) . For the most basic form of record data, there is no
explicit relationship among records or data fields, and every record (object)
has the same set of attributes. Record data is usually stored either in flat files
or in relational databases. Relational databases are certainly more than a

collection of records, but data mining often does not use any of the additional
information available in a relational database. Rather, the database serves as
a convenient place to find records. Different types of record data are
described below and are illustrated in Figure 2.2 .

Figure 2.2.
Different variations of record data.

Transaction or Market Basket Data

Transaction data is a special type of record data, where each record
(transaction) involves a set of items. Consider a grocery store. The set of
products purchased by a customer during one shopping trip constitutes a
transaction, while the individual products that were purchased are the items.
This type of data is called market basket data because the items in each
record are the products in a person’s “market basket.” Transaction data is a
collection of sets of items, but it can be viewed as a set of records whose
fields are asymmetric attributes. Most often, the attributes are binary,
indicating whether an item was purchased, but more generally, the attributes
can be discrete or continuous, such as the number of items purchased or the
amount spent on those items. Figure 2.2(b) shows a sample transaction
data set. Each row represents the purchases of a particular customer at a
particular time.

The Data Matrix

If all the data objects in a collection of data have the same fixed set of numeric
attributes, then the data objects can be thought of as points (vectors) in a
multidimensional space, where each dimension represents a distinct attribute
describing the object. A set of such data objects can be interpreted as an m by
n matrix, where there are m rows, one for each object, and n columns, one for
each attribute. (A representation that has data objects as columns and
attributes as rows is also fine.) This matrix is called a data matrix or a pattern
matrix. A data matrix is a variation of record data, but because it consists of
numeric attributes, standard matrix operation can be applied to transform and
manipulate the data. Therefore, the data matrix is the standard data format for
most statistical data. Figure 2.2(c) shows a sample data matrix.

The Sparse Data Matrix

A sparse data matrix is a special case of a data matrix where the attributes
are of the same type and are asymmetric; i.e., only non-zero values are
important. Transaction data is an example of a sparse data matrix that has
only 0–1 entries. Another common example is document data. In particular, if
the order of the terms (words) in a document is ignored—the “bag of words”
approach—then a document can be represented as a term vector, where each
term is a component (attribute) of the vector and the value of each component
is the number of times the corresponding term occurs in the document. This
representation of a collection of documents is often called a document-term
matrix. Figure 2.2(d) shows a sample document-term matrix. The
documents are the rows of this matrix, while the terms are the columns. In
practice, only the non-zero entries of sparse data matrices are stored.

Graph-Based Data
A graph can sometimes be a convenient and powerful representation for data.
We consider two specific cases: (1) the graph captures relationships among
data objects and (2) the data objects themselves are represented as graphs.

Data with Relationships among Objects

The relationships among objects frequently convey important information. In
such cases, the data is often represented as a graph. In particular, the data
objects are mapped to nodes of the graph, while the relationships among
objects are captured by the links between objects and link properties, such as
direction and weight. Consider web pages on the World Wide Web, which
contain both text and links to other pages. In order to process search queries,
web search engines collect and process web pages to extract their contents. It
is well-known, however, that the links to and from each page provide a great
deal of information about the relevance of a web page to a query, and thus,
must also be taken into consideration. Figure 2.3(a) shows a set of linked

web pages. Another important example of such graph data are the social
networks, where data objects are people and the relationships among them
are their interactions via social media.

Data with Objects That Are Graphs

If objects have structure, that is, the objects contain subobjects that have
relationships, then such objects are frequently represented as graphs. For
example, the structure of chemical compounds can be represented by a
graph, where the nodes are atoms and the links between nodes are chemical
bonds. Figure 2.3(b) shows a ball-and-stick diagram of the chemical
compound benzene, which contains atoms of carbon (black) and hydrogen
(gray). A graph representation makes it possible to determine which
substructures occur frequently in a set of compounds and to ascertain
whether the presence of any of these substructures is associated with the
presence or absence of certain chemical properties, such as melting point or
heat of formation. Frequent graph mining, which is a branch of data mining
that analyzes such data, is considered in Section 6.5.

Figure 2.3.
Different variations of graph data.

Ordered Data
For some types of data, the attributes have relationships that involve order in
time or space. Different types of ordered data are described next and are
shown in Figure 2.4 .

Sequential Transaction Data

Sequential transaction data can be thought of as an extension of transaction
data, where each transaction has a time associated with it. Consider a retail
transaction data set that also stores the time at which the transaction took
place. This time information makes it possible to find patterns such as “candy
sales peak before Halloween.” A time can also be associated with each
attribute. For example, each record could be the purchase history of a

customer, with a listing of items purchased at different times. Using this
information, it is possible to find patterns such as “people who buy DVD
players tend to buy DVDs in the period immediately following the purchase.”

Figure 2.4(a) shows an example of sequential transaction data. There are
five different times—t1, t2, t3, t4, and t5; three different customers—C1, C2,
and C3; and five different items—A, B, C, D, and E. In the top table, each row
corresponds to the items purchased at a particular time by each customer. For
instance, at time t3, customer C2 purchased items A and D. In the bottom
table, the same information is displayed, but each row corresponds to a
particular customer. Each row contains information about each transaction
involving the customer, where a transaction is considered to be a set of items
and the time at which those items were purchased. For example, customer C3
bought items A and C at time t2.

Time Series Data

Time series data is a special type of ordered data where each record is a time
series , i.e., a series of measurements taken over time. For example, a
financial data set might contain objects that are time series of the daily prices
of various stocks. As another example, consider Figure 2.4(c) , which
shows a time series of the average monthly temperature for Minneapolis
during the years 1982 to 1994. When working with temporal data, such as
time series, it is important to consider temporal autocorrelation; i.e., if two
measurements are close in time, then the values of those measurements are
often very similar.

Figure 2.4.
Different variations of ordered data.

Sequence Data

Sequence data consists of a data set that is a sequence of individual entities,
such as a sequence of words or letters. It is quite similar to sequential data,
except that there are no time stamps; instead, there are positions in an
ordered sequence. For example, the genetic information of plants and animals
can be represented in the form of sequences of nucleotides that are known as
genes. Many of the problems associated with genetic sequence data involve
predicting similarities in the structure and function of genes from similarities in
nucleotide sequences. Figure 2.4(b) shows a section of the human genetic
code expressed using the four nucleotides from which all DNA is constructed:
A, T, G, and C.

Spatial and Spatio-Temporal Data

Some objects have spatial attributes, such as positions or areas, in addition to
other types of attributes. An example of spatial data is weather data
(precipitation, temperature, pressure) that is collected for a variety of
geographical locations. Often such measurements are collected over time,
and thus, the data consists of time series at various locations. In that case, we
refer to the data as spatio-temporal data. Although analysis can be conducted
separately for each specific time or location, a more complete analysis of
spatio-temporal data requires consideration of both the spatial and temporal
aspects of the data.

An important aspect of spatial data is spatial autocorrelation; i.e., objects
that are physically close tend to be similar in other ways as well. Thus, two
points on the Earth that are close to each other usually have similar values for
temperature and rainfall. Note that spatial autocorrelation is analogous to
temporal autocorrelation.

Important examples of spatial and spatio-temporal data are the science and
engineering data sets that are the result of measurements or model output

taken at regularly or irregularly distributed points on a two- or three-
dimensional grid or mesh. For instance, Earth science data sets record the
temperature or pressure measured at points (grid cells) on latitude–longitude
spherical grids of various resolutions, e.g., by See Figure 2.4(d) . As
another example, in the simulation of the flow of a gas, the speed and
direction of flow at various instants in time can be recorded for each grid point
in the simulation. A different type of spatio-temporal data arises from tracking
the trajectories of objects, e.g., vehicles, in time and space.

Handling Non-Record Data
Most data mining algorithms are designed for record data or its variations,
such as transaction data and data matrices. Record-oriented techniques can
be applied to non-record data by extracting features from data objects and
using these features to create a record corresponding to each object.
Consider the chemical structure data that was described earlier. Given a set of
common substructures, each compound can be represented as a record with
binary attributes that indicate whether a compound contains a specific
substructure. Such a representation is actually a transaction data set, where
the transactions are the compounds and the items are the substructures.

In some cases, it is easy to represent the data in a record format, but this type
of representation does not capture all the information in the data. Consider
spatio-temporal data consisting of a time series from each point on a spatial
grid. This data is often stored in a data matrix, where each row represents a
location and each column represents a particular point in time. However, such
a representation does not explicitly capture the time relationships that are
present among attributes and the spatial relationships that exist among
objects. This does not mean that such a representation is inappropriate, but
rather that these relationships must be taken into consideration during the
analysis. For example, it would not be a good idea to use a data mining

1° 1°.

technique that ignores the temporal autocorrelation of the attributes or the
spatial autocorrelation of the data objects, i.e., the locations on the spatial

2.2 Data Quality
Data mining algorithms are often applied to data that was collected for another
purpose, or for future, but unspecified applications. For that reason, data
mining cannot usually take advantage of the significant benefits of “ad-
dressing quality issues at the source.” In contrast, much of statistics deals with
the design of experiments or surveys that achieve a prespecified level of data
quality. Because preventing data quality problems is typically not an option,
data mining focuses on (1) the detection and correction of data quality
problems and (2) the use of algorithms that can tolerate poor data quality. The
first step, detection and correction, is often called data cleaning.

The following sections discuss specific aspects of data quality. The focus is on
measurement and data collection issues, although some application-related
issues are also discussed.

2.2.1 Measurement and Data
Collection Issues

It is unrealistic to expect that data will be perfect. There may be problems due
to human error, limitations of measuring devices, or flaws in the data collection
process. Values or even entire data objects can be missing. In other cases,
there can be spurious or duplicate objects; i.e., multiple data objects that all
correspond to a single “real” object. For example, there might be two different
records for a person who has recently lived at two different addresses. Even if

all the data is present and “looks fine,” there may be inconsistencies—a
person has a height of 2 meters, but weighs only 2 kilograms.

In the next few sections, we focus on aspects of data quality that are related
to data measurement and collection. We begin with a definition of
measurement and data collection errors and then consider a variety of
problems that involve measurement error: noise, artifacts, bias, precision, and
accuracy. We conclude by discussing data quality issues that involve both
measurement and data collection problems: outliers, missing and inconsistent
values, and duplicate data.

Measurement and Data Collection Errors
The term measurement error refers to any problem resulting from the
measurement process. A common problem is that the value recorded differs
from the true value to some extent. For continuous attributes, the numerical
difference of the measured and true value is called the error. The term data
collection error refers to errors such as omitting data objects or attribute
values, or inappropriately including a data object. For example, a study of
animals of a certain species might include animals of a related species that
are similar in appearance to the species of interest. Both measurement errors
and data collection errors can be either systematic or random.

We will only consider general types of errors. Within particular domains,
certain types of data errors are commonplace, and well-developed techniques
often exist for detecting and/or correcting these errors. For example, keyboard
errors are common when data is entered manually, and as a result, many data
entry programs have techniques for detecting and, with human intervention,
correcting such errors.

Noise and Artifacts
Noise is the random component of a measurement error. It typically involves
the distortion of a value or the addition of spurious objects. Figure 2.5
shows a time series before and after it has been disrupted by random noise. If
a bit more noise were added to the time series, its shape would be lost.
Figure 2.6 shows a set of data points before and after some noise points
(indicated by ) have been added. Notice that some of the noise points are
intermixed with the non-noise points.

Figure 2.5.
Noise in a time series context.


Figure 2.6.
Noise in a spatial context.

The term noise is often used in connection with data that has a spatial or
temporal component. In such cases, techniques from signal or image
processing can frequently be used to reduce noise and thus, help to discover
patterns (signals) that might be “lost in the noise.” Nonetheless, the
elimination of noise is frequently difficult, and much work in data mining
focuses on devising robust algorithms that produce acceptable results even
when noise is present.

Data errors can be the result of a more deterministic phenomenon, such as a
streak in the same place on a set of photographs. Such deterministic
distortions of the data are often referred to as artifacts.

Precision, Bias, and Accuracy
In statistics and experimental science, the quality of the measurement process
and the resulting data are measured by precision and bias. We provide the

standard definitions, followed by a brief discussion. For the following
definitions, we assume that we make repeated measurements of the same
underlying quantity.

Definition 2.3 (Precision).
The closeness of repeated measurements (of the same
quantity) to one another.

Definition 2.4 (Bias).
A systematic variation of measurements from the quantity being

Precision is often measured by the standard deviation of a set of values, while
bias is measured by taking the difference between the mean of the set of
values and the known value of the quantity being measured. Bias can be
determined only for objects whose measured quantity is known by means
external to the current situation. Suppose that we have a standard laboratory
weight with a mass of 1g and want to assess the precision and bias of our
new laboratory scale. We weigh the mass five times, and obtain the following
five values:{ 1.015, 0.990, 1.013, 1.001, 0.986}. The mean of these values is

1.001, and hence, the bias is 0.001. The precision, as measured by the
standard deviation, is 0.013.

It is common to use the more general term, accuracy , to refer to the degree
of measurement error in data.

Definition 2.5 (Accuracy)
The closeness of measurements to the true value of the quantity
being measured.

Accuracy depends on precision and bias, but there is no specific formula for
accuracy in terms of these two quantities.

One important aspect of accuracy is the use of significant digits. The goal is
to use only as many digits to represent the result of a measurement or
calculation as are justified by the precision of the data. For example, if the
length of an object is measured with a meter stick whose smallest markings
are millimeters, then we should record the length of data only to the nearest
millimeter. The precision of such a measurement would be We do
not review the details of working with significant digits because most readers
will have encountered them in previous courses and they are covered in
considerable depth in science, engineering, and statistics textbooks.

Issues such as significant digits, precision, bias, and accuracy are sometimes
overlooked, but they are important for data mining as well as statistics and
science. Many times, data sets do not come with information about the

± 0.5mm.

precision of the data, and furthermore, the programs used for analysis return
results without any such information. Nonetheless, without some
understanding of the accuracy of the data and the results, an analyst runs the
risk of committing serious data analysis blunders.

Outliers are either (1) data objects that, in some sense, have characteristics
that are different from most of the other data objects in the data set, or (2)
values of an attribute that are unusual with respect to the typical values for
that attribute. Alternatively, they can be referred to as anomalous objects or
values. There is considerable leeway in the definition of an outlier, and many
different definitions have been proposed by the statistics and data mining
communities. Furthermore, it is important to distinguish between the notions of
noise and outliers. Unlike noise, outliers can be legitimate data objects or
values that we are interested in detecting. For instance, in fraud and network
intrusion detection, the goal is to find unusual objects or events from among a
large number of normal ones. Chapter 9 discusses anomaly detection in
more detail.

Missing Values
It is not unusual for an object to be missing one or more attribute values. In
some cases, the information was not collected; e.g., some people decline to
give their age or weight. In other cases, some attributes are not applicable to
all objects; e.g., often, forms have conditional parts that are filled out only
when a person answers a previous question in a certain way, but for simplicity,
all fields are stored. Regardless, missing values should be taken into account
during the data analysis.

There are several strategies (and variations on these strategies) for dealing
with missing data, each of which is appropriate in certain circumstances.
These strategies are listed next, along with an indication of their advantages
and disadvantages.

Eliminate Data Objects or Attributes

A simple and effective strategy is to eliminate objects with missing values.
However, even a partially specified data object contains some information,
and if many objects have missing values, then a reliable analysis can be
difficult or impossible. Nonetheless, if a data set has only a few objects that
have missing values, then it may be expedient to omit them. A related strategy
is to eliminate attributes that have missing values. This should be done with
caution, however, because the eliminated attributes may be the ones that are
critical to the analysis.

Estimate Missing Values

Sometimes missing data can be reliably estimated. For example, consider a
time series that changes in a reasonably smooth fashion, but has a few,
widely scattered missing values. In such cases, the missing values can be
estimated (interpolated) by using the remaining values. As another example,
consider a data set that has many similar data points. In this situation, the
attribute values of the points closest to the point with the missing value are
often used to estimate the missing value. If the attribute is continuous, then
the average attribute value of the nearest neighbors is used; if the attribute is
categorical, then the most commonly occurring attribute value can be taken.
For a concrete illustration, consider precipitation measurements that are
recorded by ground stations. For areas not containing a ground station, the
precipitation can be estimated using values observed at nearby ground

Ignore the Missing Value during Analysis

Many data mining approaches can be modified to ignore missing values. For
example, suppose that objects are being clustered and the similarity between
pairs of data objects needs to be calculated. If one or both objects of a pair
have missing values for some attributes, then the similarity can be calculated
by using only the attributes that do not have missing values. It is true that the
similarity will only be approximate, but unless the total number of attributes is
small or the number of missing values is high, this degree of inaccuracy may
not matter much. Likewise, many classification schemes can be modified to
work with missing values.

Inconsistent Values
Data can contain inconsistent values. Consider an address field, where both a
zip code and city are listed, but the specified zip code area is not contained in
that city. It is possible that the individual entering this information transposed
two digits, or perhaps a digit was misread when the information was scanned
from a handwritten form. Regardless of the cause of the inconsistent values, it
is important to detect and, if possible, correct such problems.

Some types of inconsistences are easy to detect. For instance, a person’s
height should not be negative. In other cases, it can be necessary to consult
an external source of information. For example, when an insurance company
processes claims for reimbursement, it checks the names and addresses on
the reimbursement forms against a database of its customers.

Once an inconsistency has been detected, it is sometimes possible to correct
the data. A product code may have “check” digits, or it may be possible to
double-check a product code against a list of known product codes, and then

correct the code if it is incorrect, but close to a known code. The correction of
an inconsistency requires additional or redundant information.

Example 2.6 (Inconsistent Sea Surface
This example illustrates an inconsistency in actual time series data that
measures the sea surface temperature (SST) at various points on the
ocean. SST data was originally collected using ocean-based
measurements from ships or buoys, but more recently, satellites have
been used to gather the data. To create a long-term data set, both sources
of data must be used. However, because the data comes from different
sources, the two parts of the data are subtly different. This discrepancy is
visually displayed in Figure 2.7 , which shows the correlation of SST
values between pairs of years. If a pair of years has a positive correlation,
then the location corresponding to the pair of years is colored white;
otherwise it is colored black. (Seasonal variations were removed from the
data since, otherwise, all the years would be highly correlated.) There is a
distinct change in behavior where the data has been put together in 1983.
Years within each of the two groups, 1958–1982 and 1983–1999, tend to
have a positive correlation with one another, but a negative correlation with
years in the other group. This does not mean that this data should not be
used, only that the analyst should consider the potential impact of such
discrepancies on the data mining analysis.

Figure 2.7.
Correlation of SST data between pairs of years. White areas indicate
positive correlation. Black areas indicate negative correlation.

Duplicate Data
A data set can include data objects that are duplicates, or almost duplicates,
of one another. Many people receive duplicate mailings because they appear
in a database multiple times under slightly different names. To detect and
eliminate such duplicates, two main issues must be addressed. First, if there
are two objects that actually represent a single object, then one or more
values of corresponding attributes are usually different, and these inconsistent
values must be resolved. Second, care needs to be taken to avoid
accidentally combining data objects that are similar, but not duplicates, such

as two distinct people with identical names. The term deduplication is often
used to refer to the process of dealing with these issues.

In some cases, two or more objects are identical with respect to the attributes
measured by the database, but they still represent different objects. Here, the
duplicates are legitimate, but can still cause problems for some algorithms if
the possibility of identical objects is not specifically accounted for in their
design. An example of this is given in Exercise 13 on page 108.

2.2.2 Issues Related to Applications

Data quality issues can also be considered from an application viewpoint as
expressed by the statement “data is of high quality if it is suitable for its
intended use.” This approach to data quality has proven quite useful,
particularly in business and industry. A similar viewpoint is also present in
statistics and the experimental sciences, with their emphasis on the careful
design of experiments to collect the data relevant to a specific hypothesis. As
with quality issues at the measurement and data collection level, many issues
are specific to particular applications and fields. Again, we consider only a few
of the general issues.


Some data starts to age as soon as it has been collected. In particular, if the
data provides a snapshot of some ongoing phenomenon or process, such as
the purchasing behavior of customers or web browsing patterns, then this
snapshot represents reality for only a limited time. If the data is out of date,
then so are the models and patterns that are based on it.


The available data must contain the information necessary for the application.
Consider the task of building a model that predicts the accident rate for
drivers. If information about the age and gender of the driver is omitted, then it
is likely that the model will have limited accuracy unless this information is
indirectly available through other attributes.

Making sure that the objects in a data set are relevant is also challenging. A
common problem is sampling bias, which occurs when a sample does not
contain different types of objects in proportion to their actual occurrence in the
population. For example, survey data describes only those who respond to the
survey. (Other aspects of sampling are discussed further in Section 2.3.2 .)
Because the results of a data analysis can reflect only the data that is present,
sampling bias will typically lead to erroneous results when applied to the
broader population.

Knowledge about the Data

Ideally, data sets are accompanied by documentation that describes different
aspects of the data; the quality of this documentation can either aid or hinder
the subsequent analysis. For example, if the documentation identifies several
attributes as being strongly related, these attributes are likely to provide highly
redundant information, and we usually decide to keep just one. (Consider
sales tax and purchase price.) If the documentation is poor, however, and fails
to tell us, for example, that the missing values for a particular field are
indicated with a -9999, then our analysis of the data may be faulty. Other
important characteristics are the precision of the data, the type of features
(nominal, ordinal, interval, ratio), the scale of measurement (e.g., meters or
feet for length), and the origin of the data.

2.3 Data Preprocessing
In this section, we consider which preprocessing steps should be applied to
make the data more suitable for data mining. Data preprocessing is a broad
area and consists of a number of different strategies and techniques that are
interrelated in complex ways. We will present some of the most important
ideas and approaches, and try to point out the interrelationships among them.
Specifically, we will discuss the following topics:

Dimensionality reduction
Feature subset selection
Feature creation
Discretization and binarization
Variable transformation

Roughly speaking, these topics fall into two categories: selecting data objects
and attributes for the analysis or for creating/changing the attributes. In both
cases, the goal is to improve the data mining analysis with respect to time,
cost, and quality. Details are provided in the following sections.

A quick note about terminology: In the following, we sometimes use synonyms
for attribute, such as feature or variable, in order to follow common usage.

2.3.1 Aggregation

Sometimes “less is more,” and this is the case with aggregation , the
combining of two or more objects into a single object. Consider a data set
consisting of transactions (data objects) recording the daily sales of products
in various store locations (Minneapolis, Chicago, Paris, …) for different days
over the course of a year. See Table 2.4 . One way to aggregate
transactions for this data set is to replace all the transactions of a single store
with a single storewide transaction. This reduces the hundreds or thousands
of transactions that occur daily at a specific store to a single daily transaction,
and the number of data objects per day is reduced to the number of stores.

Table 2.4. Data set containing information about customer purchases.

Transaction ID Item Store Location Date Price …

⋮ ⋮ ⋮ ⋮ ⋮

101123 Watch Chicago 09/06/04 $25.99 …

101123 Battery Chicago 09/06/04 $5.99 …

101124 Shoes Minneapolis 09/06/04 $75.00 …

An obvious issue is how an aggregate transaction is created; i.e., how the
values of each attribute are combined across all the records corresponding to
a particular location to create the aggregate transaction that represents the
sales of a single store or date. Quantitative attributes, such as price, are
typically aggregated by taking a sum or an average. A qualitative attribute,
such as item, can either be omitted or summarized in terms of a higher level
category, e.g., televisions versus electronics.

The data in Table 2.4 can also be viewed as a multidimensional array,
where each attribute is a dimension. From this viewpoint, aggregation is the
process of eliminating attributes, such as the type of item, or reducing the

number of values for a particular attribute; e.g., reducing the possible values
for date from 365 days to 12 months. This type of aggregation is commonly
used in Online Analytical Processing (OLAP). References to OLAP are given
in the bibliographic Notes.

There are several motivations for aggregation. First, the smaller data sets
resulting from data reduction require less memory and processing time, and
hence, aggregation often enables the use of more expensive data mining
algorithms. Second, aggregation can act as a change of scope or scale by
providing a high-level view of the data instead of a low-level view. In the
previous example, aggregating over store locations and months gives us a
monthly, per store view of the data instead of a daily, per item view. Finally, the
behavior of groups of objects or attributes is often more stable than that of
individual objects or attributes. This statement reflects the statistical fact that
aggregate quantities, such as averages or totals, have less variability than the
individual values being aggregated. For totals, the actual amount of variation
is larger than that of individual objects (on average), but the percentage of the
variation is smaller, while for means, the actual amount of variation is less
than that of individual objects (on average). A disadvantage of aggregation is
the potential loss of interesting details. In the store example, aggregating over
months loses information about which day of the week has the highest sales.

Example 2.7 (Australian Precipitation).
This example is based on precipitation in Australia from the period 1982–
1993. Figure 2.8(a) shows a histogram for the standard deviation of
average monthly precipitation for by grid cells in Australia,
while Figure 2.8(b) shows a histogram for the standard deviation of the
average yearly precipitation for the same locations. The average yearly
precipitation has less variability than the average monthly precipitation. All

3,030 0.5° 0.5°

precipitation measurements (and their standard deviations) are in

Figure 2.8.
Histograms of standard deviation for monthly and yearly precipitation in
Australia for the period 1982–1993.

2.3.2 Sampling

Sampling is a commonly used approach for selecting a subset of the data
objects to be analyzed. In statistics, it has long been used for both the
preliminary investigation of the data and the final data analysis. Sampling can
also be very useful in data mining. However, the motivations for sampling in
statistics and data mining are often different. Statisticians use sampling
because obtaining the entire set of data of interest is too expensive or time
consuming, while data miners usually sample because it is too
computationally expensive in terms of the memory or time required to process

all the data. In some cases, using a sampling algorithm can reduce the data
size to the point where a better, but more computationally expensive algorithm
can be used.

The key principle for effective sampling is the following: Using a sample will
work almost as well as using the entire data set if the sample is
representative. In turn, a sample is representative if it has approximately the
same property (of interest) as the original set of data. If the mean (average) of
the data objects is the property of interest, then a sample is representative if it
has a mean that is close to that of the original data. Because sampling is a
statistical process, the representativeness of any particular sample will vary,
and the best that we can do is choose a sampling scheme that guarantees a
high probability of getting a representative sample. As discussed next, this
involves choosing the appropriate sample size and sampling technique.

Sampling Approaches
There are many sampling techniques, but only a few of the most basic ones
and their variations will be covered here. The simplest type of sampling is
simple random sampling. For this type of sampling, there is an equal
probability of selecting any particular object. There are two variations on
random sampling (and other sampling techniques as well): (1) sampling
without replacement —as each object is selected, it is removed from the set
of all objects that together constitute the population , and (2) sampling with
replacement —objects are not removed from the population as they are
selected for the sample. In sampling with replacement, the same object can
be picked more than once. The samples produced by the two methods are not
much different when samples are relatively small compared to the data set
size, but sampling with replacement is simpler to analyze because the
probability of selecting any object remains constant during the sampling

When the population consists of different types of objects, with widely different
numbers of objects, simple random sampling can fail to adequately represent
those types of objects that are less frequent. This can cause problems when
the analysis requires proper representation of all object types. For example,
when building classification models for rare classes, it is critical that the rare
classes be adequately represented in the sample. Hence, a sampling scheme
that can accommodate differing frequencies for the object types of interest is
needed. Stratified sampling , which starts with prespecified groups of
objects, is such an approach. In the simplest version, equal numbers of
objects are drawn from each group even though the groups are of different
sizes. In another variation, the number of objects drawn from each group is
proportional to the size of that group.

Example 2.8 (Sampling and Loss of Information).
Once a sampling technique has been selected, it is still necessary to
choose the sample size. Larger sample sizes increase the probability that
a sample will be representative, but they also eliminate much of the
advantage of sampling. Conversely, with smaller sample sizes, patterns
can be missed or erroneous patterns can be detected. Figure 2.9(a)
shows a data set that contains 8000 two-dimensional points, while Figures
2.9(b) and 2.9(c) show samples from this data set of size 2000 and
500, respectively. Although most of the structure of this data set is present
in the sample of 2000 points, much of the structure is missing in the
sample of 500 points.

Figure 2.9.
Example of the loss of structure with sampling.

Example 2.9 (Determining the Proper Sample
To illustrate that determining the proper sample size requires a methodical
approach, consider the following task.

Given a set of data consisting of a small number of almost equalsized groups, find at least one

representative point for each of the groups. Assume that the objects in each group are highly

similar to each other, but not very similar to objects in different groups. Figure 2.10(a) shows

an idealized set of clusters (groups) from which these points might be drawn.

Figure 2.10.
Finding representative points from 10 groups.

This problem can be efficiently solved using sampling. One approach is to
take a small sample of data points, compute the pairwise similarities
between points, and then form groups of points that are highly similar. The
desired set of representative points is then obtained by taking one point
from each of these groups. To follow this approach, however, we need to
determine a sample size that would guarantee, with a high probability, the
desired outcome; that is, that at least one point will be obtained from each
cluster. Figure 2.10(b) shows the probability of getting one object from
each of the 10 groups as the sample size runs from 10 to 60. Interestingly,
with a sample size of 20, there is little chance (20%) of getting a sample
that includes all 10 clusters. Even with a sample size of 30, there is still a
moderate chance (almost 40%) of getting a sample that doesn’t contain
objects from all 10 clusters. This issue is further explored in the context of
clustering by Exercise 4 on page 603.

Progressive Sampling
The proper sample size can be difficult to determine, so adaptive or
progressive sampling schemes are sometimes used. These approaches
start with a small sample, and then increase the sample size until a sample of
sufficient size has been obtained. While this technique eliminates the need to
determine the correct sample size initially, it requires that there be a way to
evaluate the sample to judge if it is large enough.

Suppose, for instance, that progressive sampling is used to learn a predictive
model. Although the accuracy of predictive models increases as the sample
size increases, at some point the increase in accuracy levels off. We want to
stop increasing the sample size at this leveling-off point. By keeping track of
the change in accuracy of the model as we take progressively larger samples,
and by taking other samples close to the size of the current one, we can get
an estimate of how close we are to this leveling-off point, and thus, stop

2.3.3 Dimensionality Reduction

Data sets can have a large number of features. Consider a set of documents,
where each document is represented by a vector whose components are the
frequencies with which each word occurs in the document. In such cases,
there are typically thousands or tens of thousands of attributes (components),
one for each word in the vocabulary. As another example, consider a set of
time series consisting of the daily closing price of various stocks over a period
of 30 years. In this case, the attributes, which are the prices on specific days,
again number in the thousands.

There are a variety of benefits to dimensionality reduction. A key benefit is that
many data mining algorithms work better if the dimensionality—the number of
attributes in the data—is lower. This is partly because dimensionality reduction
can eliminate irrelevant features and reduce noise and partly because of the
curse of dimensionality, which is explained below. Another benefit is that a
reduction of dimensionality can lead to a more understandable model because
the model usually involves fewer attributes. Also, dimensionality reduction
may allow the data to be more easily visualized. Even if dimensionality
reduction doesn’t reduce the data to two or three dimensions, data is often
visualized by looking at pairs or triplets of attributes, and the number of such
combinations is greatly reduced. Finally, the amount of time and memory
required by the data mining algorithm is reduced with a reduction in

The term dimensionality reduction is often reserved for those techniques that
reduce the dimensionality of a data set by creating new attributes that are a
combination of the old attributes. The reduction of dimensionality by selecting
attributes that are a subset of the old is known as feature subset selection or
feature selection. It will be discussed in Section 2.3.4 .

In the remainder of this section, we briefly introduce two important topics: the
curse of dimensionality and dimensionality reduction techniques based on
linear algebra approaches such as principal components analysis (PCA).
More details on dimensionality reduction can be found in Appendix B.

The Curse of Dimensionality
The curse of dimensionality refers to the phenomenon that many types of data
analysis become significantly harder as the dimensionality of the data
increases. Specifically, as dimensionality increases, the data becomes
increasingly sparse in the space that it occupies. Thus, the data objects we

observe are quite possibly not a representative sample of all possible objects.
For classification, this can mean that there are not enough data objects to
allow the creation of a model that reliably assigns a class to all possible
objects. For clustering, the differences in density and in the distances between
points, which are critical for clustering, become less meaningful. (This is
discussed further in Sections 8.1.2, 8.4.6, and 8.4.8.) As a result, many
clustering and classification algorithms (and other data analysis algorithms)
have trouble with high-dimensional data leading to reduced classification
accuracy and poor quality clusters.

Linear Algebra Techniques for Dimensionality
Some of the most common approaches for dimensionality reduction,
particularly for continuous data, use techniques from linear algebra to project
the data from a high-dimensional space into a lower-dimensional space.
Principal Components Analysis (PCA) is a linear algebra technique for
continuous attributes that finds new attributes (principal components) that (1)
are linear combinations of the original attributes, (2) are orthogonal
(perpendicular) to each other, and (3) capture the maximum amount of
variation in the data. For example, the first two principal components capture
as much of the variation in the data as is possible with two orthogonal
attributes that are linear combinations of the original attributes. Singular
Value Decomposition (SVD) is a linear algebra technique that is related to
PCA and is also commonly used for dimensionality reduction. For additional
details, see Appendices A and B.

2.3.4 Feature Subset Selection

Another way to reduce the dimensionality is to use only a subset of the
features. While it might seem that such an approach would lose information,
this is not the case if redundant and irrelevant features are present.
Redundant features duplicate much or all of the information contained in one
or more other attributes. For example, the purchase price of a product and the
amount of sales tax paid contain much of the same information. Irrelevant
features contain almost no useful information for the data mining task at
hand. For instance, students’ ID numbers are irrelevant to the task of
predicting students’ grade point averages. Redundant and irrelevant features
can reduce classification accuracy and the quality of the clusters that are

While some irrelevant and redundant attributes can be eliminated immediately
by using common sense or domain knowledge, selecting the best subset of
features frequently requires a systematic approach. The ideal approach to
feature selection is to try all possible subsets of features as input to the data
mining algorithm of interest, and then take the subset that produces the best
results. This method has the advantage of reflecting the objective and bias of
the data mining algorithm that will eventually be used. Unfortunately, since the
number of subsets involving n attributes is 2 , such an approach is impractical

in most situations and alternative strategies are needed. There are three
standard approaches to feature selection: embedded, filter, and wrapper.

Embedded approaches

Feature selection occurs naturally as part of the data mining algorithm.
Specifically, during the operation of the data mining algorithm, the algorithm
itself decides which attributes to use and which to ignore. Algorithms for
building decision tree classifiers, which are discussed in Chapter 3 , often
operate in this manner.


Filter approaches

Features are selected before the data mining algorithm is run, using some
approach that is independent of the data mining task. For example, we might
select sets of attributes whose pairwise correlation is as low as possible so
that the attributes are non-redundant.

Wrapper approaches

These methods use the target data mining algorithm as a black box to find the
best subset of attributes, in a way similar to that of the ideal algorithm
described above, but typically without enumerating all possible subsets.

Because the embedded approaches are algorithm-specific, only the filter and
wrapper approaches will be discussed further here.

An Architecture for Feature Subset Selection
It is possible to encompass both the filter and wrapper approaches within a
common architecture. The feature selection process is viewed as consisting of
four parts: a measure for evaluating a subset, a search strategy that controls
the generation of a new subset of features, a stopping criterion, and a
validation procedure. Filter methods and wrapper methods differ only in the
way in which they evaluate a subset of features. For a wrapper method,
subset evaluation uses the target data mining algorithm, while for a filter
approach, the evaluation technique is distinct from the target data mining
algorithm. The following discussion provides some details of this approach,
which is summarized in Figure 2.11 .

Figure 2.11.
Flowchart of a feature subset selection process.

Conceptually, feature subset selection is a search over all possible subsets of
features. Many different types of search strategies can be used, but the
search strategy should be computationally inexpensive and should find
optimal or near optimal sets of features. It is usually not possible to satisfy
both requirements, and thus, trade-offs are necessary.

An integral part of the search is an evaluation step to judge how the current
subset of features compares to others that have been considered. This
requires an evaluation measure that attempts to determine the goodness of a
subset of attributes with respect to a particular data mining task, such as
classification or clustering. For the filter approach, such measures attempt to
predict how well the actual data mining algorithm will perform on a given set of
attributes. For the wrapper approach, where evaluation consists of actually
running the target data mining algorithm, the subset evaluation function is
simply the criterion normally used to measure the result of the data mining.

Because the number of subsets can be enormous and it is impractical to
examine them all, some sort of stopping criterion is necessary. This strategy is
usually based on one or more conditions involving the following: the number
of iterations, whether the value of the subset evaluation measure is optimal or
exceeds a certain threshold, whether a subset of a certain size has been
obtained, and whether any improvement can be achieved by the options
available to the search strategy.

Finally, once a subset of features has been selected, the results of the target
data mining algorithm on the selected subset should be validated. A
straightforward validation approach is to run the algorithm with the full set of
features and compare the full results to results obtained using the subset of
features. Hopefully, the subset of features will produce results that are better
than or almost as good as those produced when using all features. Another
validation approach is to use a number of different feature selection
algorithms to obtain subsets of features and then compare the results of
running the data mining algorithm on each subset.

Feature Weighting
Feature weighting is an alternative to keeping or eliminating features. More
important features are assigned a higher weight, while less important features
are given a lower weight. These weights are sometimes assigned based on
domain knowledge about the relative importance of features. Alternatively,
they can sometimes be determined automatically. For example, some
classification schemes, such as support vector machines (Chapter 4 ),
produce classification models in which each feature is given a weight.
Features with larger weights play a more important role in the model. The
normalization of objects that takes place when computing the cosine similarity
(Section 2.4.5 ) can also be regarded as a type of feature weighting.

2.3.5 Feature Creation

It is frequently possible to create, from the original attributes, a new set of
attributes that captures the important information in a data set much more
effectively. Furthermore, the number of new attributes can be smaller than the
number of original attributes, allowing us to reap all the previously described
benefits of dimensionality reduction. Two related methodologies for creating
new attributes are described next: feature extraction and mapping the data to
a new space.

Feature Extraction
The creation of a new set of features from the original raw data is known as
feature extraction. Consider a set of photographs, where each photograph is
to be classified according to whether it contains a human face. The raw data
is a set of pixels, and as such, is not suitable for many types of classification
algorithms. However, if the data is processed to provide higher-level features,
such as the presence or absence of certain types of edges and areas that are
highly correlated with the presence of human faces, then a much broader set
of classification techniques can be applied to this problem.

Unfortunately, in the sense in which it is most commonly used, feature
extraction is highly domain-specific. For a particular field, such as image
processing, various features and the techniques to extract them have been
developed over a period of time, and often these techniques have limited
applicability to other fields. Consequently, whenever data mining is applied to
a relatively new area, a key task is the development of new features and
feature extraction methods.

Although feature extraction is often complicated, Example 2.10 illustrates
that it can be relatively straightforward.

Example 2.10 (Density).
Consider a data set consisting of information about historical artifacts,
which, along with other information, contains the volume and mass of each
artifact. For simplicity, assume that these artifacts are made of a small
number of materials (wood, clay, bronze, gold) and that we want to classify
the artifacts with respect to the material of which they are made. In this
case, a density feature constructed from the mass and volume features,
i.e., density =mass/volume , would most directly yield an accurate
classification. Although there have been some attempts to automatically
perform such simple feature extraction by exploring basic mathematical
combinations of existing attributes, the most common approach is to
construct features using domain expertise.

Mapping the Data to a New Space
A totally different view of the data can reveal important and interesting
features. Consider, for example, time series data, which often contains
periodic patterns. If there is only a single periodic pattern and not much noise,
then the pattern is easily detected. If, on the other hand, there are a number of
periodic patterns and a significant amount of noise, then these patterns are
hard to detect. Such patterns can, nonetheless, often be detected by applying
a Fourier transform to the time series in order to change to a representation
in which frequency information is explicit. In Example 2.11 , it will not be
necessary to know the details of the Fourier transform. It is enough to know
that, for each time series, the Fourier transform produces a new data object
whose attributes are related to frequencies.

Example 2.11 (Fourier Analysis).
The time series presented in Figure 2.12(b) is the sum of three other
time series, two of which are shown in Figure 2.12(a) and have
frequencies of 7 and 17 cycles per second, respectively. The third time
series is random noise. Figure 2.12(c) shows the power spectrum that
can be computed after applying a Fourier transform to the original time
series. (Informally, the power spectrum is proportional to the square of
each frequency attribute.) In spite of the noise, there are two peaks that
correspond to the periods of the two original, non-noisy time series. Again,
the main point is that better features can reveal important aspects of the

Figure 2.12.
Application of the Fourier transform to identify the underlying frequencies
in time series data.

Many other sorts of transformations are also possible. Besides the Fourier
transform, the wavelet transform has also proven very useful for time series
and other types of data.

2.3.6 Discretization and Binarization

Some data mining algorithms, especially certain classification algorithms,
require that the data be in the form of categorical attributes. Algorithms that
find association patterns require that the data be in the form of binary
attributes. Thus, it is often necessary to transform a continuous attribute into a
categorical attribute (discretization), and both continuous and discrete
attributes may need to be transformed into one or more binary attributes
(binarization). Additionally, if a categorical attribute has a large number of
values (categories), or some values occur infrequently, then it can be
beneficial for certain data mining tasks to reduce the number of categories by
combining some of the values.

As with feature selection, the best discretization or binarization approach is
the one that “produces the best result for the data mining algorithm that will be
used to analyze the data.” It is typically not practical to apply such a criterion
directly. Consequently, discretization or binarization is performed in a way that
satisfies a criterion that is thought to have a relationship to good performance
for the data mining task being considered. In general, the best discretization
depends on the algorithm being used, as well as the other attributes being
considered. Typically, however, the discretization of each attribute is
considered in isolation.

A simple technique to binarize a categorical attribute is the following: If there
are m categorical values, then uniquely assign each original value to an
integer in the interval If the attribute is ordinal, then order must be
maintained by the assignment. (Note that even if the attribute is originally
represented using integers, this process is necessary if the integers are not in

[0, m−1].

the interval ) Next, convert each of these m integers to a binary
number. Since binary digits are required to represent these
integers, represent these binary numbers using n binary attributes. To
illustrate, a categorical variable with 5 values {awful, poor, OK, good, great}
would require three binary variables and The conversion is shown
in Table 2.5 .

Table 2.5. Conversion of a categorical attribute to three binary attributes.

Categorical Value Integer Value

awful 0 0 0 0

poor 1 0 0 1

OK 2 0 1 0

good 3 0 1 1

great 4 1 0 0

Such a transformation can cause complications, such as creating unintended
relationships among the transformed attributes. For example, in Table 2.5 ,
attributes and are correlated because information about the good value
is encoded using both attributes. Furthermore, association analysis requires
asymmetric binary attributes, where only the presence of the attribute

is important. For association problems, it is therefore necessary to
introduce one asymmetric binary attribute for each categorical value, as
shown in Table 2.6 . If the number of resulting attributes is too large, then
the techniques described in the following sections can be used to reduce the
number of categorical values before binarization.

Table 2.6. Conversion of a categorical attribute to five asymmetric binary

[0, m−1].

x1, x2, x3.

x1 x2 x3

x2 x3

(value =1).


Categorical Value Integer Value

awful 0 1 0 0 0 0

poor 1 0 1 0 0 0

OK 2 0 0 1 0 0

good 3 0 0 0 1 0

great 4 0 0 0 0 1

Likewise, for association problems, it can be necessary to replace a single
binary attribute with two asymmetric binary attributes. Consider a binary
attribute that records a person’s gender, male or female. For traditional
association rule algorithms, this information needs to be transformed into two
asymmetric binary attributes, one that is a 1 only when the person is male and
one that is a 1 only when the person is female. (For asymmetric binary
attributes, the information representation is somewhat inefficient in that two
bits of storage are required to represent each bit of information.)

Discretization of Continuous Attributes
Discretization is typically applied to attributes that are used in classification or
association analysis. Transformation of a continuous attribute to a categorical
attribute involves two subtasks: deciding how many categories,n , to have and
determining how to map the values of the continuous attribute to these
categories. In the first step, after the values of the continuous attribute are
sorted, they are then divided into n intervals by specifying split points. In
the second, rather trivial step, all the values in one interval are mapped to the
same categorical value. Therefore, the problem of discretization is one of

x1 x2 x3 x4 x5


deciding how many split points to choose and where to place them. The result
can be represented either as a set of intervals

where and can be or respectively, or equivalently, as a
series of inequalities

Unsupervised Discretization

A basic distinction between discretization methods for classification is whether
class information is used (supervised) or not (unsupervised). If class
information is not used, then relatively simple approaches are common. For
instance, the equal width approach divides the range of the attribute into a
user-specified number of intervals each having the same width. Such an
approach can be badly affected by outliers, and for that reason, an equal
frequency (equal depth) approach, which tries to put the same number of
objects into each interval, is often preferred. As another example of
unsupervised discretization, a clustering method, such as K-means (see
Chapter 7 ), can also be used. Finally, visually inspecting the data can
sometimes be an effective approach.

Example 2.12 (Discretization Techniques).
This example demonstrates how these approaches work on an actual data
set. Figure 2.13(a) shows data points belonging to four different groups,
along with two outliers—the large dots on either end. The techniques of the
previous paragraph were applied to discretize the x values of these data
points into four categorical values. (Points in the data set have a random y
component to make it easy to see how many points are in each group.)
Visually inspecting the data works quite well, but is not automatic, and
thus, we focus on the other three approaches. The split points produced by
the techniques equal width, equal frequency, and K-means are shown in

{(x0, x1], (x1, x2],…, (xn
−1, xn)}, x0 xn +∞ −∞,

x0<x≤x1, …, xn−1<x<xn.

Figures 2.13(b) , 2.13(c) , and 2.13(d) , respectively. The split points
are represented as dashed lines.

Figure 2.13.
Different discretization techniques.

In this particular example, if we measure the performance of a
discretization technique by the extent to which different objects that clump
together have the same categorical value, then K-means performs best,
followed by equal frequency, and finally, equal width. More generally, the
best discretization will depend on the application and often involves
domain-specific discretization. For example, the discretization of people
into low income, middle income, and high income is based on economic

Supervised Discretization

If classification is our application and class labels are known for some data
objects, then discretization approaches that use class labels often produce
better classification. This should not be surprising, since an interval
constructed with no knowledge of class labels often contains a mixture of
class labels. A conceptually simple approach is to place the splits in a way
that maximizes the purity of the intervals, i.e., the extent to which an interval
contains a single class label. In practice, however, such an approach requires
potentially arbitrary decisions about the purity of an interval and the minimum
size of an interval.

To overcome such concerns, some statistically based approaches start with
each attribute value in a separate interval and create larger intervals by
merging adjacent intervals that are similar according to a statistical test. An
alternative to this bottom-up approach is a top-down approach that starts by
bisecting the initial values so that the resulting two intervals give minimum
entropy. This technique only needs to consider each value as a possible split
point, because it is assumed that intervals contain ordered sets of values. The
splitting process is then repeated with another interval, typically choosing the
interval with the worst (highest) entropy, until a user-specified number of
intervals is reached, or a stopping criterion is satisfied.

Entropy-based approaches are one of the most promising approaches to
discretization, whether bottom-up or top-down. First, it is necessary to define
entropy. Let k be the number of different class labels, m be the number of

values in the i interval of a partition, and m be the number of values of class
j in interval i. Then the entropy e of the i interval is given by the equation

where is the probability (fraction of values) of class j in the
interval. The total entropy, e, of the partition is the weighted average of the
individual interval entropies, i.e.,

where m is the number of values, is the fraction of values in the
interval, and n is the number of intervals. Intuitively, the entropy of an interval
is a measure of the purity of an interval. If an interval contains only values of
one class (is perfectly pure), then the entropy is 0 and it contributes nothing to
the overall entropy. If the classes of values in an interval occur equally often
(the interval is as impure as possible), then the entropy is a maximum.

Example 2.13 (Discretization of Two Attributes).
The top-down method based on entropy was used to independently
discretize both the x and y attributes of the two-dimensional data shown in
Figure 2.14 . In the first discretization, shown in Figure 2.14(a) , the x
and y attributes were both split into three intervals. (The dashed lines
indicate the split points.) In the second discretization, shown in Figure
2.14(b) , the x and y attributes were both split into five intervals.




ei=−∑j=1kpijlog2 pij,

pij=mij/mi ith


wi=mi/m ith

Figure 2.14.
Discretizing x and y attributes for four groups (classes) of points.

This simple example illustrates two aspects of discretization. First, in two
dimensions, the classes of points are well separated, but in one dimension,
this is not so. In general, discretizing each attribute separately often
guarantees suboptimal results. Second, five intervals work better than three,
but six intervals do not improve the discretization much, at least in terms of
entropy. (Entropy values and results for six intervals are not shown.)
Consequently, it is desirable to have a stopping criterion that automatically
finds the right number of partitions.

Categorical Attributes with Too Many Values
Categorical attributes can sometimes have too many values. If the categorical
attribute is an ordinal attribute, then techniques similar to those for continuous
attributes can be used to reduce the number of categories. If the categorical
attribute is nominal, however, then other approaches are needed. Consider a

university that has a large number of departments. Consequently, a
department name attribute might have dozens of different values. In this
situation, we could use our knowledge of the relationships among different
departments to combine departments into larger groups, such as engineering,
social sciences, or biological sciences. If domain knowledge does not serve
as a useful guide or such an approach results in poor classification
performance, then it is necessary to use a more empirical approach, such as
grouping values together only if such a grouping results in improved
classification accuracy or achieves some other data mining objective.

2.3.7 Variable Transformation

A variable transformation refers to a transformation that is applied to all the
values of a variable. (We use the term variable instead of attribute to adhere
to common usage, although we will also refer to attribute transformation on
occasion.) In other words, for each object, the transformation is applied to the
value of the variable for that object. For example, if only the magnitude of a
variable is important, then the values of the variable can be transformed by
taking the absolute value. In the following section, we discuss two important
types of variable transformations: simple functional transformations and

Simple Functions
For this type of variable transformation, a simple mathematical function is
applied to each value individually. If x is a variable, then examples of such
transformations include or In statistics, variable
transformations, especially sqrt, log, and 1/x, are often used to transform data
that does not have a Gaussian (normal) distribution into data that does. While

xk, log x, ex, x, 1/x, sin x, |x|.

this can be important, other reasons often take precedence in data mining.
Suppose the variable of interest is the number of data bytes in a session, and
the number of bytes ranges from 1 to 1 billion. This is a huge range, and it can
be advantageous to compress it by using a log transformation. In this case,
sessions that transferred and bytes would be more similar to each
other than sessions that transferred 10 and 1000 bytes
For some applications, such as network intrusion detection, this may be what
is desired, since the first two sessions most likely represent transfers of large
files, while the latter two sessions could be two quite distinct types of

Variable transformations should be applied with caution because they change
the nature of the data. While this is what is desired, there can be problems if
the nature of the transformation is not fully appreciated. For instance, the
transformation 1/x reduces the magnitude of values that are 1 or larger, but
increases the magnitude of values between 0 and 1. To illustrate, the values
{1, 2, 3} go to but the values go to {1, 2, 3}. Thus, for
all sets of values, the transformation 1/x reverses the order. To help clarify the
effect of a transformation, it is important to ask questions such as the
following: What is the desired property of the transformed attribute? Does the
order need to be maintained? Does the transformation apply to all values,
especially negative values and 0? What is the effect of the transformation on
the values between 0 and 1? Exercise 17 on page 109 explores other
aspects of variable transformation.

Normalization or Standardization
The goal of standardization or normalization is to make an entire set of values
have a particular property. A traditional example is that of “standardizing a
variable” in statistics. If is the mean (average) of the attribute values and
is their standard deviation, then the transformation creates a new


108 109
(9−8=1 versus 3−1=3).

{ 1, 12, 13 }, { 1, 12, 13 }

x¯ sx

variable that has a mean of 0 and a standard deviation of 1. If different
variables are to be used together, e.g., for clustering, then such a
transformation is often necessary to avoid having a variable with large values
dominate the results of the analysis. To illustrate, consider comparing people
based on two variables: age and income. For any two people, the difference in
income will likely be much higher in absolute terms (hundreds or thousands of
dollars) than the difference in age (less than 150). If the differences in the
range of values of age and income are not taken into account, then the
comparison between people will be dominated by differences in income. In
particular, if the similarity or dissimilarity of two people is calculated using the
similarity or dissimilarity measures defined later in this chapter, then in many
cases, such as that of Euclidean distance, the income values will dominate
the calculation.

The mean and standard deviation are strongly affected by outliers, so the
above transformation is often modified. First, the mean is replaced by the
median, i.e., the middle value. Second, the standard deviation is replaced by
the absolute standard deviation. Specifically, if x is a variable, then the
absolute standard deviation of x is given by where is the
value of the variable, m is the number of objects, and is either the mean

or median. Other approaches for computing estimates of the location (center)
and spread of a set of values in the presence of outliers are described in
statistics books. These more robust measures can also be used to define a
standardization transformation.

σA=∑i=1m|xi−μ|, xi
ith μ

2.4 Measures of Similarity and
Similarity and dissimilarity are important because they are used by a number
of data mining techniques, such as clustering, nearest neighbor classification,
and anomaly detection. In many cases, the initial data set is not needed once
these similarities or dissimilarities have been computed. Such approaches can
be viewed as transforming the data to a similarity (dissimilarity) space and
then performing the analysis. Indeed, kernel methods are a powerful
realization of this idea. These methods are introduced in Section 2.4.7 and
are discussed more fully in the context of classification in Section 4.9.4.

We begin with a discussion of the basics: high-level definitions of similarity
and dissimilarity, and a discussion of how they are related. For convenience,
the term proximity is used to refer to either similarity or dissimilarity. Since
the proximity between two objects is a function of the proximity between the
corresponding attributes of the two objects, we first describe how to measure
the proximity between objects having only one attribute.

We then consider proximity measures for objects with multiple attributes. This
includes measures such as the Jaccard and cosine similarity measures, which
are useful for sparse data, such as documents, as well as correlation and
Euclidean distance, which are useful for non-sparse (dense) data, such as
time series or multi-dimensional points. We also consider mutual information,
which can be applied to many types of data and is good for detecting
nonlinear relationships. In this discussion, we restrict ourselves to objects with
relatively homogeneous attribute types, typically binary or continuous.

Next, we consider several important issues concerning proximity measures.
This includes how to compute proximity between objects when they have
heterogeneous types of attributes, and approaches to account for differences
of scale and correlation among variables when computing distance between
numerical objects. The section concludes with a brief discussion of how to
select the right proximity measure.

Although this section focuses on the computation of proximity between data
objects, proximity can also be computed between attributes. For example, for
the document-term matrix of Figure 2.2(d) , the cosine measure can be
used to compute similarity between a pair of documents or a pair of terms
(words). Knowing that two variables are strongly related can, for example, be
helpful for eliminating redundancy. In particular, the correlation and mutual
information measures discussed later are often used for that purpose.

2.4.1 Basics

Informally, the similarity between two objects is a numerical measure of the
degree to which the two objects are alike. Consequently, similarities are
higher for pairs of objects that are more alike. Similarities are usually non-
negative and are often between 0 (no similarity) and 1 (complete similarity).

The dissimilarity between two objects is a numerical measure of the degree
to which the two objects are different. Dissimilarities are lower for more similar
pairs of objects. Frequently, the term distance is used as a synonym for
dissimilarity, although, as we shall see, distance often refers to a special class

of dissimilarities. Dissimilarities sometimes fall in the interval [0, 1], but it is
also common for them to range from 0 to ∞.

Transformations are often applied to convert a similarity to a dissimilarity, or
vice versa, or to transform a proximity measure to fall within a particular
range, such as [0,1]. For instance, we may have similarities that range from 1
to 10, but the particular algorithm or software package that we want to use
may be designed to work only with dissimilarities, or it may work only with
similarities in the interval [0,1]. We discuss these issues here because we will
employ such transformations later in our discussion of proximity. In addition,
these issues are relatively independent of the details of specific proximity

Frequently, proximity measures, especially similarities, are defined or
transformed to have values in the interval [0,1]. Informally, the motivation for
this is to use a scale in which a proximity value indicates the fraction of
similarity (or dissimilarity) between two objects. Such a transformation is often
relatively straightforward. For example, if the similarities between objects
range from 1 (not at all similar) to 10 (completely similar), we can make them
fall within the range [0, 1] by using the transformation where s and
s′ are the original and new similarity values, respectively. In the more general
case, the transformation of similarities to the interval [0, 1] is given by the
expression where max_s and min_s are the
maximum and minimum similarity values, respectively. Likewise, dissimilarity
measures with a finite range can be mapped to the interval [0,1] by using the
formula This is an example of a linear
transformation, which preserves the relative distances between points. In
other words, if points, and are twice as far apart as points, and
the same will be true after a linear transformation.




x1 x2, x3 x4,

However, there can be complications in mapping proximity measures to the
interval [0, 1] using a linear transformation. If, for example, the proximity
measure originally takes values in the interval then max_d is not defined
and a nonlinear transformation is needed. Values will not have the same
relationship to one another on the new scale. Consider the transformation

for a dissimilarity measure that ranges from 0 to The
dissimilarities 0, 0.5, 2, 10, 100, and 1000 will be transformed into the new
dissimilarities 0, 0.33, 0.67, 0.90, 0.99, and 0.999, respectively. Larger values
on the original dissimilarity scale are compressed into the range of values
near 1, but whether this is desirable depends on the application.

Note that mapping proximity measures to the interval [0, 1] can also change
the meaning of the proximity measure. For example, correlation, which is
discussed later, is a measure of similarity that takes values in the interval

Mapping these values to the interval [0,1] by taking the absolute value
loses information about the sign, which can be important in some applications.
See Exercise 22 on page 111 .

Transforming similarities to dissimilarities and vice versa is also relatively
straightforward, although we again face the issues of preserving meaning and
changing a linear scale into a nonlinear scale. If the similarity (or dissimilarity)
falls in the interval [0,1], then the dissimilarity can be defined as

Another simple approach is to define similarity as the negative
of the dissimilarity (or vice versa). To illustrate, the dissimilarities 0, 1, 10, and
100 can be transformed into the similarities and

The similarities resulting from the negation transformation are not restricted to
the range [0, 1], but if that is desired, then transformations such as

or can be used. For the
transformation the dissimilarities 0, 1, 10, 100 are transformed into 1,


d=d/(1+d) ∞.

[−1, 1].


0, −1, −10, −100,

s=1d+1, s=e−d, s=1−d−min_dmax_d−min_d

0.5, 0.09, 0.01, respectively. For they become 1.00, 0.37, 0.00, 0.00,
respectively, while for they become 1.00, 0.99,
0.90, 0.00, respectively. In this discussion, we have focused on converting
dissimilarities to similarities. Conversion in the opposite direction is considered
in Exercise 23 on page 111 .

In general, any monotonic decreasing function can be used to convert
dissimilarities to similarities, or vice versa. Of course, other factors also must
be considered when transforming similarities to dissimilarities, or vice versa,
or when transforming the values of a proximity measure to a new scale. We
have mentioned issues related to preserving meaning, distortion of scale, and
requirements of data analysis tools, but this list is certainly not exhaustive.

2.4.2 Similarity and Dissimilarity
between Simple Attributes

The proximity of objects with a number of attributes is typically defined by
combining the proximities of individual attributes, and thus, we first discuss
proximity between objects having a single attribute. Consider objects
described by one nominal attribute. What would it mean for two such objects
to be similar? Because nominal attributes convey only information about the
distinctness of objects, all we can say is that two objects either have the same
value or they do not. Hence, in this case similarity is traditionally defined as 1
if attribute values match, and as 0 otherwise. A dissimilarity would be defined
in the opposite way: 0 if the attribute values match, and 1 if they do not.

For objects with a single ordinal attribute, the situation is more complicated
because information about order should be taken into account. Consider an


attribute that measures the quality of a product, e.g., a candy bar, on the scale
{poor, fair, OK, good, wonderful}. It would seem reasonable that a product,
P1, which is rated wonderful, would be closer to a product P2, which is rated
good, than it would be to a product P3, which is rated OK. To make this
observation quantitative, the values of the ordinal attribute are often mapped
to successive integers, beginning at 0 or 1, e.g.,

Then, or, if
we want the dissimilarity to fall between 0 and A
similarity for ordinal attributes can then be defined as

This definition of similarity (dissimilarity) for an ordinal attribute should make
the reader a bit uneasy since this assumes equal intervals between
successive values of the attribute, and this is not necessarily so. Otherwise,
we would have an interval or ratio attribute. Is the difference between the
values fair and good really the same as that between the values OK and
wonderful? Probably not, but in practice, our options are limited, and in the
absence of more information, this is the standard approach for defining
proximity between ordinal attributes.

For interval or ratio attributes, the natural measure of dissimilarity between
two objects is the absolute difference of their values. For example, we might
compare our current weight and our weight a year ago by saying “I am ten
pounds heavier.” In cases such as these, the dissimilarities typically range
from 0 to rather than from 0 to 1. The similarity of interval or ratio attributes
is typically expressed by transforming a dissimilarity into a similarity, as
previously described.

Table 2.7 summarizes this discussion. In this table, x and y are two objects
that have one attribute of the indicated type. Also, d(x, y) and s(x, y) are the
dissimilarity and similarity between x and y, respectively. Other approaches
are possible; these are the most common ones.

{poor=0, fair=1, OK=2, good=3, wonderful=4}. d(P1, P2)=3−2=1
d(P1, P2)=3−24=0.25.



Table 2.7. Similarity and dissimilarity for simple attributes


Dissimilarity Similarity


Ordinal (values mapped to
integers 0 to , where n is the
number of values)

or Ratio

The following two sections consider more complicated measures of proximity
between objects that involve multiple attributes: (1) dissimilarities between
data objects and (2) similarities between data objects. This division allows us
to more naturally display the underlying motivations for employing various
proximity measures. We emphasize, however, that similarities can be
transformed into dissimilarities and vice versa using the approaches described

2.4.3 Dissimilarities between Data

In this section, we discuss various kinds of dissimilarities. We begin with a
discussion of distances, which are dissimilarities with certain properties, and
then provide examples of more general kinds of dissimilarities.


d={ 0if x=y1if x≠y s={ 1if x=y0if x≠y



d=|x−y| s=−d, s=11+d, s=e−d,s=1−d−min_dmax_d−min_d

We first present some examples, and then offer a more formal description of
distances in terms of the properties common to all distances. The Euclidean
distance ,d , between two points, x and y , in one-, two-, three-, or higher-
dimensional space, is given by the following familiar formula:

where n is the number of dimensions and and are, respectively, the
attributes (components) of x and y. We illustrate this formula with Figure
2.15 and Tables 2.8 and 2.9 , which show a set of points, the x and y
coordinates of these points, and the distance matrix containing the pairwise
distances of these points.

Figure 2.15.
Four two-dimensional points.

The Euclidean distance measure given in Equation 2.1 is generalized by
the Minkowski distance metric shown in Equation 2.2 ,

d(x,y)=∑k=1n(xk−yk)2, (2.1)

xk yk kth

d(x,y)=(∑k=1n|xk−yk|r)1/r, (2.2)

where r is a parameter. The following are the three most common examples of
Minkowski distances.

City block (Manhattan, taxicab, norm) distance. A common
example is the Hamming distance , which is the number of bits that is
different between two objects that have only binary attributes, i.e., between
two binary vectors.

Euclidean distance ( norm).
Supremum ( or norm) distance. This is the maximum

difference between any attribute of the objects. More formally, the
distance is defined by Equation 2.3

The r parameter should not be confused with the number of dimensions (at-
tributes) n. The Euclidean, Manhattan, and supremum distances are defined
for all values of n: 1, 2, 3, …, and specify different ways of combining the
differences in each dimension (attribute) into an overall distance.

Tables 2.10 and 2.11 , respectively, give the proximity matrices for the
and distances using data from Table 2.8 . Notice that all these distance
matrices are symmetric; i.e., the entry is the same as the entry. In
Table 2.9 , for instance, the fourth row of the first column and the fourth
column of the first row both contain the value 5.1.

Table 2.8. x and y coordinates of four points.

point x coordinate y coordinate

p1 0 2

p2 2 0

p3 3 1

r=1. L1

r=2. L2
r=∞. Lmax L∞


d(x, y)=limr→∞(∑k=1n|xk−yk|r)1/r. (2.3)


ijth jith

p4 5 1

Table 2.9. Euclidean distance matrix for Table 2.8 .

p1 p2 p3 p4

p1 0.0 2.8 3.2 5.1

p2 2.8 0.0 1.4 3.2

p3 3.2 1.4 0.0 2.0

p4 5.1 3.2 2.0 0.0

Table 2.10. distance matrix for Table 2.8 .

L p1 p2 p3 p4

p1 0.0 4.0 4.0 6.0

p2 4.0 0.0 2.0 4.0

p3 4.0 2.0 0.0 2.0

p4 6.0 4.0 2.0 0.0

Table 2.11. distance matrix for Table 2.8 .

p1 p2 p3 p4

p1 0.0 2.0 3.0 5.0

p2 2.0 0.0 1.0 3.0

p3 3.0 1.0 0.0 2.0





p4 5.0 3.0 2.0 0.0

Distances, such as the Euclidean distance, have some well-known properties.
If d(x, y) is the distance between two points, x and y, then the following
properties hold.

1. Positivity
a. for all x and y,
b. only if

2. Symmetry for all x and y.
3. Triangle Inequality for all points x , y , and z.

Measures that satisfy all three properties are known as metrics. Some people
use the term distance only for dissimilarity measures that satisfy these
properties, but that practice is often violated. The three properties described
here are useful, as well as mathematically pleasing. Also, if the triangle
inequality holds, then this property can be used to increase the efficiency of
techniques (including clustering) that depend on distances possessing this
property. (See Exercise 25 .) Nonetheless, many dissimilarities do not
satisfy one or more of the metric properties. Example 2.14 illustrates such
a measure.

Example 2.14 (Non-metric Dissimilarities: Set
This example is based on the notion of the difference of two sets, as
defined in set theory. Given two sets A and B, is the set of elements of
A that are not in

d(x, y)≥0
d(x, y)=0 x=y.

d(x, y)=d(y, x)
d(x, z)≤d(x, y)+d(y, z)


B. For example, if and then and
the empty set. We can define the distance d between two sets A

and B as where size is a function returning the number
of elements in a set. This distance measure, which is an integer value
greater than or equal to 0, does not satisfy the second part of the positivity
property, the symmetry property, or the triangle inequality. However, these
properties can be made to hold if the dissimilarity measure is modified as
follows: See Exercise 21 on page 110 .

2.4.4 Similarities between Data

For similarities, the triangle inequality (or the analogous property) typically
does not hold, but symmetry and positivity typically do. To be explicit, if s(x, y)
is the similarity between points x and y, then the typical properties of
similarities are the following:

1. only if
2. for all x and y. (Symmetry)

There is no general analog of the triangle inequality for similarity measures. It
is sometimes possible, however, to show that a similarity measure can easily
be converted to a metric distance. The cosine and Jaccard similarity
measures, which are discussed shortly, are two examples. Also, for specific
similarity measures, it is possible to derive mathematical bounds on the
similarity between two objects that are similar in spirit to the triangle inequality.

Example 2.15 (A Non-symmetric Similarity

A={1, 2, 3, 4} B={2, 3, 4}, A−B={1} B

d(A, B)=size(A−B),

d(A, B)=size(A−B)+size(B−A).

s(x, y)=1 x=y. (0≤s≤1)
s(x, y)=s(y, x)

Consider an experiment in which people are asked to classify a small set
of characters as they flash on a screen. The confusion matrix for this
experiment records how often each character is classified as itself, and
how often each is classified as another character. Using the confusion
matrix, we can define a similarity measure between a character x and a
character y as the number of times that x is misclassified asy , but note
that this measure is not symmetric. For example, suppose that “0”
appeared 200 times and was classified as a “0” 160 times, but as an “o” 40
times. Likewise, suppose that “o” appeared 200 times and was classified
as an “o” 170 times, but as “0” only 30 times. Then, but

In such situations, the similarity measure can be made
symmetric by setting where s indicates the
new similarity measure.

2.4.5 Examples of Proximity Measures

This section provides specific examples of some similarity and dissimilarity

Similarity Measures for Binary Data
Similarity measures between objects that contain only binary attributes are
called similarity coefficients , and typically have values between 0 and 1. A
value of 1 indicates that the two objects are completely similar, while a value
of 0 indicates that the objects are not at all similar. There are many rationales
for why one coefficient is better than another in specific instances.

s(o, 0)=30.


Let x and y be two objects that consist of n binary attributes. The comparison
of two such objects, i.e., two binary vectors, leads to the following four
quantities (frequencies):

Simple Matching Coefficient

One commonly used similarity coefficient is the simple matching coefficient
(SMC), which is defined as

This measure counts both presences and absences equally. Consequently,
the SMC could be used to find students who had answered questions similarly
on a test that consisted only of true/false questions.

Jaccard Coefficient

Suppose that x and y are data objects that represent two rows (two
transactions) of a transaction matrix (see Section 2.1.2 ). If each
asymmetric binary attribute corresponds to an item in a store, then a 1
indicates that the item was purchased, while a 0 indicates that the product
was not purchased. Because the number of products not purchased by any
customer far outnumbers the number of products that were purchased, a
similarity measure such as SMC would say that all transactions are very
similar. As a result, the Jaccard coefficient is frequently used to handle objects
consisting of asymmetric binary attributes. The Jaccard coefficient , which is
often symbolized by j, is given by the following equation:

f00=the number of attributes where x is 0 and y is 0f01= the number of attributes where

SMC=number of matching attribute valuesnumber of attributes=f11+f00f01+f10(2.4)

J=number of matching presencesnumber of attributes not involved in 00 matches(2.5)

Example 2.16 (The SMC and Jaccard Similarity
To illustrate the difference between these two similarity measures, we
calculate SMC and j for the following two binary vectors.

Cosine Similarity
Documents are often represented as vectors, where each component
(attribute) represents the frequency with which a particular term (word) occurs
in the document. Even though documents have thousands or tens of
thousands of attributes (terms), each document is sparse since it has
relatively few nonzero attributes. Thus, as with transaction data, similarity
should not depend on the number of shared 0 values because any two
documents are likely to “not contain” many of the same words, and therefore,
if 0–0 matches are counted, most documents will be highly similar to most
other documents. Therefore, a similarity measure for documents needs to
ignores 0–0 matches like the Jaccard measure, but also must be able to
handle non-binary vectors. The cosine similarity , defined next, is one of the
most common measures of document similarity. If x and y are two document
vectors, then

x = (1, 0, 0, 0, 0, 0, 0, 0, 0, 0)y = (0, 0, 0, 0, 0, 0, 1, 0, 0, 1)

f01=2the number of attributes where x was 0 and y was 1f10=1the number of attributes where



where ′ indicates vector or matrix transpose and indicates the inner
product of the two vectors,

and is the length of vector

The inner product of two vectors works well for asymmetric attributes since it
depends only on components that are non-zero in both vectors. Hence, the
similarity between two documents depends only upon the words that appear
in both of them.

Example 2.17 (Cosine Similarity between Two
Document Vectors).
This example calculates the cosine similarity for the following two data
objects, which might represent document vectors:

As indicated by Figure 2.16 , cosine similarity really is a measure of the
(cosine of the) angle between x and y. Thus, if the cosine similarity is 1, the
angle between x and y is and x and y are the same except for length. If
the cosine similarity is 0, then the angle between x and y is and they do
not share any terms (words).

cos(x, y)=⟨ x, y ⟩∥x∥∥y∥=x′y∥x∥∥y∥, (2.6)

⟨ x, y ⟩

⟨ x, y ⟩=∑k=1nxkyk=x′y, (2.7)

∥x∥ x, ∥x∥=∑k=1nxk2=⟨ x, x ⟩=x′x.

x=(3, 2, 0, 5, 0, 0, 0, 2, 0, 0)y=(1, 0, 0, 0, 0, 0, 0, 1, 0, 2)

⟨ x, y


Figure 2.16.
Geometric illustration of the cosine measure.

Equation 2.6 also can be written as Equation 2.8 .

where and Dividing x and y by their lengths normalizes them
to have a length of 1. This means that cosine similarity does not take the
length of the two data objects into account when computing similarity.
(Euclidean distance might be a better choice when length is important.) For
vectors with a length of 1, the cosine measure can be calculated by taking a
simple inner product. Consequently, when many cosine similarities between
objects are being computed, normalizing the objects to have unit length can
reduce the time required.

Extended Jaccard Coefficient (Tanimoto
The extended Jaccard coefficient can be used for document data and that
reduces to the Jaccard coefficient in the case of binary attributes. This
coefficient, which we shall represent as EJ, is defined by the following

cos(x, y)=⟨ x∥x∥, y∥y∥ ⟩=⟨ x′, y′ ⟩, (2.8)

x′=x/∥x∥ y′=y/∥y∥.

EJ(x, y)=⟨ x, y ⟩ǁ x ǁ2+ǁ y ǁ2−⟨ x, y ⟩=x′yǁ x ǁ2+ǁ y ǁ2−x′y. (2.9)

Correlation is frequently used to measure the linear relationship between two
sets of values that are observed together. Thus, correlation can measure the
relationship between two variables (height and weight) or between two objects
(a pair of temperature time series). Correlation is used much more frequently
to measure the similarity between attributes since the values in two data
objects come from different attributes, which can have very different attribute
types and scales. There are many types of correlation, and indeed correlation
is sometimes used in a general sense to mean the relationship between two
sets of values that are observed together. In this discussion, we will focus on a
measure appropriate for numerical values.

Specifically, Pearson’s correlation between two sets of numerical values,
i.e., two vectors, x and y, is defined by the following equation:

where we use the following standard statistical notation and definitions:

corr(x, y)=covariance(x, y)standard_deviation(x)×standard_deviation(y)=sxysx (2.10)

covariance(x, y)=sxy=1n−1∑k=1n(xk−x¯)(yk−y¯) (2.11)



x¯=1n∑k=1nxk is the mean of x

y¯=1n∑k=1nyk is the mean of y

Example 2.18 (Perfect Correlation).
Correlation is always in the range to 1. A correlation of means
that x and y have a perfect positive (negative) linear relationship; that is,

where a and b are constants. The following two vectors x and y
illustrate cases where the correlation is and respectively. In the first
case, the means of x and y were chosen to be 0, for simplicity.

Example 2.19 (Nonlinear Relationships).
If the correlation is 0, then there is no linear relationship between the two
sets of values. However, nonlinear relationships can still exist. In the
following example, but their correlation is 0.

Example 2.20 (Visualizing Correlation).
It is also easy to judge the correlation between two vectors x and y by
plotting pairs of corresponding values of x and y in a scatter plot. Figure
2.17 shows a number of these scatter plots when x and y consist of a
set of 30 pairs of values that are randomly generated (with a normal
distribution) so that the correlation of x and y ranges from to 1. Each
circle in a plot represents one of the 30 pairs of x and y values; its x
coordinate is the value of that pair for x, while its y coordinate is the value
of the same pair for y.

−1 1 (−1)

−1 +1,

x=(−3, 6, 0, 3, −6)y=(1, −2, 0, −1, 2)corr(x, y)=−1xk=−3yk

x=(3, 6, 0, 3, 6)y=(1, 2, 0, 1, 2)corr(x, y)=1xk=3yk


x=(−3, −2, −1, 0, 1, 2, 3)y=(9, 4, 1, 0, 1, 4, 9)


Figure 2.17.
Scatter plots illustrating correlations from to 1.

If we transform x and y by subtracting off their means and then normalizing
them so that their lengths are 1, then their correlation can be calculated by
taking the dot product. Let us refer to these transformed vectors of x and y as
and , respectively. (Notice that this transformation is not the same as the

standardization used in other contexts, where we subtract the means and
divide by the standard deviations, as discussed in Section 2.3.7 .) This
transformation highlights an interesting relationship between the correlation
measure and the cosine measure. Specifically, the correlation between x and
y is identical to the cosine between and However, the cosine between x
and y is not the same as the cosine between and even though they both
have the same correlation measure. In general, the correlation between two


x′ y′

x′ y′.
x′ y′,

vectors is equal to the cosine measure only in the special case when the
means of the two vectors are 0.

Differences Among Measures For Continuous
In this section, we illustrate the difference among the three proximity
measures for continuous attributes that we have just defined: cosine,
correlation, and Minkowski distance. Specifically, we consider two types of
data transformations that are commonly used, namely, scaling (multiplication)
by a constant factor and translation (addition) by a constant value. A proximity
measure is considered to be invariant to a data transformation if its value
remains unchanged even after performing the transformation. Table 2.12
compares the behavior of cosine, correlation, and Minkowski distance
measures regarding their invariance to scaling and translation operations. It
can be seen that while correlation is invariant to both scaling and translation,
cosine is only invariant to scaling but not to translation. Minkowski distance
measures, on the other hand, are sensitive to both scaling and translation and
are thus invariant to neither.

Table 2.12. Properties of cosine, correlation, and Minkowski distance

Property Cosine Correlation Minkowski Distance

Invariant to scaling (multiplication) Yes Yes No

Invariant to translation (addition) No Yes No

Let us consider an example to demonstrate the significance of these
differences among different proximity measures.

Example 2.21 (Comparing proximity measures).
Consider the following two vectors x and y with seven numeric attributes.

It can be seen that both x and y have 4 non-zero values, and the values in
the two vectors are mostly the same, except for the third and the fourth
components. The cosine, correlation, and Euclidean distance between the
two vectors can be computed as follows.

Not surprisingly, x and y have a cosine and correlation measure close to 1,
while the Euclidean distance between them is small, indicating that they
are quite similar. Now let us consider the vector which is a scaled
version of y (multiplied by a constant factor of 2), and the vector which
is constructed by translating y by 5 units as follows.

We are interested in finding whether and show the same proximity
with x as shown by the original vector y. Table 2.13 shows the different
measures of proximity computed for the pairs and It
can be seen that the value of correlation between x and y remains
unchanged even after replacing y with or However, the value of
cosine remains equal to 0.9667 when computed for (x, y) and but
significantly reduces to 0.7940 when computed for This highlights

x=(1, 2, 4, 3, 0, 0, 0)y=(1, 2, 3, 4, 0, 0, 0)

cos(x, y)=2930×30=0.9667correlation(x, y)=2.35711.5811×1.5811=0.9429Euclidean distance
x−y ǁ=1.4142


ys=2×y=(2, 4, 6, 8, 0, 0, 0)

yt=y+5=(6, 7, 8, 9, 5, 5, 5)

ys yt

(x, y), (x, ys), (x, yt).

ys yt.
(x, ys),

(x, yt).

the fact that cosine is invariant to the scaling operation but not to the
translation operation, in contrast with the correlation measure. The
Euclidean distance, on the other hand, shows different values for all three
pairs of vectors, as it is sensitive to both scaling and translation.

Table 2.13. Similarity between and

Measure (x, y)

Cosine 0.9667 0.9667 0.7940

Correlation 0.9429 0.9429 0.9429

Euclidean Distance 1.4142 5.8310 14.2127

We can observe from this example that different proximity measures
behave differently when scaling or translation operations are applied on
the data. The choice of the right proximity measure thus depends on the
desired notion of similarity between data objects that is meaningful for a
given application. For example, if x and y represented the frequencies of
different words in a document-term matrix, it would be meaningful to use a
proximity measure that remains unchanged when y is replaced by
because is just a scaled version of y with the same distribution of words
occurring in the document. However, is different from y, since it contains
a large number of words with non-zero frequencies that do not occur in y.
Because cosine is invariant to scaling but not to translation, it will be an
ideal choice of proximity measure for this application.

Consider a different scenario in which x represents a location’s
temperature measured on the Celsius scale for seven days. Let and
be the temperatures measured on those days at a different location, but

using three different measurement scales. Note that different units of

(x, y), (x, ys), (x, yt).

(x, ys) (x, yt)



y, ys,

temperature have different offsets (e.g. Celsius and Kelvin) and different
scaling factors (e.g. Celsius and Fahrenheit). It is thus desirable to use a
proximity measure that captures the proximity between temperature values
without being affected by the measurement scale. Correlation would then
be the ideal choice of proximity measure for this application, as it is
invariant to both scaling and translation.

As another example, consider a scenario where x represents the amount
of precipitation (in cm) measured at seven locations. Let and be
estimates of the precipitation at these locations, which are predicted using
three different models. Ideally, we would like to choose a model that
accurately reconstructs the measurements in x without making any error. It
is evident that y provides a good approximation of the values in x, whereas
and provide poor estimates of precipitation, even though they do

capture the trend in precipitation across locations. Hence, we need to
choose a proximity measure that penalizes any difference in the model
estimates from the actual observations, and is sensitive to both the scaling
and translation operations. The Euclidean distance satisfies this property
and thus would be the right choice of proximity measure for this
application. Indeed, the Euclidean distance is commonly used in
computing the accuracy of models, which will be discussed later in
Chapter 3 .

2.4.6 Mutual Information

Like correlation, mutual information is used as a measure of similarity
between two sets of paired values that is sometimes used as an alternative to
correlation, particularly when a nonlinear relationship is suspected between
the pairs of values. This measure comes from information theory, which is the

y, ys, yt

ys yt

study of how to formally define and quantify information. Indeed, mutual
information is a measure of how much information one set of values provides
about another, given that the values come in pairs, e.g., height and weight. If
the two sets of values are independent, i.e., the value of one tells us nothing
about the other, then their mutual information is 0. On the other hand, if the
two sets of values are completely dependent, i.e., knowing the value of one
tells us the value of the other and vice-versa, then they have maximum mutual
information. Mutual information does not have a maximum value, but we will
define a normalized version of it that ranges between 0 and 1.

To define mutual information, we consider two sets of values, X and Y , which
occur in pairs (X, Y). We need to measure the average information in a single
set of values, i.e., either in X or in Y , and in the pairs of their values. This is
commonly measured by entropy. More specifically, assume X and Y are
discrete, that is, X can take m distinct values, and Y can take n
distinct values, Then their individual and joint entropy can be
defined in terms of the probabilities of each value and pair of values as

where if the probability of a value or combination of values is 0, then
is conventionally taken to be 0.

The mutual information of X and Y can now be defined straightforwardly:

u1, u2, …, um
v1, v2, …, vn.

H(X)=−∑j=1mP(X=uj)log2 P(X=uj) (2.12)

H(Y)=−∑k=1nP(Y=vk)log2 P(Y=vk) (2.13)

H(X, Y)=−∑j=1m∑k=1nP(X=uj, Y=vk)log2 P(X=uj, Y=vk) (2.14)

0 log2(0)

I(X, Y)=H(X)+H(Y)−H(X, Y) (2.15)

Note that H(X, Y) is symmetric, i.e., and thus mutual
information is also symmetric, i.e.,

Practically, X and Y are either the values in two attributes or two rows of the
same data set. In Example 2.22 , we will represent those values as two
vectors x and y and calculate the probability of each value or pair of values
from the frequency with which values or pairs of values occur in x, y and

where is the component of x and is the component of y. Let
us illustrate using a previous example.

Example 2.22 (Evaluating Nonlinear
Relationships with Mutual Information).
Recall Example 2.19 where but their correlation was 0.

From Figure 2.22 , Although a variety
of approaches to normalize mutual information are possible—see
Bibliographic Notes—for this example, we will apply one that divides the
mutual information by and produces a result between 0
and 1. This yields a value of Thus, we can see
that x and y are strongly related. They are not perfectly related because
given a value of y there is, except for some ambiguity about the value
of x. Notice that for the normalized mutual information would be 1.

Figure 2.18.
Computation of mutual information.

Table 2.14. Entropy for x

H(X, Y)=H(Y, X),
I(X, Y)=I(Y).

(xi, yi), xi ith yi ith


x=(−3, −2, −1, 0, 1, 2, 3)y=(9, 4, 1, 0, 1, 4, 9)

I(x, y)=H(x)+H(y)−H(x, y)=1.9502.

log2(min(m, n))


xj P(x=xj) −P(x=xj)log2 P(x=xj)

1/7 0.4011

1/7 0.4011

1/7 0.4011

0 1/7 0.4011

1 1/7 0.4011

2 1/7 0.4011

3 1/7 0.4011

H(x) 2.8074

Table 2.15. Entropy for y

9 2/7 0.5164

4 2/7 0.5164

1 2/7 0.5164

0 1/7 0.4011

H(y) 1.9502

Table 2.16. Joint entropy for x and y

9 1/7 0.4011

4 1/7 0.4011




yk P(y=yk) −P(y=yk)log2 P(y=yk)

xj yk P(x=xj, y=xk) −P(x=xj, y=xk)log2 P(x=xj, y=xk)



1 1/7 0.4011

0 0 1/7 0.4011

1 1 1/7 0.4011

2 4 1/7 0.4011

3 9 1/7 0.4011

H(x, y) 2.8074

2.4.7 Kernel Functions*

It is easy to understand how similarity and distance might be useful in an
application such as clustering, which tries to group similar objects together.
What is much less obvious is that many other data analysis tasks, including
predictive modeling and dimensionality reduction, can be expressed in terms
of pairwise “proximities” of data objects. More specifically, many data analysis
problems can be mathematically formulated to take as input, a kernel matrix,
K, which can be considered a type of proximity matrix. Thus, an initial
preprocessing step is used to convert the input data into a kernel matrix,
which is the input to the data analysis algorithm.

More formally, if a data set has m data objects, then K is an m by m matrix. If
and are the and data objects, respectively, then the entry of

K, is computed by a kernel function:


xi xj ith jth kij, ijth

kij=κ(xi, xj) (2.16)

As we will see in the material that follows, the use of a kernel matrix allows
both wider applicability of an algorithm to various kinds of data and an ability
to model nonlinear relationships with algorithms that are designed only for
detecting linear relationships.

Kernels make an algorithm data independent

If an algorithm uses a kernel matrix, then it can be used with any type of data
for which a kernel function can be designed. This is illustrated by Algorithm
2.1. Although only some data analysis algorithms can be modified to use a
kernel matrix as input, this approach is extremely powerful because it allows
such an algorithm to be used with almost any type of data for which an
appropriate kernel function can be defined. Thus, a classification algorithm
can be used, for example, with record data, string data, or graph data. If an
algorithm can be reformulated to use a kernel matrix, then its applicability to
different types of data increases dramatically. As we will see in later chapters,
many clustering, classification, and anomaly detection algorithms work only
with similarities or distances, and thus, can be easily modified to work with

Algorithm 2.1 Basic kernel algorithm.
1. Read in the m data objects in the data set.
2. Compute the kernel matrix, K by applying the kernel function,

to each pair of data objects.
3. Run the data analysis algorithm with K as input.
4. Return the analysis result, e.g., predicted class or cluster labels.

Mapping data into a higher dimensional data space can


allow modeling of nonlinear relationships

There is yet another, equally important, aspect of kernel based data
algorithms—their ability to model nonlinear relationships with algorithms that
model only linear relationships. Typically, this works by first transforming
(mapping) the data from a lower dimensional data space to a higher
dimensional space.

Example 2.23 (Mapping Data to a Higher
Dimensional Space).
Consider the relationship between two variables x and y given by the
following equation, which defines an ellipse in two dimensions (Figure
2.19(a) ):

Figure 2.19.
Mapping data to a higher dimensional space: two to three dimensions.

We can map our two dimensional data to three dimensions by creating
three new variables, u, v, and w, which are defined as follows:

As a result, we can now express Equation 2.17 as a linear one. This
equation describes a plane in three dimensions. Points on the ellipse will
lie on that plane, while points inside and outside the ellipse will lie on
opposite sides of the plane. See Figure 2.19(b) . The viewpoint of this
3D plot is along the surface of the separating plane so that the plane
appears as a line.

The Kernel Trick

The approach illustrated above shows the value in mapping data to higher
dimensional space, an operation that is integral to kernel-based methods.
Conceptually, we first define a function that maps data points x and y to
data points and in a higher dimensional space such that the inner
product gives the desired measure of proximity of x and y. It may seem
that we have potentially sacrificed a great deal by using such an approach,
because we can greatly expand the size of our data, increase the
computational complexity of our analysis, and encounter problems with the
curse of dimensionality by computing similarity in a high-dimensional space.
However, this is not the case since these problems can be avoided by defining
a kernel function that can compute the same similarity value, but with the
data points in the original space, i.e., This is known as
the kernel trick. Despite the name, the kernel trick has a very solid

4×2+9xy+7y2=10 (2.17)


4u+9v+7w=10 (2.18)

φ(x) φ(y)

⟨x, y⟩

κ(x, y)=⟨ φ(x), φ(y) ⟩.

mathematical foundation and is a remarkably powerful approach for data

Not every function of a pair of data objects satisfies the properties needed for
a kernel function, but it has been possible to design many useful kernels for a
wide variety of data types. For example, three common kernel functions are
the polynomial, Gaussian (radial basis function (RBF)), and sigmoid kernels. If
x and y are two data objects, specifically, two data vectors, then these two
kernel functions can be expressed as follows, respectively:

where and are constants, d is an integer parameter that gives the
polynomial degree, is the length of the vector and is a
parameter that governs the “spread” of a Gaussian.

Example 2.24 (The Polynomial Kernel).
Note that the kernel functions presented in the previous section are
computing the same similarity value as would be computed if we actually
mapped the data to a higher dimensional space and then computed an
inner product there. For example, for the polynomial kernel of degree 2, let
be the function that maps a two-dimensional data vector to the

higher dimensional space. Specifically, let

κ(x, y)−(x′y+c)d (2.19)

κ(x, y)=exp(−ǁ x−y ǁ/2σ2) (2.20)

κ(x, y)=tanh(αx′y+c) (2.21)

α c≥0
ǁ x−y ǁ x−y σ>0

φ x=(x1, x2)

φ(x)=(x12, x22, 2x1x2, 2cx1, 2cx2, c). (2.22)

For the higher dimensional space, let the proximity be defined as the inner
product of and i.e., Then, as previously
mentioned, it can be shown that

where is defined by Equation 2.19 above. Specifically, if
and then

More generally, the kernel trick depends on defining and so that
Equation 2.23 holds. This has been done for a wide variety of kernels.

This discussion of kernel-based approaches was intended only to provide a
brief introduction to this topic and has omitted many details. A fuller discussion
of the kernel-based approach is provided in Section 4.9.4, which discusses
these issues in the context of nonlinear support vector machines for
classification. More general references for the kernel based analysis can be
found in the Bibliographic Notes of this chapter.

2.4.8 Bregman Divergence*

This section provides a brief description of Bregman divergences, which are a
family of proximity functions that share some common properties. As a result,
it is possible to construct general data mining algorithms, such as clustering
algorithms, that work with any Bregman divergence. A concrete example is
the K-means clustering algorithm (Section 7.2). Note that this section requires
knowledge of vector calculus.

φ(x) φ(y), ⟨ φ(x), φ(y) ⟩.

κ(x, y)=⟨ φ(x), φ(y) ⟩ (2.23)

κ x=(x1, x2)
y=(y1, y2),

κ(x, y)=⟨ x, y ⟩=x′y=(x12y12, x22y22, 2x1x2y1y2, 2cx1y1, 2cx2y2, c2).(2.24)

κ φ

Bregman divergences are loss or distortion functions. To understand the idea
of a loss function, consider the following. Let x and y be two points, where y is
regarded as the original point and x is some distortion or approximation of it.
For example, x may be a point that was generated by adding random noise to
y. The goal is to measure the resulting distortion or loss that results if y is
approximated by x. Of course, the more similar x and y are, the smaller the
loss or distortion. Thus, Bregman divergences can be used as dissimilarity

More formally, we have the following definition.

Definition 2.6 (Bregman divergence)
Given a strictly convex function (with a few modest restrictions
that are generally satisfied), the Bregman divergence (loss
function) generated by that function is given by the
following equation:

where is the gradient of evaluated at is the vector
difference between x and y, and is the inner
product between and For points in Euclidean space,
the inner product is just the dot product.

D(x, y) can be written as where
and represents the equation of a plane that is tangent to the function at y.


D(x, y)

D(x, y)=ϕ(x)−ϕ(y)−⟨ ∇ϕ(y), (x−y) ⟩ (2.25)

∇ϕ(y) ϕ y, x−y,
⟨ ∇ϕ(y), (x−y) ⟩

∇ϕ(y) (x−y).

D(x, y)=ϕ(x)−L(x), L(x)=ϕ(y)+⟨ ∇ϕ(y), (x−y) ⟩

Using calculus terminology, L(x) is the linearization of around the point y,
and the Bregman divergence is just the difference between a function and a
linear approximation to that function. Different Bregman divergences are
obtained by using different choices for

Example 2.25.
We provide a concrete example using squared Euclidean distance, but
restrict ourselves to one dimension to simplify the mathematics. Let x and
y be real numbers and be the real-valued function, In that
case, the gradient reduces to the derivative, and the dot product reduces
to multiplication. Specifically, Equation 2.25 becomes Equation 2.26 .

The graph for this example, with is shown in Figure 2.20 . The
Bregman divergence is shown for two values of x: and



ϕ(t) ϕ(t)=t2.

D(x,y)=x2−y2−2y(x−y)=(x−y)2 (2.26)

x=2 x=3.

Figure 2.20.
Illustration of Bregman divergence.

2.4.9 Issues in Proximity Calculation

This section discusses several important issues related to proximity
measures: (1) how to handle the case in which attributes have different scales
and/or are correlated, (2) how to calculate proximity between objects that are
composed of different types of attributes, e.g., quantitative and qualitative, (3)
and how to handle proximity calculations when attributes have different
weights; i.e., when not all attributes contribute equally to the proximity of

Standardization and Correlation for Distance

An important issue with distance measures is how to handle the situation
when attributes do not have the same range of values. (This situation is often
described by saying that “the variables have different scales.”) In a previous
example, Euclidean distance was used to measure the distance between
people based on two attributes: age and income. Unless these two attributes
are standardized, the distance between two people will be dominated by

A related issue is how to compute distance when there is correlation between
some of the attributes, perhaps in addition to differences in the ranges of
values. A generalization of Euclidean distance, the Mahalanobis distance, is
useful when attributes are correlated, have different ranges of values (different
variances), and the distribution of the data is approximately Gaussian
(normal). Correlated variables have a large impact on standard distance
measures since a change in any of the correlated variables is reflected in a
change in all the correlated variables. Specifically, the Mahalanobis distance
between two objects (vectors) x and y is defined as

where is the inverse of the covariance matrix of the data. Note that the
covariance matrix is the matrix whose entry is the covariance of the
and attributes as defined by Equation 2.11 .

Example 2.26.
In Figure 2.21 , there are 1000 points, whose x and y attributes have a
correlation of 0.6. The distance between the two large points at the
opposite ends of the long axis of the ellipse is 14.7 in terms of Euclidean

Mahalanobis(x, y)=(x−y)′∑−1(x−y), (2.27)

∑ ijth ith


distance, but only 6 with respect to Mahalanobis distance. This is because
the Mahalanobis distance gives less emphasis to the direction of largest
variance. In practice, computing the Mahalanobis distance is expensive,
but can be worthwhile for data whose attributes are correlated. If the
attributes are relatively uncorrelated, but have different ranges, then
standardizing the variables is sufficient.

Figure 2.21.
Set of two-dimensional points. The Mahalanobis distance between the two
points represented by large dots is 6; their Euclidean distance is 14.7.

Combining Similarities for Heterogeneous

The previous definitions of similarity were based on approaches that assumed
all the attributes were of the same type. A general approach is needed when
the attributes are of different types. One straightforward approach is to
compute the similarity between each attribute separately using Table 2.7 ,
and then combine these similarities using a method that results in a similarity
between 0 and 1. One possible approach is to define the overall similarity as
the average of all the individual attribute similarities. Unfortunately, this
approach does not work well if some of the attributes are asymmetric
attributes. For example, if all the attributes are asymmetric binary attributes,
then the similarity measure suggested previously reduces to the simple
matching coefficient, a measure that is not appropriate for asymmetric binary
attributes. The easiest way to fix this problem is to omit asymmetric attributes
from the similarity calculation when their values are 0 for both of the objects
whose similarity is being computed. A similar approach also works well for
handling missing values.

In summary, Algorithm 2.2 is effective for computing an overall similarity
between two objects, x and y, with different types of attributes. This procedure
can be easily modified to work with dissimilarities.

Algorithm 2.2 Similarities of heterogeneous

1: For the attribute, compute a similarity, in the
range [0, 1].
2: Define an indicator variable, for the attribute as

kth sk(x, y),

δk, kth

0if the kth attribute is an asymmetric attribute andboth objects have a value of

Using Weights
In much of the previous discussion, all attributes were treated equally when
computing proximity. This is not desirable when some attributes are more
important to the definition of proximity than others. To address these
situations, the formulas for proximity can be modified by weighting the
contribution of each attribute.

With attribute weights, (2.28) becomes

The definition of the Minkowski distance can also be modified as follows:

2.4.10 Selecting the Right Proximity

A few general observations may be helpful. First, the type of proximity
measure should fit the type of data. For many types of dense, continuous
data, metric distance measures such as Euclidean distance are often used.
Proximity between continuous attributes is most often expressed in terms of

3: Compute the overall similarity between the two objects using
the following formula:
similarity (x, y)=∑k=1nδksk(x, y)∑k=1nδk(2.28)


similarity (x, y)=∑k=1nwkδksk(x, y)∑k=1nwkδk. (2.29)

d (x, y)=(∑k=1nwk|xk−yk|r)1/r. (2.30)

differences, and distance measures provide a well-defined way of combining
these differences into an overall proximity measure. Although attributes can
have different scales and be of differing importance, these issues can often be
dealt with as described earlier, such as normalization and weighting of

For sparse data, which often consists of asymmetric attributes, we typically
employ similarity measures that ignore 0–0 matches. Conceptually, this
reflects the fact that, for a pair of complex objects, similarity depends on the
number of characteristics they both share, rather than the number of
characteristics they both lack. The cosine, Jaccard, and extended Jaccard
measures are appropriate for such data.

There are other characteristics of data vectors that often need to be
considered. Invariance to scaling (multiplication) and to translation (addition)
were previously discussed with respect to Euclidean distance and the cosine
and correlation measures. The practical implications of such considerations
are that, for example, cosine is more suitable for sparse document data where
only scaling is important, while correlation works better for time series, where
both scaling and translation are important. Euclidean distance or other types
of Minkowski distance are most appropriate when two data vectors are to
match as closely as possible across all components (features).

In some cases, transformation or normalization of the data is needed to obtain
a proper similarity measure. For instance, time series can have trends or
periodic patterns that significantly impact similarity. Also, a proper computation
of similarity often requires that time lags be taken into account. Finally, two
time series may be similar only over specific periods of time. For example,
there is a strong relationship between temperature and the use of natural gas,
but only during the heating season.

Practical consideration can also be important. Sometimes, one or more
proximity measures are already in use in a particular field, and thus, others
will have answered the question of which proximity measures should be used.
Other times, the software package or clustering algorithm being used can
drastically limit the choices. If efficiency is a concern, then we may want to
choose a proximity measure that has a property, such as the triangle
inequality, that can be used to reduce the number of proximity calculations.
(See Exercise 25 .)

However, if common practice or practical restrictions do not dictate a choice,
then the proper choice of a proximity measure can be a time-consuming task
that requires careful consideration of both domain knowledge and the purpose
for which the measure is being used. A number of different similarity
measures may need to be evaluated to see which ones produce results that
make the most sense.

2.5 Bibliographic Notes
It is essential to understand the nature of the data that is being analyzed, and
at a fundamental level, this is the subject of measurement theory. In particular,
one of the initial motivations for defining types of attributes was to be precise
about which statistical operations were valid for what sorts of data. We have
presented the view of measurement theory that was initially described in a
classic paper by S. S. Stevens [112]. (Tables 2.2 and 2.3 are derived
from those presented by Stevens [113].) While this is the most common view
and is reasonably easy to understand and apply, there is, of course, much
more to measurement theory. An authoritative discussion can be found in a
three-volume series on the foundations of measurement theory [88, 94, 114].
Also of interest is a wide-ranging article by Hand [77], which discusses
measurement theory and statistics, and is accompanied by comments from
other researchers in the field. Numerous critiques and extensions of the
approach of Stevens have been made [66, 97, 117]. Finally, many books and
articles describe measurement issues for particular areas of science and

Data quality is a broad subject that spans every discipline that uses data.
Discussions of precision, bias, accuracy, and significant figures can be found
in many introductory science, engineering, and statistics textbooks. The view
of data quality as “fitness for use” is explained in more detail in the book by
Redman [103]. Those interested in data quality may also be interested in
MIT’s Information Quality (MITIQ) Program [95, 118]. However, the knowledge
needed to deal with specific data quality issues in a particular domain is often
best obtained by investigating the data quality practices of researchers in that

Aggregation is a less well-defined subject than many other preprocessing
tasks. However, aggregation is one of the main techniques used by the
database area of Online Analytical Processing (OLAP) [68, 76, 102]. There
has also been relevant work in the area of symbolic data analysis (Bock and
Diday [64]). One of the goals in this area is to summarize traditional record
data in terms of symbolic data objects whose attributes are more complex
than traditional attributes. Specifically, these attributes can have values that
are sets of values (categories), intervals, or sets of values with weights
(histograms). Another goal of symbolic data analysis is to be able to perform
clustering, classification, and other kinds of data analysis on data that consists
of symbolic data objects.

Sampling is a subject that has been well studied in statistics and related fields.
Many introductory statistics books, such as the one by Lindgren [90], have
some discussion about sampling, and entire books are devoted to the subject,
such as the classic text by Cochran [67]. A survey of sampling for data mining
is provided by Gu and Liu [74], while a survey of sampling for databases is
provided by Olken and Rotem [98]. There are a number of other data mining
and database-related sampling references that may be of interest, including
papers by Palmer and Faloutsos [100], Provost et al. [101], Toivonen [115],
and Zaki et al. [119].

In statistics, the traditional techniques that have been used for dimensionality
reduction are multidimensional scaling (MDS) (Borg and Groenen [65],
Kruskal and Uslaner [89]) and principal component analysis (PCA) (Jolliffe
[80]), which is similar to singular value decomposition (SVD) (Demmel [70]).
Dimensionality reduction is discussed in more detail in Appendix B.

Discretization is a topic that has been extensively investigated in data mining.
Some classification algorithms work only with categorical data, and
association analysis requires binary data, and thus, there is a significant

motivation to investigate how to best binarize or discretize continuous
attributes. For association analysis, we refer the reader to work by Srikant and
Agrawal [111], while some useful references for discretization in the area of
classification include work by Dougherty et al. [71], Elomaa and Rousu [72],
Fayyad and Irani [73], and Hussain et al. [78].

Feature selection is another topic well investigated in data mining. A broad
coverage of this topic is provided in a survey by Molina et al. [96] and two
books by Liu and Motada [91, 92]. Other useful papers include those by Blum
and Langley [63], Kohavi and John [87], and Liu et al. [93].

It is difficult to provide references for the subject of feature transformations
because practices vary from one discipline to another. Many statistics books
have a discussion of transformations, but typically the discussion is restricted
to a particular purpose, such as ensuring the normality of a variable or making
sure that variables have equal variance. We offer two references: Osborne
[99] and Tukey [116].

While we have covered some of the most commonly used distance and
similarity measures, there are hundreds of such measures and more are
being created all the time. As with so many other topics in this chapter, many
of these measures are specific to particular fields, e.g., in the area of time
series see papers by Kalpakis et al. [81] and Keogh and Pazzani [83].
Clustering books provide the best general discussions. In particular, see the
books by Anderberg [62], Jain and Dubes [79], Kaufman and Rousseeuw [82],
and Sneath and Sokal [109].

Information-based measures of similarity have become more popular lately
despite the computational difficulties and expense of calculating them. A good
introduction to information theory is provided by Cover and Thomas [69].
Computing the mutual information for continuous variables can be

straightforward if they follow a well-know distribution, such as Gaussian.
However, this is often not the case, and many techniques have been
developed. As one example, the article by Khan, et al. [85] compares various
methods in the context of comparing short time series. See also the
information and mutual information packages for R and Matlab. Mutual
information has been the subject of considerable recent attention due to paper
by Reshef, et al. [104, 105] that introduced an alternative measure, albeit one
based on mutual information, which was claimed to have superior properties.
Although this approach had some early support, e.g., [110], others have
pointed out various limitations [75, 86, 108].

Two popular books on the topic of kernel methods are [106] and [107]. The
latter also has a website with links to kernel-related materials [84]. In addition,
many current data mining, machine learning, and statistical learning textbooks
have some material about kernel methods. Further references for kernel
methods in the context of support vector machine classifiers are provided in
the bibliographic Notes of Section 4.9.4.

[62] M. R. Anderberg. Cluster Analysis for Applications. Academic Press, New

York, December 1973.

[63] A. Blum and P. Langley. Selection of Relevant Features and Examples in
Machine Learning. Artificial Intelligence, 97(1–2):245–271, 1997.

[64] H. H. Bock and E. Diday. Analysis of Symbolic Data: Exploratory Methods
for Extracting Statistical Information from Complex Data (Studies in
Classification, Data Analysis, and Knowledge Organization). Springer-
Verlag Telos, January 2000.

[65] I. Borg and P. Groenen. Modern Multidimensional Scaling—Theory and
Applications. Springer-Verlag, February 1997.

[66] N. R. Chrisman. Rethinking levels of measurement for cartography.
Cartography and Geographic Information Systems, 25(4):231–242, 1998.

[67] W. G. Cochran. Sampling Techniques. John Wiley & Sons, 3rd edition,
July 1977.

[68] E. F. Codd, S. B. Codd, and C. T. Smalley. Providing OLAP (On-line
Analytical Processing) to User- Analysts: An IT Mandate. White Paper, E.F.
Codd and Associates, 1993.

[69] T. M. Cover and J. A. Thomas. Elements of information theory. John
Wiley & Sons, 2012.

[70] J. W. Demmel. Applied Numerical Linear Algebra. Society for Industrial &
Applied Mathematics, September 1997.

[71] J. Dougherty, R. Kohavi, and M. Sahami. Supervised and Unsupervised
Discretization of Continuous Features. In Proc. of the 12th Intl. Conf. on
Machine Learning, pages 194–202, 1995.

[72] T. Elomaa and J. Rousu. General and Efficient Multisplitting of Numerical
Attributes. Machine Learning, 36(3):201–244, 1999.

[73] U. M. Fayyad and K. B. Irani. Multi-interval discretization of
continuousvalued attributes for classification learning. In Proc. 13th Int.
Joint Conf. on Artificial Intelligence, pages 1022–1027. Morgan Kaufman,

[74] F. H. Gaohua Gu and H. Liu. Sampling and Its Application in Data Mining:
A Survey. Technical Report TRA6/00, National University of Singapore,
Singapore, 2000.

[75] M. Gorfine, R. Heller, and Y. Heller. Comment on Detecting novel
associations in large data sets. Unpublished (available at http://emotion.
technion. ac. il/ gorfinm/files/science6. pdf on 11 Nov. 2012), 2012.

[76] J. Gray, S. Chaudhuri, A. Bosworth, A. Layman, D. Reichart, M.
Venkatrao, F. Pellow, and H. Pirahesh. Data Cube: A Relational
Aggregation Operator Generalizing Group-By, Cross-Tab, and Sub-Totals.
Journal Data Mining and Knowledge Discovery, 1(1): 29–53, 1997.

[77] D. J. Hand. Statistics and the Theory of Measurement. Journal of the
Royal Statistical Society: Series A (Statistics in Society), 159(3):445–492,

[78] F. Hussain, H. Liu, C. L. Tan, and M. Dash. TRC6/99: Discretization: an
enabling technique. Technical report, National University of Singapore,
Singapore, 1999.

[79] A. K. Jain and R. C. Dubes. Algorithms for Clustering Data. Prentice Hall
Advanced Reference Series. Prentice Hall, March 1988.

[80] I. T. Jolliffe. Principal Component Analysis. Springer Verlag, 2nd edition,
October 2002.

[81] K. Kalpakis, D. Gada, and V. Puttagunta. Distance Measures for Effective
Clustering of ARIMA Time-Series. In Proc. of the 2001 IEEE Intl. Conf. on
Data Mining, pages 273–280. IEEE Computer Society, 2001.

[82] L. Kaufman and P. J. Rousseeuw. Finding Groups in Data: An
Introduction to Cluster Analysis. Wiley Series in Probability and Statistics.
John Wiley and Sons, New York, November 1990.

[83] E. J. Keogh and M. J. Pazzani. Scaling up dynamic time warping for
datamining applications. In KDD, pages 285–289, 2000.

[84] Kernel Methods for Pattern Analysis Website. http://www.kernel-
methods.net/, 2014.

[85] S. Khan, S. Bandyopadhyay, A. R. Ganguly, S. Saigal, D. J. Erickson III,
V. Protopopescu, and G. Ostrouchov. Relative performance of mutual
information estimation methods for quantifying the dependence among
short and noisy data. Physical Review E, 76(2):026209, 2007.

[86] J. B. Kinney and G. S. Atwal. Equitability, mutual information, and the
maximal information coefficient. Proceedings of the National Academy of
Sciences, 2014.

[87] R. Kohavi and G. H. John. Wrappers for Feature Subset Selection.
Artificial Intelligence, 97(1–2):273–324, 1997.

[88] D. Krantz, R. D. Luce, P. Suppes, and A. Tversky. Foundations of
Measurements: Volume 1: Additive and polynomial representations.
Academic Press, New York, 1971.

[89] J. B. Kruskal and E. M. Uslaner. Multidimensional Scaling. Sage
Publications, August 1978.

[90] B. W. Lindgren. Statistical Theory. CRC Press, January 1993.

[91] H. Liu and H. Motoda, editors. Feature Extraction, Construction and
Selection: A Data Mining Perspective. Kluwer International Series in
Engineering and Computer Science, 453. Kluwer Academic Publishers,
July 1998.

[92] H. Liu and H. Motoda. Feature Selection for Knowledge Discovery and
Data Mining. Kluwer International Series in Engineering and Computer
Science, 454. Kluwer Academic Publishers, July 1998.

[93] H. Liu, H. Motoda, and L. Yu. Feature Extraction, Selection, and
Construction. In N. Ye, editor, The Handbook of Data Mining, pages 22–
41. Lawrence Erlbaum Associates, Inc., Mahwah, NJ, 2003.

[94] R. D. Luce, D. Krantz, P. Suppes, and A. Tversky. Foundations of
Measurements: Volume 3: Representation, Axiomatization, and
Invariance. Academic Press, New York, 1990.

[95] MIT Information Quality (MITIQ) Program. http://mitiq.mit.edu/, 2014.

[96] L. C. Molina, L. Belanche, and A. Nebot. Feature Selection Algorithms: A
Survey and Experimental Evaluation. In Proc. of the 2002 IEEE Intl. Conf.
on Data Mining, 2002.

[97] F. Mosteller and J. W. Tukey. Data analysis and regression: a second
course in statistics. Addison-Wesley, 1977.

[98] F. Olken and D. Rotem. Random Sampling from Databases—A Survey.
Statistics & Computing, 5(1):25–42, March 1995.

[99] J. Osborne. Notes on the Use of Data Transformations. Practical
Assessment, Research & Evaluation, 28(6), 2002.

[100] C. R. Palmer and C. Faloutsos. Density biased sampling: An improved
method for data mining and clustering. ACM SIGMOD Record, 29(2):82–
92, 2000.

[101] F. J. Provost, D. Jensen, and T. Oates. Efficient Progressive Sampling.
In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining,
pages 23–32, 1999.

[102] R. Ramakrishnan and J. Gehrke. Database Management Systems.
McGraw-Hill, 3rd edition, August 2002.

[103] T. C. Redman. Data Quality: The Field Guide. Digital Press, January

[104] D. Reshef, Y. Reshef, M. Mitzenmacher, and P. Sabeti. Equitability
analysis of the maximal information coefficient, with comparisons. arXiv
preprint arXiv:1301.6314, 2013.

[105] D. N. Reshef, Y. A. Reshef, H. K. Finucane, S. R. Grossman, G.
McVean, P. J. Turnbaugh, E. S. Lander, M. Mitzenmacher, and P. C.

Sabeti. Detecting novel associations in large data sets. science,
334(6062):1518–1524, 2011.

[106] B. Schölkopf and A. J. Smola. Learning with kernels: support vector
machines, regularization, optimization, and beyond. MIT press, 2002.

[107] J. Shawe-Taylor and N. Cristianini. Kernel methods for pattern analysis.
Cambridge university press, 2004.

[108] N. Simon and R. Tibshirani. Comment on” Detecting Novel Associations
In Large Data Sets” by Reshef Et Al, Science Dec 16, 2011. arXiv preprint
arXiv:1401.7645, 2014.

[109] P. H. A. Sneath and R. R. Sokal. Numerical Taxonomy. Freeman, San
Francisco, 1971.

[110] T. Speed. A correlation for the 21st century. Science, 334(6062):1502–
1503, 2011.

[111] R. Srikant and R. Agrawal. Mining Quantitative Association Rules in
Large Relational Tables. In Proc. of 1996 ACM-SIGMOD Intl. Conf. on
Management of Data, pages 1–12, Montreal, Quebec, Canada, August

[112] S. S. Stevens. On the Theory of Scales of Measurement. Science,
103(2684):677–680, June 1946.

[113] S. S. Stevens. Measurement. In G. M. Maranell, editor, Scaling: A
Sourcebook for Behavioral Scientists, pages 22–41. Aldine Publishing Co.,
Chicago, 1974.

[114] P. Suppes, D. Krantz, R. D. Luce, and A. Tversky. Foundations of
Measurements: Volume 2: Geometrical, Threshold, and Probabilistic
Representations. Academic Press, New York, 1989.

[115] H. Toivonen. Sampling Large Databases for Association Rules. In
VLDB96, pages 134–145. Morgan Kaufman, September 1996.

[116] J. W. Tukey. On the Comparative Anatomy of Transformations. Annals of
Mathematical Statistics, 28(3):602–632, September 1957.

[117] P. F. Velleman and L. Wilkinson. Nominal, ordinal, interval, and ratio
typologies are misleading. The American Statistician, 47(1):65–72, 1993.

[118] R. Y. Wang, M. Ziad, Y. W. Lee, and Y. R. Wang. Data Quality. The
Kluwer International Series on Advances in Database Systems, Volume
23. Kluwer Academic Publishers, January 2001.

[119] M. J. Zaki, S. Parthasarathy, W. Li, and M. Ogihara. Evaluation of
Sampling for Data Mining of Association Rules. Technical Report TR617,
Rensselaer Polytechnic Institute, 1996.

2.6 Exercises
1. In the initial example of Chapter 2 , the statistician says, “Yes, fields 2
and 3 are basically the same.” Can you tell from the three lines of sample data
that are shown why she says that?

2. Classify the following attributes as binary, discrete, or continuous. Also
classify them as qualitative (nominal or ordinal) or quantitative (interval or
ratio). Some cases may have more than one interpretation, so briefly indicate
your reasoning if you think there may be some ambiguity.

Example: Age in years. Answer: Discrete, quantitative, ratio

a. Time in terms of AM or PM.

b. Brightness as measured by a light meter.

c. Brightness as measured by people’s judgments.

d. Angles as measured in degrees between 0 and 360.

e. Bronze, Silver, and Gold medals as awarded at the Olympics.

f. Height above sea level.

g. Number of patients in a hospital.

h. ISBN numbers for books. (Look up the format on the Web.)

i. Ability to pass light in terms of the following values: opaque, translucent,

j. Military rank.

k. Distance from the center of campus.

l. Density of a substance in grams per cubic centimeter.

m. Coat check number. (When you attend an event, you can often give your
coat to someone who, in turn, gives you a number that you can use to
claim your coat when you leave.)

3. You are approached by the marketing director of a local company, who
believes that he has devised a foolproof way to measure customer
satisfaction. He explains his scheme as follows: “It’s so simple that I can’t
believe that no one has thought of it before. I just keep track of the number of
customer complaints for each product. I read in a data mining book that
counts are ratio attributes, and so, my measure of product satisfaction must
be a ratio attribute. But when I rated the products based on my new customer
satisfaction measure and showed them to my boss, he told me that I had
overlooked the obvious, and that my measure was worthless. I think that he
was just mad because our bestselling product had the worst satisfaction since
it had the most complaints. Could you help me set him straight?”

a. Who is right, the marketing director or his boss? If you answered, his
boss, what would you do to fix the measure of satisfaction?

b. What can you say about the attribute type of the original product
satisfaction attribute?

4. A few months later, you are again approached by the same marketing
director as in Exercise 3 . This time, he has devised a better approach to
measure the extent to which a customer prefers one product over other similar
products. He explains, “When we develop new products, we typically create
several variations and evaluate which one customers prefer. Our standard
procedure is to give our test subjects all of the product variations at one time
and then ask them to rank the product variations in order of preference.
However, our test subjects are very indecisive, especially when there are

more than two products. As a result, testing takes forever. I suggested that we
perform the comparisons in pairs and then use these comparisons to get the
rankings. Thus, if we have three product variations, we have the customers
compare variations 1 and 2, then 2 and 3, and finally 3 and 1. Our testing time
with my new procedure is a third of what it was for the old procedure, but the
employees conducting the tests complain that they cannot come up with a
consistent ranking from the results. And my boss wants the latest product
evaluations, yesterday. I should also mention that he was the person who
came up with the old product evaluation approach. Can you help me?”

a. Is the marketing director in trouble? Will his approach work for generating
an ordinal ranking of the product variations in terms of customer
preference? Explain.

b. Is there a way to fix the marketing director’s approach? More generally,
what can you say about trying to create an ordinal measurement scale
based on pairwise comparisons?

c. For the original product evaluation scheme, the overall rankings of each
product variation are found by computing its average over all test
subjects. Comment on whether you think that this is a reasonable
approach. What other approaches might you take?

5. Can you think of a situation in which identification numbers would be useful
for prediction?

6. An educational psychologist wants to use association analysis to analyze
test results. The test consists of 100 questions with four possible answers

a. How would you convert this data into a form suitable for association

b. In particular, what type of attributes would you have and how many of
them are there?

7. Which of the following quantities is likely to show more temporal
autocorrelation: daily rainfall or daily temperature? Why?

8. Discuss why a document-term matrix is an example of a data set that has
asymmetric discrete or asymmetric continuous features.

9. Many sciences rely on observation instead of (or in addition to) designed
experiments. Compare the data quality issues involved in observational
science with those of experimental science and data mining.

10. Discuss the difference between the precision of a measurement and the
terms single and double precision, as they are used in computer science,
typically to represent floating-point numbers that require 32 and 64 bits,

11. Give at least two advantages to working with data stored in text files
instead of in a binary format.

12. Distinguish between noise and outliers. Be sure to consider the following

a. Is noise ever interesting or desirable? Outliers?

b. Can noise objects be outliers?

c. Are noise objects always outliers?

d. Are outliers always noise objects?

e. Can noise make a typical value into an unusual one, or vice versa?

Algorithm 2.3 Algorithm for finding k-

nearest neighbors.

13. Consider the problem of finding the K-nearest neighbors of a data object.
A programmer designs Algorithm 2.3 for this task.

a. Describe the potential problems with this algorithm if there are duplicate
objects in the data set. Assume the distance function will return a
distance of 0 only for objects that are the same.

b. How would you fix this problem?

14. The following attributes are measured for members of a herd of Asian
elephants: weight, height, tusk length, trunk length, and ear area. Based on
these measurements, what sort of proximity measure from Section 2.4
would you use to compare or group these elephants? Justify your answer and
explain any special circumstances.

15. You are given a set of m objects that is divided into k groups, where the i
group is of size If the goal is to obtain a sample of size what is the
difference between the following two sampling schemes? (Assume sampling
with replacement.)

: for to number of data objects do

: Find the distances of the object to all other objects.

3: Sort these distances in decreasing order.

(Keep track of which object is associated with each distance.)

4: return the objects associated with the first k distances of the
sorted list

5: end for

1 i=1

2 ith


mi. n<m,

a. We randomly select elements from each group.

b. We randomly select n elements from the data set, without regard for the
group to which an object belongs.

16. Consider a document-term matrix, where is the frequency of the
word (term) in the document and m is the number of documents. Consider
the variable transformation that is defined by

where is the number of documents in which the term appears, which is
known as the document frequency of the term. This transformation is known
as the inverse document frequency transformation.

a. What is the effect of this transformation if a term occurs in one
document? In every document?

b. What might be the purpose of this transformation?

17. Assume that we apply a square root transformation to a ratio attribute x to
obtain the new attribute As part of your analysis, you identify an interval (a,
b) in which has a linear relationship to another attributey.

a. What is the corresponding interval (A, B) in terms of x ?

b. Give an equation that relates y to x.

18. This exercise compares and contrasts some similarity and distance

a. For binary data, the L1 distance corresponds to the Hamming distance;
that is, the number of bits that are different between two binary vectors.
The Jaccard similarity is a measure of the similarity between two binary


tfij ith

tfij′=tfij×logmdfi, (2.31)

dfi ith


vectors. Compute the Hamming distance and the Jaccard similarity
between the following two binary vectors.

b. Which approach, Jaccard or Hamming distance, is more similar to the
Simple Matching Coefficient, and which approach is more similar to the
cosine measure? Explain. (Note: The Hamming measure is a distance,
while the other three measures are similarities, but don’t let this confuse

c. Suppose that you are comparing how similar two organisms of different
species are in terms of the number of genes they share. Describe which
measure, Hamming or Jaccard, you think would be more appropriate for
comparing the genetic makeup of two organisms. Explain. (Assume that
each animal is represented as a binary vector, where each attribute is 1 if
a particular gene is present in the organism and 0 otherwise.)

d. If you wanted to compare the genetic makeup of two organisms of the
same species, e.g., two human beings, would you use the Hamming
distance, the Jaccard coefficient, or a different measure of similarity or
distance? Explain. (Note that two human beings share of the
same genes.)

19. For the following vectors, x and y, calculate the indicated similarity or
distance measures.

a. cosine, correlation, Euclidean

b. cosine, correlation, Euclidean, Jaccard

c. cosine, correlation, Euclidean

d. cosine, correlation, Jaccard



x=(1, 1, 1, 1), y=(2, 2, 2, 2)

x=(0, 1, 0, 1), y=(1, 0, 1, 0)

x=(0, −1, 0, 1), y=(1, 0, −1, 0)

x=(1, 1, 0, 1, 0, 1), y=(1, 1, 1, 0, 0, 1)

e. cosine, correlation

20. Here, we further explore the cosine and correlation measures.

a. What is the range of values possible for the cosine measure?

b. If two objects have a cosine measure of 1, are they identical? Explain.

c. What is the relationship of the cosine measure to correlation, if any?
(Hint: Look at statistical measures such as mean and standard deviation
in cases where cosine and correlation are the same and different.)

d. Figure 2.22(a) shows the relationship of the cosine measure to
Euclidean distance for 100,000 randomly generated points that have
been normalized to have an L2 length of 1. What general observation can
you make about the relationship between Euclidean distance and cosine
similarity when vectors have an L2 norm of 1?

Figure 2.22.
Graphs for Exercise 20 .

x=(2, −1, 0, 2, 0, −3), y=( −1, 1, −1, 0, 0, −1)

e. Figure 2.22(b) shows the relationship of correlation to Euclidean
distance for 100,000 randomly generated points that have been
standardized to have a mean of 0 and a standard deviation of 1. What
general observation can you make about the relationship between
Euclidean distance and correlation when the vectors have been
standardized to have a mean of 0 and a standard deviation of 1?

f. Derive the mathematical relationship between cosine similarity and
Euclidean distance when each data object has an L length of 1.

g. Derive the mathematical relationship between correlation and Euclidean
distance when each data point has been been standardized by
subtracting its mean and dividing by its standard deviation.

21. Show that the set difference metric given by

satisfies the metric axioms given on page 77 . A and B are sets and is
the set difference.

22. Discuss how you might map correlation values from the interval to
the interval [0, 1]. Note that the type of transformation that you use might
depend on the application that you have in mind. Thus, consider two
applications: clustering time series and predicting the behavior of one time
series given another.

23. Given a similarity measure with values in the interval [0, 1], describe two
ways to transform this similarity value into a dissimilarity value in the interval

24. Proximity is typically defined between a pair of objects.


d(A, B)=size(A−B)+size(B−A) (2.32)


[−1, 1]

[0, ∞].

a. Define two ways in which you might define the proximity among a group
of objects.

b. How might you define the distance between two sets of points in
Euclidean space?

c. How might you define the proximity between two sets of data objects?
(Make no assumption about the data objects, except that a proximity
measure is defined between any pair of objects.)

25. You are given a set of points s in Euclidean space, as well as the distance
of each point in s to a point x. (It does not matter if )

a. If the goal is to find all points within a specified distance of point
explain how you could use the triangle inequality and the already
calculated distances to x to potentially reduce the number of distance
calculations necessary? Hint: The triangle inequality,

can be rewritten as

b. In general, how would the distance between x and y affect the number of
distance calculations?

c. Suppose that you can find a small subset of points from the original
data set, such that every point in the data set is within a specified
distance of at least one of the points in and that you also have the
pairwise distance matrix for Describe a technique that uses this
information to compute, with a minimum of distance calculations, the set
of all points within a distance of of a specified point from the data set.

26. Show that 1 minus the Jaccard similarity is a distance measure between
two data objects, x and y, that satisfies the metric axioms given on page
77 . Specifically,

ε y, y≠x,

d(x, z)≤d(x, y)+d(y, x), d(x, y)≥d(x, z)−d(y, z).


ε S′,


d(x, y)=1−J(x, y).

27. Show that the distance measure defined as the angle between two data
vectors, x and y, satisfies the metric axioms given on page 77 . Specifically,

28. Explain why computing the proximity between two attributes is often
simpler than computing the similarity between two objects.

d(x, y)=arccos(cos(x, y)).

3 Classification: Basic Concepts and

Humans have an innate ability to classify things into
categories, e.g., mundane tasks such as filtering spam
email messages or more specialized tasks such as
recognizing celestial objects in telescope images (see
Figure 3.1 ). While manual classification often suffices
for small and simple data sets with only a few
attributes, larger and more complex data sets require
an automated solution.

Figure 3.1.
Classification of galaxies from telescope images taken
from the NASA website.

This chapter introduces the basic concepts of
classification and describes some of its key issues such
as model overfitting, model selection, and model
evaluation. While these topics are illustrated using a
classification technique known as decision tree
induction, most of the discussion in this chapter is also
applicable to other classification techniques, many of
which are covered in Chapter 4 .

3.1 Basic Concepts
Figure 3.2 illustrates the general idea behind classification. The data for a
classification task consists of a collection of instances (records). Each such
instance is characterized by the tuple ( , y), where is the set of attribute
values that describe the instance and y is the class label of the instance. The
attribute set can contain attributes of any type, while the class label y must
be categorical.

Figure 3.2.
A schematic illustration of a classification task.

A classification model is an abstract representation of the relationship
between the attribute set and the class label. As will be seen in the next two
chapters, the model can be represented in many ways, e.g., as a tree, a
probability table, or simply, a vector of real-valued parameters. More formally,
we can express it mathematically as a target function f that takes as input the
attribute set and produces an output corresponding to the predicted class
label. The model is said to classify an instance ( , y) correctly if .

Table 3.1 shows examples of attribute sets and class labels for various
classification tasks. Spam filtering and tumor identification are examples of
binary classification problems, in which each data instance can be categorized
into one of two classes. If the number of classes is larger than 2, as in the


galaxy classification example, then it is called a multiclass classification

Table 3.1. Examples of classification tasks.

Task Attribute set Class label

Spam filtering Features extracted from email message header
and content

spam or non-spam


Features extracted from magnetic resonance
imaging (MRI) scans

malignant or benign


Features extracted from telescope images elliptical, spiral, or

We illustrate the basic concepts of classification in this chapter with the
following two examples.

3.1. Example Vertebrate Classification
Table 3.2 shows a sample data set for classifying vertebrates into
mammals, reptiles, birds, fishes, and amphibians. The attribute set
includes characteristics of the vertebrate such as its body temperature,
skin cover, and ability to fly. The data set can also be used for a binary
classification task such as mammal classification, by grouping the reptiles,
birds, fishes, and amphibians into a single category called non-mammals.

Table 3.2. A sample data for the vertebrate classification problem.







Hibernates Class

human warm-


hair yes no no yes no mammal

3.2. Example Loan Borrower Classification
Consider the problem of predicting whether a loan borrower will repay the
loan or default on the loan payments. The data set used to build the


python cold-blooded scales no no no no yes reptile

salmon cold-blooded scales no yes no no no fish

whale warm-

hair yes yes no no no mammal

frog cold-blooded none no semi no yes yes amphibian

komodo cold-blooded scales no no no yes no reptile


bat warm-

hair yes no yes yes yes mammal

pigeon warm-

feathers no no yes yes no bird

cat warm-

fur yes no no yes no mammal

leopard cold-blooded scales yes yes no no no fish


turtle cold-blooded scales no semi no yes no reptile

penguin warm-

feathers no semi no yes no bird

porcupine warm-

quills yes no no yes yes mammal

eel cold-blooded scales no yes no no no fish

salamander cold-blooded none no semi no yes yes amphibian

classification model is shown in Table 3.3 . The attribute set includes
personal information of the borrower such as marital status and annual
income, while the class label indicates whether the borrower had defaulted
on the loan payments.

Table 3.3. A sample data for the loan borrower classification problem.

ID Home Owner Marital Status Annual Income Defaulted?

1 Yes Single 125000 No

2 No Married 100000 No

3 No Single 70000 No

4 Yes Married 120000 No

5 No Divorced 95000 Yes

6 No Single 60000 No

7 Yes Divorced 220000 No

8 No Single 85000 Yes

9 No Married 75000 No

10 No Single 90000 Yes

A classification model serves two important roles in data mining. First, it is
used as a predictive model to classify previously unlabeled instances. A
good classification model must provide accurate predictions with a fast
response time. Second, it serves as a descriptive model to identify the
characteristics that distinguish instances from different classes. This is
particularly useful for critical applications, such as medical diagnosis, where it

is insufficient to have a model that makes a prediction without justifying how it
reaches such a decision.

For example, a classification model induced from the vertebrate data set
shown in Table 3.2 can be used to predict the class label of the following

In addition, it can be used as a descriptive model to help determine
characteristics that define a vertebrate as a mammal, a reptile, a bird, a fish,
or an amphibian. For example, the model may identify mammals as warm-
blooded vertebrates that give birth to their young.

There are several points worth noting regarding the previous example. First,
although all the attributes shown in Table 3.2 are qualitative, there are no
restrictions on the type of attributes that can be used as predictor variables.
The class label, on the other hand, must be of nominal type. This
distinguishes classification from other predictive modeling tasks such as
regression, where the predicted value is often quantitative. More information
about regression can be found in Appendix D.

Another point worth noting is that not all attributes may be relevant to the
classification task. For example, the average length or weight of a vertebrate
may not be useful for classifying mammals, as these attributes can show
same value for both mammals and non-mammals. Such an attribute is
typically discarded during preprocessing. The remaining attributes might not
be able to distinguish the classes by themselves, and thus, must be used in








Hibernates Class


cold-blooded scales no no no yes yes ?

concert with other attributes. For instance, the Body Temperature attribute is
insufficient to distinguish mammals from other vertebrates. When it is used
together with Gives Birth, the classification of mammals improves significantly.
However, when additional attributes, such as Skin Cover are included, the
model becomes overly specific and no longer covers all mammals. Finding the
optimal combination of attributes that best discriminates instances from
different classes is the key challenge in building classification models.

3.2 General Framework for
Classification is the task of assigning labels to unlabeled data instances and a
classifier is used to perform such a task. A classifier is typically described in
terms of a model as illustrated in the previous section. The model is created
using a given a set of instances, known as the training set, which contains
attribute values as well as class labels for each instance. The systematic
approach for learning a classification model given a training set is known as a
learning algorithm. The process of using a learning algorithm to build a
classification model from the training data is known as induction. This
process is also often described as “learning a model” or “building a model.”
This process of applying a classification model on unseen test instances to
predict their class labels is known as deduction. Thus, the process of
classification involves two steps: applying a learning algorithm to training data
to learn a model, and then applying the model to assign labels to unlabeled
instances. Figure 3.3 illustrates the general framework for classification.

Figure 3.3.
General framework for building a classification model.

A classification technique refers to a general approach to classification, e.g.,
the decision tree technique that we will study in this chapter. This classification
technique like most others, consists of a family of related models and a
number of algorithms for learning these models. In Chapter 4 , we will study
additional classification techniques, including neural networks and support
vector machines.

A couple notes on terminology. First, the terms “classifier” and “model” are
often taken to be synonymous. If a classification technique builds a single,

global model, then this is fine. However, while every model defines a classifier,
not every classifier is defined by a single model. Some classifiers, such as k-
nearest neighbor classifiers, do not build an explicit model (Section 4.3 ),
while other classifiers, such as ensemble classifiers, combine the output of a
collection of models (Section 4.10 ). Second, the term “classifier” is often
used in a more general sense to refer to a classification technique. Thus, for
example, “decision tree classifier” can refer to the decision tree classification
technique or a specific classifier built using that technique. Fortunately, the
meaning of “classifier” is usually clear from the context.

In the general framework shown in Figure 3.3 , the induction and deduction
steps should be performed separately. In fact, as will be discussed later in
Section 3.6 , the training and test sets should be independent of each other
to ensure that the induced model can accurately predict the class labels of
instances it has never encountered before. Models that deliver such predictive
insights are said to have good generalization performance. The
performance of a model (classifier) can be evaluated by comparing the
predicted labels against the true labels of instances. This information can be
summarized in a table called a confusion matrix. Table 3.4 depicts the
confusion matrix for a binary classification problem. Each entry denotes the
number of instances from class i predicted to be of class j. For example, is
the number of instances from class 0 incorrectly predicted as class 1. The
number of correct predictions made by the model is and the number
of incorrect predictions is .

Table 3.4. Confusion matrix for a binary classification problem.

Predicted Class

Actual Class



Class=1 Class=0

Class=1 f11 f10

Although a confusion matrix provides the information needed to determine
how well a classification model performs, summarizing this information into a
single number makes it more convenient to compare the relative performance
of different models. This can be done using an evaluation metric such as
accuracy, which is computed in the following way:

Accuracy =

For binary classification problems, the accuracy of a model is given by

Error rate is another related metric, which is defined as follows for binary
classification problems:

The learning algorithms of most classification techniques are designed to
learn models that attain the highest accuracy, or equivalently, the lowest error
rate when applied to the test set. We will revisit the topic of model evaluation
in Section 3.6 .

Class=0 f01 f00

Accuracy=Number of correct predictionsTotal number of predictions. (3.1)

Accuracy=f11+f00f11+f10+f01+f00. (3.2)

Error rate=Number of wrong predictionsTotal number of predictions=f10+f01f11(3.3)

3.3 Decision Tree Classifier
This section introduces a simple classification technique known as the
decision tree classifier. To illustrate how a decision tree works, consider the
classification problem of distinguishing mammals from non-mammals using
the vertebrate data set shown in Table 3.2 . Suppose a new species is
discovered by scientists. How can we tell whether it is a mammal or a non-
mammal? One approach is to pose a series of questions about the
characteristics of the species. The first question we may ask is whether the
species is cold- or warm-blooded. If it is cold-blooded, then it is definitely not a
mammal. Otherwise, it is either a bird or a mammal. In the latter case, we
need to ask a follow-up question: Do the females of the species give birth to
their young? Those that do give birth are definitely mammals, while those that
do not are likely to be non-mammals (with the exception of egg-laying
mammals such as the platypus and spiny anteater).

The previous example illustrates how we can solve a classification problem by
asking a series of carefully crafted questions about the attributes of the test
instance. Each time we receive an answer, we could ask a follow-up question
until we can conclusively decide on its class label. The series of questions and
their possible answers can be organized into a hierarchical structure called a
decision tree. Figure 3.4 shows an example of the decision tree for the
mammal classification problem. The tree has three types of nodes:

A root node, with no incoming links and zero or more outgoing links.
Internal nodes, each of which has exactly one incoming link and two or
more outgoing links.
Leaf or terminal nodes, each of which has exactly one incoming link and
no outgoing links.

Every leaf node in the decision tree is associated with a class label. The non-
terminal nodes, which include the root and internal nodes, contain attribute
test conditions that are typically defined using a single attribute. Each
possible outcome of the attribute test condition is associated with exactly one
child of this node. For example, the root node of the tree shown in Figure
3.4 uses the attribute to define an attribute test condition
that has two outcomes, warm and cold, resulting in two child nodes.

Figure 3.4.
A decision tree for the mammal classification problem.

Given a decision tree, classifying a test instance is straightforward. Starting
from the root node, we apply its attribute test condition and follow the
appropriate branch based on the outcome of the test. This will lead us either
to another internal node, for which a new attribute test condition is applied, or
to a leaf node. Once a leaf node is reached, we assign the class label
associated with the node to the test instance. As an illustration, Figure 3.5

traces the path used to predict the class label of a flamingo. The path
terminates at a leaf node labeled as .

Figure 3.5.
Classifying an unlabeled vertebrate. The dashed lines represent the outcomes
of applying various attribute test conditions on the unlabeled vertebrate. The
vertebrate is eventually assigned to the class.

3.3.1 A Basic Algorithm to Build a
Decision Tree

Many possible decision trees that can be constructed from a particular data
set. While some trees are better than others, finding an optimal one is
computationally expensive due to the exponential size of the search space.
Efficient algorithms have been developed to induce a reasonably accurate,

albeit suboptimal, decision tree in a reasonable amount of time. These
algorithms usually employ a greedy strategy to grow the decision tree in a top-
down fashion by making a series of locally optimal decisions about which
attribute to use when partitioning the training data. One of the earliest method
is Hunt’s algorithm, which is the basis for many current implementations of
decision tree classifiers, including ID3, C4.5, and CART. This subsection
presents Hunt’s algorithm and describes some of the design issues that must
be considered when building a decision tree.

Hunt’s Algorithm
In Hunt’s algorithm, a decision tree is grown in a recursive fashion. The tree
initially contains a single root node that is associated with all the training
instances. If a node is associated with instances from more than one class, it
is expanded using an attribute test condition that is determined using a
splitting criterion. A child leaf node is created for each outcome of the
attribute test condition and the instances associated with the parent node are
distributed to the children based on the test outcomes. This node expansion
step can then be recursively applied to each child node, as long as it has
labels of more than one class. If all the instances associated with a leaf node
have identical class labels, then the node is not expanded any further. Each
leaf node is assigned a class label that occurs most frequently in the training
instances associated with the node.

To illustrate how the algorithm works, consider the training set shown in Table
3.3 for the loan borrower classification problem. Suppose we apply Hunt’s
algorithm to fit the training data. The tree initially contains only a single leaf
node as shown in Figure 3.6(a) . This node is labeled as Defaulted = No,
since the majority of the borrowers did not default on their loan payments. The
training error of this tree is 30% as three out of the ten training instances have

the class label . The leaf node can therefore be further
expanded because it contains training instances from more than one class.

Figure 3.6.
Hunt’s algorithm for building decision trees.

Let Home Owner be the attribute chosen to split the training instances. The
justification for choosing this attribute as the attribute test condition will be
discussed later. The resulting binary split on the Home Owner attribute is
shown in Figure 3.6(b) . All the training instances for which Home Owner =
Yes are propagated to the left child of the root node and the rest are
propagated to the right child. Hunt’s algorithm is then recursively applied to
each child. The left child becomes a leaf node labeled , since

Defaulted = Yes

Defaulted = No

all instances associated with this node have identical class label
. The right child has instances from each class label. Hence,

we split it further. The resulting subtrees after recursively expanding the right
child are shown in Figures 3.6(c) and (d) .

Hunt’s algorithm, as described above, makes some simplifying assumptions
that are often not true in practice. In the following, we describe these
assumptions and briefly discuss some of the possible ways for handling them.

1. Some of the child nodes created in Hunt’s algorithm can be empty if
none of the training instances have the particular attribute values. One
way to handle this is by declaring each of them as a leaf node with a
class label that occurs most frequently among the training instances
associated with their parent nodes.

2. If all training instances associated with a node have identical attribute
values but different class labels, it is not possible to expand this node
any further. One way to handle this case is to declare it a leaf node and
assign it the class label that occurs most frequently in the training
instances associated with this node.

Design Issues of Decision Tree Induction
Hunt’s algorithm is a generic procedure for growing decision trees in a greedy
fashion. To implement the algorithm, there are two key design issues that
must be addressed.

1. What is the splitting criterion? At each recursive step, an attribute
must be selected to partition the training instances associated with a
node into smaller subsets associated with its child nodes. The splitting
criterion determines which attribute is chosen as the test condition and

Defaulted = No

how the training instances should be distributed to the child nodes. This
will be discussed in Sections 3.3.2 and 3.3.3 .

2. What is the stopping criterion? The basic algorithm stops expanding
a node only when all the training instances associated with the node
have the same class labels or have identical attribute values. Although
these conditions are sufficient, there are reasons to stop expanding a
node much earlier even if the leaf node contains training instances from
more than one class. This process is called early termination and the
condition used to determine when a node should be stopped from
expanding is called a stopping criterion. The advantages of early
termination are discussed in Section 3.4 .

3.3.2 Methods for Expressing Attribute
Test Conditions

Decision tree induction algorithms must provide a method for expressing an
attribute test condition and its corresponding outcomes for different attribute

Binary Attributes

The test condition for a binary attribute generates two potential outcomes, as
shown in Figure 3.7 .

Figure 3.7.
Attribute test condition for a binary attribute.

Nominal Attributes

Since a nominal attribute can have many values, its attribute test condition
can be expressed in two ways, as a multiway split or a binary split as shown in
Figure 3.8 . For a multiway split (Figure 3.8(a) ), the number of outcomes
depends on the number of distinct values for the corresponding attribute. For
example, if an attribute such as marital status has three distinct values—
single, married, or divorced—its test condition will produce a three-way split. It
is also possible to create a binary split by partitioning all values taken by the
nominal attribute into two groups. For example, some decision tree
algorithms, such as CART, produce only binary splits by considering all

ways of creating a binary partition of k attribute values. Figure 3.8(b)
illustrates three different ways of grouping the attribute values for marital
status into two subsets.


Figure 3.8.
Attribute test conditions for nominal attributes.

Ordinal Attributes

Ordinal attributes can also produce binary or multi-way splits. Ordinal attribute
values can be grouped as long as the grouping does not violate the order
property of the attribute values. Figure 3.9 illustrates various ways of
splitting training records based on the Shirt Size attribute. The groupings
shown in Figures 3.9(a) and (b) preserve the order among the attribute
values, whereas the grouping shown in Figure 3.9(c) violates this property
because it combines the attribute values Small and Large into the same
partition while Medium and Extra Large are combined into another partition.

Figure 3.9.
Different ways of grouping ordinal attribute values.

Continuous Attributes

For continuous attributes, the attribute test condition can be expressed as a
comparison test (e.g., ) producing a binary split, or as a range query of the
form , for producing a multiway split. The difference
between these approaches is shown in Figure 3.10 . For the binary split,
any possible value v between the minimum and maximum attribute values in
the training data can be used for constructing the comparison test .
However, it is sufficient to only consider distinct attribute values in the training
set as candidate split positions. For the multiway split, any possible collection
of attribute value ranges can be used, as long as they are mutually exclusive
and cover the entire range of attribute values between the minimum and
maximum values observed in the training set. One approach for constructing
multiway splits is to apply the discretization strategies described in Section
2.3.6 on page 63. After discretization, a new ordinal value is assigned to
each discretized interval, and the attribute test condition is then defined using
this newly constructed ordinal attribute.

vi≤A<vi+1 i=1, …, k,


Figure 3.10.
Test condition for continuous attributes.

3.3.3 Measures for Selecting an
Attribute Test Condition

There are many measures that can be used to determine the goodness of an
attribute test condition. These measures try to give preference to attribute test
conditions that partition the training instances into purer subsets in the child
nodes, which mostly have the same class labels. Having purer nodes is useful
since a node that has all of its training instances from the same class does not
need to be expanded further. In contrast, an impure node containing training
instances from multiple classes is likely to require several levels of node
expansions, thereby increasing the depth of the tree considerably. Larger
trees are less desirable as they are more susceptible to model overfitting, a
condition that may degrade the classification performance on unseen
instances, as will be discussed in Section 3.4 . They are also difficult to
interpret and incur more training and test time as compared to smaller trees.

In the following, we present different ways of measuring the impurity of a node
and the collective impurity of its child nodes, both of which will be used to
identify the best attribute test condition for a node.

Impurity Measure for a Single Node
The impurity of a node measures how dissimilar the class labels are for the
data instances belonging to a common node. Following are examples of
measures that can be used to evaluate the impurity of a node t:

where pi(t) is the relative frequency of training instances that belong to class i
at node t, c is the total number of classes, and in entropy
calculations. All three measures give a zero impurity value if a node contains
instances from a single class and maximum impurity if the node has equal
proportion of instances from multiple classes.

Figure 3.11 compares the relative magnitude of the impurity measures
when applied to binary classification problems. Since there are only two
classes, . The horizontal axis p refers to the fraction of instances
that belong to one of the two classes. Observe that all three measures attain
their maximum value when the class distribution is uniform (i.e.,

) and minimum value when all the instances belong to a single
class (i.e., either or equals to 1). The following examples illustrate
how the values of the impurity measures vary as we alter the class

Entropy=−∑i=0c−1pi(t) log2pi(t), (3.4)

Gini index=1−∑i=0c−1pi(t)2, (3.5)

Classification error=1−maxi[pi(t)], (3.6)

0 log2 0=0


p0(t) p1(t)

Figure 3.11.
Comparison among the impurity measures for binary classification problems.

Node Count



Node Count



Node Count


N1 Gini=1−(0/6)2−(6/6)2=0

Class=0 Entropy=−(0/6) log2(0/6)−(6/6) log2(6/6)=0

Class=1 Error=1−max[0/6, 6/6]=0

N2 Gini=1−(1/6)2−(5/6)2=0.278

Class=0 Entropy=−(1/6) log2(1/6)−(5/6) log2(5/6)=0.650

Class=1 Error=1−max[1/6, 5/6]=0.167

N3 Gini=1−(3/6)2−(3/6)2=0.5

Class=0 Entropy=−(3/6) log2(3/6)−(3/6) log2(3/6)=1


Based on these calculations, node has the lowest impurity value, followed
by and . This example, along with Figure 3.11 , shows the
consistency among the impurity measures, i.e., if a node has lower
entropy than node , then the Gini index and error rate of will also be
lower than that of . Despite their agreement, the attribute chosen as
splitting criterion by the impurity measures can still be different (see Exercise
6 on page 187).

Collective Impurity of Child Nodes
Consider an attribute test condition that splits a node containing N training
instances into k children, , where every child node represents a
partition of the data resulting from one of the k outcomes of the attribute test
condition. Let be the number of training instances associated with a child
node , whose impurity value is . Since a training instance in the parent
node reaches node for a fraction of times, the collective impurity of
the child nodes can be computed by taking a weighted sum of the impurities
of the child nodes, as follows:

3.3. Example Weighted Entropy
Consider the candidate attribute test condition shown in Figures 3.12(a)
and (b) for the loan borrower classification problem. Splitting on the
Home Owner attribute will generate two child nodes

Class=1 Error=1−max[6/6, 3/6]=0.5

N2 N3

N2 N1

{v1, v2, ⋯ ,vk}

vj I(vj)

vj N(vj)/N

I(children)=∑j=1kN(vj)NI(vj), (3.7)

Figure 3.12.
Examples of candidate attribute test conditions.

whose weighted entropy can be calculated as follows:

Splitting on Marital Status, on the other hand, leads to three child nodes
with a weighted entropy given by

Thus, Marital Status has a lower weighted entropy than Home Owner.

Identifying the best attribute test condition
To determine the goodness of an attribute test condition, we need to compare
the degree of impurity of the parent node (before splitting) with the weighted
degree of impurity of the child nodes (after splitting). The larger their

I(Home Owner=yes)=03log203−33log233=0I(Home Owner=no)=
−37log237−47log247=0.985I(Home Owner=310×0+710×0.985=0.690

I(Marital Status=Single)=
−25log225−35log235=0.971I(Marital Status=Married)=
−03log203−33log233=0I(Marital Status=Divorced)=
−12log212−12log212=1.000I(Marital Status)=510×0.971+310×0+210×1=0.686

difference, the better the test condition. This difference, , also termed as the
gain in purity of an attribute test condition, can be defined as follows:

Figure 3.13.
Splitting criteria for the loan borrower classification problem using Gini index.

where I(parent) is the impurity of a node before splitting and I(children) is the
weighted impurity measure after splitting. It can be shown that the gain is non-
negative since for any reasonable measure such as those
presented above. The higher the gain, the purer are the classes in the child
nodes relative to the parent node. The splitting criterion in the decision tree
learning algorithm selects the attribute test condition that shows the maximum
gain. Note that maximizing the gain at a given node is equivalent to
minimizing the weighted impurity measure of its children since I(parent) is the
same for all candidate attribute test conditions. Finally, when entropy is used


Δ=I(parent)−I(children), (3.8)


as the impurity measure, the difference in entropy is commonly known as
information gain, .

In the following, we present illustrative approaches for identifying the best
attribute test condition given qualitative or quantitative attributes.

Splitting of Qualitative Attributes
Consider the first two candidate splits shown in Figure 3.12 involving
qualitative attributes and . The initial class
distribution at the parent node is (0.3, 0.7), since there are 3 instances of
class and 7 instances of class in the training data. Thus,

The information gains for Home Owner and Marital Status are each given by

The information gain for Marital Status is thus higher due to its lower weighted
entropy, which will thus be considered for splitting.

Binary Splitting of Qualitative Attributes
Consider building a decision tree using only binary splits and the Gini index as
the impurity measure. Figure 3.13 shows examples of four candidate
splitting criteria for the and attributes. Since there
are 3 borrowers in the training set who defaulted and 7 others who repaid their
loan (see Table in Figure 3.13 ), the Gini index of the parent node before
splitting is



Δinfo(Home Owner)=0.881−0.690=0.191Δinfo(Marital Status)=0.881−0.686=0.195

If is chosen as the splitting attribute, the Gini index for the child
nodes and are 0 and 0.490, respectively. The weighted average Gini
index for the children is

where the weights represent the proportion of training instances assigned to
each child. The gain using as splitting attribute is

. Similarly, we can apply a binary split on the
attribute. However, since is a nominal attribute with

three outcomes, there are three possible ways to group the attribute values
into a binary split. The weighted average Gini index of the children for each
candidate binary split is shown in Figure 3.13 . Based on these results,

and the last binary split using are clearly the best
candidates, since they both produce the lowest weighted average Gini index.
Binary splits can also be used for ordinal attributes, if the binary partitioning of
the attribute values does not violate the ordering property of the values.

Binary Splitting of Quantitative Attributes
Consider the problem of identifying the best binary split for
the preceding loan approval classification problem. As discussed previously,
even though can take any value between the minimum and maximum values
of annual income in the training set, it is sufficient to only consider the annual
income values observed in the training set as candidate split positions. For
each candidate , the training set is scanned once to count the number of
borrowers with annual income less than or greater than along with their class
proportions. We can then compute the Gini index at each candidate split


N1 N2



Annual Income≤τ



position and choose the that produces the lowest value. Computing the Gini
index at each candidate split position requires O(N) operations, where N is the
number of training instances. Since there are at most N possible candidates,
the overall complexity of this brute-force method is . It is possible to
reduce the complexity of this problem to O(N log N) by using a method
described as follows (see illustration in Figure 3.14 ). In this method, we
first sort the training instances based on their annual income, a one-time cost
that requires O(N log N) operations. The candidate split positions are given by
the midpoints between every two adjacent sorted values: $55,000, $65,000,
$72,500, and so on. For the first candidate, since none of the instances has
an annual income less than or equal to $55,000, the Gini index for the child
node with is equal to zero. In contrast, there are 3
training instances of class and instances of class No with annual
income greater than $55,000. The Gini index for this node is 0.420. The
weighted average Gini index for the first candidate split position, , is
equal to .

Figure 3.14.
Splitting continuous attributes.

For the next candidate, , the class distribution of its child nodes can
be obtained with a simple update of the distribution for the previous candidate.
This is because, as increases from $55,000 to $65,000, there is only one



Annual Income< $55,000




training instance affected by the change. By examining the class label of the
affected training instance, the new class distribution is obtained. For example,
as increases to $65,000, there is only one borrower in the training set, with
an annual income of $60,000, affected by this change. Since the class label
for the borrower is , the count for class increases from 0 to 1 (for

) and decreases from 7 to 6 (for
), as shown in Figure 3.14 . The distribution for the

class remains unaffected. The updated Gini index for this candidate split
position is 0.400.

This procedure is repeated until the Gini index for all candidates are found.
The best split position corresponds to the one that produces the lowest Gini
index, which occurs at . Since the Gini index at each candidate split
position can be computed in O(1) time, the complexity of finding the best split
position is O(N) once all the values are kept sorted, a one-time operation that
takes O(N log N) time. The overall complexity of this method is thus O(N log
N), which is much smaller than the time taken by the brute-force
method. The amount of computation can be further reduced by considering
only candidate split positions located between two adjacent sorted instances
with different class labels. For example, we do not need to consider candidate
split positions located between $60,000 and $75,000 because all three
instances with annual income in this range ($60,000, $70,000, and $75,000)
have the same class labels. Choosing a split position within this range only
increases the degree of impurity, compared to a split position located outside
this range. Therefore, the candidate split positions at and

can be ignored. Similarly, we do not need to consider the candidate
split positions at $87,500, $92,500, $110,000, $122,500, and $172,500
because they are located between two adjacent instances with the same
labels. This strategy reduces the number of candidate split positions to
consider from 9 to 2 (excluding the two boundary cases and



Annual Income≤$65,000
Annual Income>$65,000





Gain Ratio
One potential limitation of impurity measures such as entropy and Gini index
is that they tend to favor qualitative attributes with large number of distinct
values. Figure 3.12 shows three candidate attributes for partitioning the
data set given in Table 3.3 . As previously mentioned, the attribute

is a better choice than the attribute , because it provides a
larger information gain. However, if we compare them against ,
the latter produces the purest partitions with the maximum information gain,
since the weighted entropy and Gini index is equal to zero for its children. Yet,

is not a good attribute for splitting because it has a unique value
for each instance. Even though a test condition involving will
accurately classify every instance in the training data, we cannot use such a
test condition on new test instances with values that haven’t been
seen before during training. This example suggests having a low impurity
value alone is insufficient to find a good attribute test condition for a node. As
we will see later in Section 3.4 , having more number of child nodes can
make a decision tree more complex and consequently more susceptible to
overfitting. Hence, the number of children produced by the splitting attribute
should also be taken into consideration while deciding the best attribute test

There are two ways to overcome this problem. One way is to generate only
binary decision trees, thus avoiding the difficulty of handling attributes with
varying number of partitions. This strategy is employed by decision tree
classifiers such as CART. Another way is to modify the splitting criterion to
take into account the number of partitions produced by the attribute. For
example, in the C4.5 decision tree algorithm, a measure known as gain ratio
is used to compensate for attributes that produce a large number of child
nodes. This measure is computed as follows:

where is the number of instances assigned to node and k is the total
number of splits. The split information measures the entropy of splitting a
node into its child nodes and evaluates if the split results in a larger number of
equally-sized child nodes or not. For example, if every partition has the same
number of instances, then and the split information would be
equal to log k. Thus, if an attribute produces a large number of splits, its split
information is also large, which in turn, reduces the gain ratio.

3.4. Example Gain Ratio
Consider the data set given in Exercise 2 on page 185. We want to select
the best attribute test condition among the following three attributes:

, , and . The entropy before splitting is

If is used as attribute test condition:

If is used as attribute test condition:

Finally, if is used as attribute test condition:

Gain ratio=ΔinfoSplit Info=Entropy(Parent)−∑i=1kN(vi)NEntropy(vi)


N(vi) vi



]×2=0.971Gain Ratio=1−0.971−1020log21020−1020log21020=0.0291=0.029

]=0.380Gain Ratio=1−0.380−420log2420−820log2820−820log2820=0.6201.52

Thus, even though has the highest information gain, its gain
ratio is lower than since it produces a larger number of splits.

3.3.4 Algorithm for Decision Tree

Algorithm 3.1 presents a pseudocode for decision tree induction algorithm.
The input to this algorithm is a set of training instances E along with the
attribute set F . The algorithm works by recursively selecting the best attribute
to split the data (Step 7) and expanding the nodes of the tree (Steps 11 and
12) until the stopping criterion is met (Step 1). The details of this algorithm are
explained below.

1. The function extends the decision tree by creating a new
node. A node in the decision tree either has a test condition, denoted
as node.test cond, or a class label, denoted as node.label.

2. The function determines the attribute test condition
for partitioning the training instances associated with a node. The
splitting attribute chosen depends on the impurity measure used. The
popular measures include entropy and the Gini index.

3. The function determines the class label to be assigned to a
leaf node. For each leaf node t, let denote the fraction of training
instances from class i associated with the node t. The label assigned to

]×20=0Gain Ratio=1−0−120log2120×20=14.32=0.23


the leaf node is typically the one that occurs most frequently in the
training instances that are associated with this node.

Algorithm 3.1 A skeleton decision tree
induction algorithm.

where the argmax operator returns the class i that maximizes .
Besides providing the information needed to determine the class label

leaf.label=argmaxi p(i|t), (3.10)


of a leaf node, can also be used as a rough estimate of the
probability that an instance assigned to the leaf node t belongs to class
i. Sections 4.11.2 and 4.11.4 in the next chapter describe how
such probability estimates can be used to determine the performance
of a decision tree under different cost functions.

4. The function is used to terminate the tree-growing
process by checking whether all the instances have identical class
label or attribute values. Since decision tree classifiers employ a top-
down, recursive partitioning approach for building a model, the number
of training instances associated with a node decreases as the depth of
the tree increases. As a result, a leaf node may contain too few training
instances to make a statistically significant decision about its class
label. This is known as the data fragmentation problem. One way to
avoid this problem is to disallow splitting of a node when the number of
instances associated with the node fall below a certain threshold. A
more systematic way to control the size of a decision tree (number of
leaf nodes) will be discussed in Section 3.5.4 .

3.3.5 Example Application: Web Robot

Consider the task of distinguishing the access patterns of web robots from
those generated by human users. A web robot (also known as a web crawler)
is a software program that automatically retrieves files from one or more
websites by following the hyperlinks extracted from an initial set of seed
URLs. These programs have been deployed for various purposes, from
gathering web pages on behalf of search engines to more malicious activities
such as spamming and committing click frauds in online advertisements.


Figure 3.15.
Input data for web robot detection.

The web robot detection problem can be cast as a binary classification task.
The input data for the classification task is a web server log, a sample of
which is shown in Figure 3.15(a) . Each line in the log file corresponds to a
request made by a client (i.e., a human user or a web robot) to the web
server. The fields recorded in the web log include the client’s IP address,
timestamp of the request, URL of the requested file, size of the file, and user
agent, which is a field that contains identifying information about the client.

For human users, the user agent field specifies the type of web browser or
mobile device used to fetch the files, whereas for web robots, it should
technically contain the name of the crawler program. However, web robots
may conceal their true identities by declaring their user agent fields to be
identical to known browsers. Therefore, user agent is not a reliable field to
detect web robots.

The first step toward building a classification model is to precisely define a
data instance and associated attributes. A simple approach is to consider
each log entry as a data instance and use the appropriate fields in the log file
as its attribute set. This approach, however, is inadequate for several reasons.
First, many of the attributes are nominal-valued and have a wide range of
domain values. For example, the number of unique client IP addresses, URLs,
and referrers in a log file can be very large. These attributes are undesirable
for building a decision tree because their split information is extremely high
(see Equation (3.9) ). In addition, it might not be possible to classify test
instances containing IP addresses, URLs, or referrers that are not present in
the training data. Finally, by considering each log entry as a separate data
instance, we disregard the sequence of web pages retrieved by the client—a
critical piece of information that can help distinguish web robot accesses from
those of a human user.

A better alternative is to consider each web session as a data instance. A web
session is a sequence of requests made by a client during a given visit to the
website. Each web session can be modeled as a directed graph, in which the
nodes correspond to web pages and the edges correspond to hyperlinks
connecting one web page to another. Figure 3.15(b) shows a graphical
representation of the first web session given in the log file. Every web session
can be characterized using some meaningful attributes about the graph that
contain discriminatory information. Figure 3.15(c) shows some of the
attributes extracted from the graph, including the depth and breadth of its

corresponding tree rooted at the entry point to the website. For example, the
depth and breadth of the tree shown in Figure 3.15(b) are both equal to

The derived attributes shown in Figure 3.15(c) are more informative than
the original attributes given in the log file because they characterize the
behavior of the client at the website. Using this approach, a data set
containing 2916 instances was created, with equal numbers of sessions due
to web robots (class 1) and human users (class 0). 10% of the data were
reserved for training while the remaining 90% were used for testing. The
induced decision tree is shown in Figure 3.16 , which has an error rate
equal to 3.8% on the training set and 5.3% on the test set. In addition to its
low error rate, the tree also reveals some interesting properties that can help
discriminate web robots from human users:

1. Accesses by web robots tend to be broad but shallow, whereas
accesses by human users tend to be more focused (narrow but deep).

2. Web robots seldom retrieve the image pages associated with a web

3. Sessions due to web robots tend to be long and contain a large number
of requested pages.

4. Web robots are more likely to make repeated requests for the same
web page than human users since the web pages retrieved by human
users are often cached by the browser.

3.3.6 Characteristics of Decision Tree

The following is a summary of the important characteristics of decision tree
induction algorithms.

1. Applicability: Decision trees are a nonparametric approach for
building classification models. This approach does not require any prior
assumption about the probability distribution governing the class and
attributes of the data, and thus, is applicable to a wide variety of data
sets. It is also applicable to both categorical and continuous data
without requiring the attributes to be transformed into a common
representation via binarization, normalization, or standardization.
Unlike some binary classifiers described in Chapter 4 , it can also
deal with multiclass problems without the need to decompose them into
multiple binary classification tasks. Another appealing feature of
decision tree classifiers is that the induced trees, especially the shorter
ones, are relatively easy to interpret. The accuracies of the trees are
also quite comparable to other classification techniques for many
simple data sets.

2. Expressiveness: A decision tree provides a universal representation
for discrete-valued functions. In other words, it can encode any function
of discrete-valued attributes. This is because every discrete-valued
function can be represented as an assignment table, where every
unique combination of discrete attributes is assigned a class label.
Since every combination of attributes can be represented as a leaf in
the decision tree, we can always find a decision tree whose label
assignments at the leaf nodes matches with the assignment table of
the original function. Decision trees can also help in providing compact
representations of functions when some of the unique combinations of
attributes can be represented by the same leaf node. For example,
Figure 3.17 shows the assignment table of the Boolean function

involving four binary attributes, resulting in a total of
possible assignments. The tree shown in Figure 3.17 shows


a compressed encoding of this assignment table. Instead of requiring a
fully-grown tree with 16 leaf nodes, it is possible to encode the function
using a simpler tree with only 7 leaf nodes. Nevertheless, not all
decision trees for discrete-valued attributes can be simplified. One
notable example is the parity function, whose value is 1 when there is
an even number of true values among its Boolean attributes, and 0
otherwise. Accurate modeling of such a function requires a full decision
tree with nodes, where d is the number of Boolean attributes (see
Exercise 1 on page 185).


Figure 3.16.
Decision tree model for web robot detection.

Figure 3.17.
Decision tree for the Boolean function .

3. Computational Efficiency: Since the number of possible decision
trees can be very large, many decision tree algorithms employ a
heuristic-based approach to guide their search in the vast hypothesis
space. For example, the algorithm presented in Section 3.3.4 uses
a greedy, top-down, recursive partitioning strategy for growing a
decision tree. For many data sets, such techniques quickly construct a
reasonably good decision tree even when the training set size is very
large. Furthermore, once a decision tree has been built, classifying a
test record is extremely fast, with a worst-case complexity of O(w),
where w is the maximum depth of the tree.

4. Handling Missing Values: A decision tree classifier can handle
missing attribute values in a number of ways, both in the training and
the test sets. When there are missing values in the test set, the
classifier must decide which branch to follow if the value of a splitting


node attribute is missing for a given test instance. One approach,
known as the probabilistic split method, which is employed by the
C4.5 decision tree classifier, distributes the data instance to every child
of the splitting node according to the probability that the missing
attribute has a particular value. In contrast, the CART algorithm uses
the surrogate split method, where the instance whose splitting
attribute value is missing is assigned to one of the child nodes based
on the value of another non-missing surrogate attribute whose splits
most resemble the partitions made by the missing attribute. Another
approach, known as the separate class method is used by the CHAID
algorithm, where the missing value is treated as a separate categorical
value distinct from other values of the splitting attribute. Figure 3.18
shows an example of the three different ways for handling missing
values in a decision tree classifier. Other strategies for dealing with
missing values are based on data preprocessing, where the instance
with missing value is either imputed with the mode (for categorical
attribute) or mean (for continuous attribute) value or discarded before
the classifier is trained.

Figure 3.18.
Methods for handling missing attribute values in decision tree classifier.

During training, if an attribute v has missing values in some of the
training instances associated with a node, we need a way to measure
the gain in purity if v is used for splitting. One simple way is to exclude
instances with missing values of v in the counting of instances
associated with every child node, generated for every possible
outcome of v.Further, if v is chosen as the attribute test condition at a
node, training instances with missing values of v can be propagated to
the child nodes using any of the methods described above for handling
missing values in test instances.

5. Handling Interactions among Attributes: Attributes are considered
interacting if they are able to distinguish between classes when used
together, but individually they provide little or no information. Due to the
greedy nature of the splitting criteria in decision trees, such attributes
could be passed over in favor of other attributes that are not as useful.
This could result in more complex decision trees than necessary.
Hence, decision trees can perform poorly when there are interactions
among attributes.
To illustrate this point, consider the three-dimensional data shown in
Figure 3.19(a) , which contains 2000 data points from one of two
classes, denoted as and in the diagram. Figure 3.19(b) shows
the distribution of the two classes in the two-dimensional space
involving attributes X and Y , which is a noisy version of the XOR
Boolean function. We can see that even though the two classes are
well-separated in this two-dimensional space, neither of the two
attributes contain sufficient information to distinguish between the two
classes when used alone. For example, the entropies of the following
attribute test conditions: and , are close to 1, indicating that
neither X nor Y provide any reduction in the impurity measure when
used individually. X and Y thus represent a case of interaction among
attributes. The data set also contains a third attribute, Z, in which both
classes are distributed uniformly, as shown in Figures 3.19(c) and

+ ∘

X≤10 Y≤10

3.19(d) , and hence, the entropy of any split involving Z is close to 1.
As a result, Z is as likely to be chosen for splitting as the interacting but
useful attributes, X and Y . For further illustration of this issue, readers
are referred to Example 3.7 in Section 3.4.1 and Exercise 7 at
the end of this chapter.

Figure 3.19.
Example of a XOR data involving X and Y , along with an irrelevant
attribute Z.

6. Handling Irrelevant Attributes: An attribute is irrelevant if it is not
useful for the classification task. Since irrelevant attributes are poorly
associated with the target class labels, they will provide little or no gain
in purity and thus will be passed over by other more relevant features.
Hence, the presence of a small number of irrelevant attributes will not
impact the decision tree construction process. However, not all
attributes that provide little to no gain are irrelevant (see Figure
3.19 ). Hence, if the classification problem is complex (e.g., involving
interactions among attributes) and there are a large number of
irrelevant attributes, then some of these attributes may be accidentally
chosen during the tree-growing process, since they may provide a
better gain than a relevant attribute just by random chance. Feature
selection techniques can help to improve the accuracy of decision trees
by eliminating the irrelevant attributes during preprocessing. We will
investigate the issue of too many irrelevant attributes in Section
3.4.1 .

7. Handling Redundant Attributes: An attribute is redundant if it is
strongly correlated with another attribute in the data. Since redundant
attributes show similar gains in purity if they are selected for splitting,
only one of them will be selected as an attribute test condition in the
decision tree algorithm. Decision trees can thus handle the presence of
redundant attributes.

8. Using Rectilinear Splits: The test conditions described so far in this
chapter involve using only a single attribute at a time. As a
consequence, the tree-growing procedure can be viewed as the
process of partitioning the attribute space into disjoint regions until
each region contains records of the same class. The border between
two neighboring regions of different classes is known as a decision
boundary. Figure 3.20 shows the decision tree as well as the
decision boundary for a binary classification problem. Since the test
condition involves only a single attribute, the decision boundaries are

rectilinear; i.e., parallel to the coordinate axes. This limits the
expressiveness of decision trees in representing decision boundaries of
data sets with continuous attributes. Figure 3.21 shows a two-
dimensional data set involving binary classes that cannot be perfectly
classified by a decision tree whose attribute test conditions are defined
based on single attributes. The binary classes in the data set are
generated from two skewed Gaussian distributions, centered at (8,8)
and (12,12), respectively. The true decision boundary is represented by
the diagonal dashed line, whereas the rectilinear decision boundary
produced by the decision tree classifier is shown by the thick solid line.
In contrast, an oblique decision tree may overcome this limitation by
allowing the test condition to be specified using more than one
attribute. For example, the binary classification data shown in Figure
3.21 can be easily represented by an oblique decision tree with a
single root node with test condition

Figure 3.20.


Example of a decision tree and its decision boundaries for a two-
dimensional data set.

Figure 3.21.
Example of data set that cannot be partitioned optimally using a
decision tree with single attribute test conditions. The true decision
boundary is shown by the dashed line.

Although an oblique decision tree is more expressive and can produce
more compact trees, finding the optimal test condition is
computationally more expensive.

9. Choice of Impurity Measure: It should be noted that the choice of
impurity measure often has little effect on the performance of decision
tree classifiers since many of the impurity measures are quite
consistent with each other, as shown in Figure 3.11 on page 129.
Instead, the strategy used to prune the tree has a greater impact on the
final tree than the choice of impurity measure.

3.4 Model Overfitting
Methods presented so far try to learn classification models that show the
lowest error on the training set. However, as we will show in the following
example, even if a model fits well over the training data, it can still show poor
generalization performance, a phenomenon known as model overfitting.

Figure 3.22.
Examples of training and test sets of a two-dimensional classification problem.

Figure 3.23.
Effect of varying tree size (number of leaf nodes) on training and test errors.

3.5. Example Overfitting and Underfitting of
Decision Trees
Consider the two-dimensional data set shown in Figure 3.22(a) . The
data set contains instances that belong to two separate classes,
represented as and , respectively, where each class has 5400
instances. All instances belonging to the class were generated from a
uniform distribution. For the class, 5000 instances were generated from
a Gaussian distribution centered at (10,10) with unit variance, while the
remaining 400 instances were sampled from the same uniform distribution
as the class. We can see from Figure 3.22(a) that the class can be
largely distinguished from the class by drawing a circle of appropriate
size centered at (10,10). To learn a classifier using this two-dimensional
data set, we randomly sampled 10% of the data for training and used the
remaining 90% for testing. The training set, shown in Figure 3.22(b) ,
looks quite representative of the overall data. We used Gini index as the

+ ∘


∘ +

impurity measure to construct decision trees of increasing sizes (number of
leaf nodes), by recursively expanding a node into child nodes till every leaf
node was pure, as described in Section 3.3.4 .

Figure 3.23(a) shows changes in the training and test error rates as the
size of the tree varies from 1 to 8. Both error rates are initially large when
the tree has only one or two leaf nodes. This situation is known as model
underfitting. Underfitting occurs when the learned decision tree is too
simplistic, and thus, incapable of fully representing the true relationship
between the attributes and the class labels. As we increase the tree size
from 1 to 8, we can observe two effects. First, both the error rates
decrease since larger trees are able to represent more complex decision
boundaries. Second, the training and test error rates are quite close to
each other, which indicates that the performance on the training set is fairly
representative of the generalization performance. As we further increase
the size of the tree from 8 to 150, the training error continues to steadily
decrease till it eventually reaches zero, as shown in Figure 3.23(b) .
However, in a striking contrast, the test error rate ceases to decrease any
further beyond a certain tree size, and then it begins to increase. The
training error rate thus grossly under-estimates the test error rate once the
tree becomes too large. Further, the gap between the training and test
error rates keeps on widening as we increase the tree size. This behavior,
which may seem counter-intuitive at first, can be attributed to the
phenomena of model overfitting.

3.4.1 Reasons for Model Overfitting

Model overfitting is the phenomena where, in the pursuit of minimizing the
training error rate, an overly complex model is selected that captures specific

patterns in the training data but fails to learn the true nature of relationships
between attributes and class labels in the overall data. To illustrate this,
Figure 3.24 shows decision trees and their corresponding decision
boundaries (shaded rectangles represent regions assigned to the class) for
two trees of sizes 5 and 50. We can see that the decision tree of size 5
appears quite simple and its decision boundaries provide a reasonable
approximation to the ideal decision boundary, which in this case corresponds
to a circle centered around the Gaussian distribution at (10, 10). Although its
training and test error rates are non-zero, they are very close to each other,
which indicates that the patterns learned in the training set should generalize
well over the test set. On the other hand, the decision tree of size 50 appears
much more complex than the tree of size 5, with complicated decision
boundaries. For example, some of its shaded rectangles (assigned the
class) attempt to cover narrow regions in the input space that contain only one
or two training instances. Note that the prevalence of instances in such
regions is highly specific to the training set, as these regions are mostly
dominated by – instances in the overall data. Hence, in an attempt to perfectly
fit the training data, the decision tree of size 50 starts fine tuning itself to
specific patterns in the training data, leading to poor performance on an
independently chosen test set.



+ +

Figure 3.24.
Decision trees with different model complexities.

Figure 3.25.
Performance of decision trees using 20% data for training (twice the original
training size).

There are a number of factors that influence model overfitting. In the following,
we provide brief descriptions of two of the major factors: limited training size
and high model complexity. Though they are not exhaustive, the interplay
between them can help explain most of the common model overfitting
phenomena in real-world applications.

Limited Training Size
Note that a training set consisting of a finite number of instances can only
provide a limited representation of the overall data. Hence, it is possible that
the patterns learned from a training set do not fully represent the true patterns
in the overall data, leading to model overfitting. In general, as we increase the
size of a training set (number of training instances), the patterns learned from
the training set start resembling the true patterns in the overall data. Hence,

the effect of overfitting can be reduced by increasing the training size, as
illustrated in the following example.

3.6 Example Effect of Training Size
Suppose that we use twice the number of training instances than what we
had used in the experiments conducted in Example 3.5 . Specifically, we
use 20% data for training and use the remainder for testing. Figure
3.25(b) shows the training and test error rates as the size of the tree is
varied from 1 to 150. There are two major differences in the trends shown
in this figure and those shown in Figure 3.23(b) (using only 10% of the
data for training). First, even though the training error rate decreases with
increasing tree size in both figures, its rate of decrease is much smaller
when we use twice the training size. Second, for a given tree size, the gap
between the training and test error rates is much smaller when we use
twice the training size. These differences suggest that the patterns learned
using 20% of data for training are more generalizable than those learned
using 10% of data for training.

Figure 3.25(a) shows the decision boundaries for the tree of size 50,
learned using 20% of data for training. In contrast to the tree of the same
size learned using 10% data for training (see Figure 3.24(d) ), we can
see that the decision tree is not capturing specific patterns of noisy
instances in the training set. Instead, the high model complexity of 50 leaf
nodes is being effectively used to learn the boundaries of the instances
centered at (10, 10).

High Model Complexity
Generally, a more complex model has a better ability to represent complex
patterns in the data. For example, decision trees with larger number of leaf



nodes can represent more complex decision boundaries than decision trees
with fewer leaf nodes. However, an overly complex model also has a tendency
to learn specific patterns in the training set that do not generalize well over
unseen instances. Models with high complexity should thus be judiciously
used to avoid overfitting.

One measure of model complexity is the number of “parameters” that need to
be inferred from the training set. For example, in the case of decision tree
induction, the attribute test conditions at internal nodes correspond to the
parameters of the model that need to be inferred from the training set. A
decision tree with larger number of attribute test conditions (and consequently
more leaf nodes) thus involves more “parameters” and hence is more

Given a class of models with a certain number of parameters, a learning
algorithm attempts to select the best combination of parameter values that
maximizes an evaluation metric (e.g., accuracy) over the training set. If the
number of parameter value combinations (and hence the complexity) is large,
the learning algorithm has to select the best combination from a large number
of possibilities, using a limited training set. In such cases, there is a high
chance for the learning algorithm to pick a spurious combination of
parameters that maximizes the evaluation metric just by random chance. This
is similar to the multiple comparisons problem (also referred as multiple
testing problem) in statistics.

As an illustration of the multiple comparisons problem, consider the task of
predicting whether the stock market will rise or fall in the next ten trading days.
If a stock analyst simply makes random guesses, the probability that her
prediction is correct on any trading day is 0.5. However, the probability that
she will predict correctly at least nine out of ten times is

which is extremely low.

Suppose we are interested in choosing an investment advisor from a pool of
200 stock analysts. Our strategy is to select the analyst who makes the most
number of correct predictions in the next ten trading days. The flaw in this
strategy is that even if all the analysts make their predictions in a random
fashion, the probability that at least one of them makes at least nine correct
predictions is

which is very high. Although each analyst has a low probability of predicting at
least nine times correctly, considered together, we have a high probability of
finding at least one analyst who can do so. However, there is no guarantee in
the future that such an analyst will continue to make accurate predictions by
random guessing.

How does the multiple comparisons problem relate to model overfitting? In the
context of learning a classification model, each combination of parameter
values corresponds to an analyst, while the number of training instances
corresponds to the number of days. Analogous to the task of selecting the
best analyst who makes the most accurate predictions on consecutive days,
the task of a learning algorithm is to select the best combination of parameters
that results in the highest accuracy on the training set. If the number of
parameter combinations is large but the training size is small, it is highly likely
for the learning algorithm to choose a spurious parameter combination that
provides high training accuracy just by random chance. In the following
example, we illustrate the phenomena of overfitting due to multiple
comparisons in the context of decision tree induction.



Figure 3.26.
Example of a two-dimensional (X-Y) data set.

Figure 3.27.

Training and test error rates illustrating the effect of multiple comparisons
problem on model overfitting.

3.7. Example Multiple Comparisons and
Consider the two-dimensional data set shown in Figure 3.26 containing
500 and 500 instances, which is similar to the data shown in Figure
3.19 . In this data set, the distributions of both classes are well-separated
in the two-dimensional (XY) attribute space, but none of the two attributes
(X or Y) are sufficiently informative to be used alone for separating the two
classes. Hence, splitting the data set based on any value of an X or Y
attribute will provide close to zero reduction in an impurity measure.
However, if X and Y attributes are used together in the splitting criterion
(e.g., splitting X at 10 and Y at 10), the two classes can be effectively

+ ∘

Figure 3.28.
Decision tree with 6 leaf nodes using X and Y as attributes. Splits have
been numbered from 1 to 5 in order of other occurrence in the tree.

Figure 3.27(a) shows the training and test error rates for learning
decision trees of varying sizes, when 30% of the data is used for training
and the remainder of the data for testing. We can see that the two classes
can be separated using a small number of leaf nodes. Figure 3.28
shows the decision boundaries for the tree with six leaf nodes, where the
splits have been numbered according to their order of appearance in the
tree. Note that the even though splits 1 and 3 provide trivial gains, their
consequent splits (2, 4, and 5) provide large gains, resulting in effective
discrimination of the two classes.

Assume we add 100 irrelevant attributes to the two-dimensional X-Y data.
Learning a decision tree from this resultant data will be challenging
because the number of candidate attributes to choose for splitting at every
internal node will increase from two to 102. With such a large number of
candidate attribute test conditions to choose from, it is quite likely that
spurious attribute test conditions will be selected at internal nodes because
of the multiple comparisons problem. Figure 3.27(b) shows the training
and test error rates after adding 100 irrelevant attributes to the training set.
We can see that the test error rate remains close to 0.5 even after using 50
leaf nodes, while the training error rate keeps on declining and eventually
becomes 0.

3.5 Model Selection
There are many possible classification models with varying levels of model
complexity that can be used to capture patterns in the training data. Among
these possibilities, we want to select the model that shows lowest
generalization error rate. The process of selecting a model with the right level
of complexity, which is expected to generalize well over unseen test
instances, is known as model selection. As described in the previous
section, the training error rate cannot be reliably used as the sole criterion for
model selection. In the following, we present three generic approaches to
estimate the generalization performance of a model that can be used for
model selection. We conclude this section by presenting specific strategies for
using these approaches in the context of decision tree induction.

3.5.1 Using a Validation Set

Note that we can always estimate the generalization error rate of a model by
using “out-of-sample” estimates, i.e. by evaluating the model on a separate
validation set that is not used for training the model. The error rate on the
validation set, termed as the validation error rate, is a better indicator of
generalization performance than the training error rate, since the validation set
has not been used for training the model. The validation error rate can be
used for model selection as follows.

Given a training set D.train, we can partition D.train into two smaller subsets,
D.tr and D.val, such that D.tr is used for training while D.val is used as the
validation set. For example, two-thirds of D.train can be reserved as D.tr for

training, while the remaining one-third is used as D.val for computing
validation error rate. For any choice of classification model m that is trained on
D.tr, we can estimate its validation error rate on D.val, . The model
that shows the lowest value of can then be selected as the preferred
choice of model.

The use of validation set provides a generic approach for model selection.
However, one limitation of this approach is that it is sensitive to the sizes of
D.tr and D.val, obtained by partitioning D.train. If the size of D.tr is too small, it
may result in the learning of a poor classification model with sub-standard
performance, since a smaller training set will be less representative of the
overall data. On the other hand, if the size of D.val is too small, the validation
error rate might not be reliable for selecting models, as it would be computed
over a small number of instances.

Figure 3.29.


Class distribution of validation data for the two decision trees shown in Figure
3.30 .

3.8. Example Validation Error
In the following example, we illustrate one possible approach for using a
validation set in decision tree induction. Figure 3.29 shows the
predicted labels at the leaf nodes of the decision trees generated in Figure
3.30 . The counts given beneath the leaf nodes represent the proportion
of data instances in the validation set that reach each of the nodes. Based
on the predicted labels of the nodes, the validation error rate for the left
tree is , while the validation error rate for the right
tree is . Based on their validation error rates, the right
tree is preferred over the left one.

3.5.2 Incorporating Model Complexity

Since the chance for model overfitting increases as the model becomes more
complex, a model selection approach should not only consider the training
error rate but also the model complexity. This strategy is inspired by a well-
known principle known as Occam’s razor or the principle of parsimony,
which suggests that given two models with the same errors, the simpler model
is preferred over the more complex model. A generic approach to account for
model complexity while estimating generalization performance is formally
described as follows.

Given a training set D.train, let us consider learning a classification model m
that belongs to a certain class of models, . For example, if represents the
set of all possible decision trees, then m can correspond to a specific decision



tree learned from the training set. We are interested in estimating the
generalization error rate of m, gen.error(m). As discussed previously, the
training error rate of m, train.error(m, D.train), can under-estimate
gen.error(m) when the model complexity is high. Hence, we represent
gen.error(m) as a function of not just the training error rate but also the model
complexity of as follows:

where is a hyper-parameter that strikes a balance between minimizing
training error and reducing model complexity. A higher value of gives more
emphasis to the model complexity in the estimation of generalization
performance. To choose the right value of , we can make use of the
validation set in a similar way as described in 3.5.1 . For example, we can
iterate through a range of values of and for every possible value, we can
learn a model on a subset of the training set, D.tr, and compute its validation
error rate on a separate subset, D.val. We can then select the value of that
provides the lowest validation error rate.

Equation 3.11 provides one possible approach for incorporating model
complexity into the estimate of generalization performance. This approach is
at the heart of a number of techniques for estimating generalization
performance, such as the structural risk minimization principle, the Akaike’s
Information Criterion (AIC), and the Bayesian Information Criterion (BIC). The
structural risk minimization principle serves as the building block for learning
support vector machines, which will be discussed later in Chapter 4 . For
more details on AIC and BIC, see the Bibliographic Notes.

In the following, we present two different approaches for estimating the
complexity of a model, . While the former is specific to decision
trees, the latter is more generic and can be used with any class of models.

M, complexity(M),

gen.error(m)=train.error(m, D.train)+α×complexity(M), (3.11)






Estimating the Complexity of Decision Trees
In the context of decision trees, the complexity of a decision tree can be
estimated as the ratio of the number of leaf nodes to the number of training
instances. Let k be the number of leaf nodes and be the number of
training instances. The complexity of a decision tree can then be described as

. This reflects the intuition that for a larger training size, we can learn a
decision tree with larger number of leaf nodes without it becoming overly
complex. The generalization error rate of a decision tree T can then be
computed using Equation 3.11 as follows:

where err(T) is the training error of the decision tree and is a hyper-
parameter that makes a trade-off between reducing training error and
minimizing model complexity, similar to the use of in Equation 3.11 .
can be viewed as the relative cost of adding a leaf node relative to incurring a
training error. In the literature on decision tree induction, the above approach
for estimating generalization error rate is also termed as the pessimistic
error estimate. It is called pessimistic as it assumes the generalization error
rate to be worse than the training error rate (by adding a penalty term for
model complexity). On the other hand, simply using the training error rate as
an estimate of the generalization error rate is called the optimistic error
estimate or the resubstitution estimate.

3.9. Example Generalization Error Estimates
Consider the two binary decision trees, and , shown in Figure
3.30 . Both trees are generated from the same training data and is
generated by expanding three leaf nodes of . The counts shown in the
leaf nodes of the trees represent the class distribution of the training





α Ω



instances. If each leaf node is labeled according to the majority class of
training instances that reach the node, the training error rate for the left
tree will be , while the training error rate for the right
tree will be . Based on their training error rates alone,
would preferred over , even though is more complex (contains

larger number of leaf nodes) than .

Figure 3.30.
Example of two decision trees generated from the same training data.

Now, assume that the cost associated with each leaf node is . Then,
the generalization error estimate for will be

and the generalization error estimate for will be






errgen(TR) =624+0.5×424=824=0.3333.

Since has a lower generalization error rate, it will still be preferred over
. Note that implies that a node should always be expanded into

its two child nodes if it improves the prediction of at least one training
instance, since expanding a node is less costly than misclassifying a
training instance. On the other hand, if , then the generalization error
rate for is and for is

. In this case, will be preferred over
because it has a lower generalization error rate. This example illustrates
that different choices of can change our preference of decision trees
based on their generalization error estimates. However, for a given choice
of , the pessimistic error estimate provides an approach for modeling the
generalization performance on unseen test instances. The value of can
be selected with the help of a validation set.

Minimum Description Length Principle
Another way to incorporate model complexity is based on an information-
theoretic approach known as the minimum description length or MDL
principle. To illustrate this approach, consider the example shown in Figure
3.31 . In this example, both person and person are given a set of
instances with known attribute values . Assume person A knows the class
label y for every instance, while person has no such information. would
like to share the class information with by sending a message containing
the labels. The message would contain bits of information, where N is
the number of instances.

TR Ω=0.5

TL errgen(TL)=11/24=0.458 TR

errgen(TR)=10/24=0.417 TR TL




Figure 3.31.
An illustration of the minimum description length principle.

Alternatively, instead of sending the class labels explicitly, can build a
classification model from the instances and transmit it to . can then apply
the model to determine the class labels of the instances. If the model is 100%
accurate, then the cost for transmission is equal to the number of bits required
to encode the model. Otherwise, must also transmit information about
which instances are misclassified by the model so that can reproduce the
same class labels. Thus, the overall transmission cost, which is equal to the
total description length of the message, is

where the first term on the right-hand side is the number of bits needed to
encode the misclassified instances, while the second term is the number of
bits required to encode the model. There is also a hyper-parameter that
trades-off the relative costs of the misclassified instances and the model.

Cost(model, data)=Cost(data|model)+α×Cost(model), (3.12)


Notice the familiarity between this equation and the generic equation for
generalization error rate presented in Equation 3.11 . A good model must
have a total description length less than the number of bits required to encode
the entire sequence of class labels. Furthermore, given two competing
models, the model with lower total description length is preferred. An example
showing how to compute the total description length of a decision tree is given
in Exercise 10 on page 189.

3.5.3 Estimating Statistical Bounds

Instead of using Equation 3.11 to estimate the generalization error rate of a
model, an alternative way is to apply a statistical correction to the training
error rate of the model that is indicative of its model complexity. This can be
done if the probability distribution of training error is available or can be
assumed. For example, the number of errors committed by a leaf node in a
decision tree can be assumed to follow a binomial distribution. We can thus
compute an upper bound limit to the observed training error rate that can be
used for model selection, as illustrated in the following example.

3.10. Example Statistical Bounds on Training
Consider the left-most branch of the binary decision trees shown in Figure
3.30 . Observe that the left-most leaf node of has been expanded
into two child nodes in . Before splitting, the training error rate of the
node is . By approximating a binomial distribution with a normal
distribution, the following upper bound of the training error rate e can be



where is the confidence level, is the standardized value from a
standard normal distribution, and N is the total number of training
instances used to compute e. By replacing and , the
upper bound for the error rate is , which
corresponds to errors. If we expand the node into its child
nodes as shown in , the training error rates for the child nodes are

and , respectively. Using Equation (3.13) , the
upper bounds of these error rates are and

, respectively. The overall training error of the
child nodes is , which is larger than the estimated
error for the corresponding node in , suggesting that it should not be

3.5.4 Model Selection for Decision

Building on the generic approaches presented above, we present two
commonly used model selection strategies for decision tree induction.

Prepruning (Early Stopping Rule)

In this approach, the tree-growing algorithm is halted before generating a fully
grown tree that perfectly fits the entire training data. To do this, a more
restrictive stopping condition must be used; e.g., stop expanding a leaf node
when the observed gain in the generalization error estimate falls below a
certain threshold. This estimate of the generalization error rate can be

eupper(N, e, α)=e+zα/222N+zα/2e(1−e)N+zα/224N21+zα/22N, (3.13)

α zα/2

α=25%,N=7, e=2/7
eupper(7, 2/7, 0.25)=0.503


1/4=0.250 1/3=0.333
eupper(4, 1/4,0.25)=0.537

eupper(3, 1/3, 0.25)=0.650


computed using any of the approaches presented in the preceding three
subsections, e.g., by using pessimistic error estimates, by using validation
error estimates, or by using statistical bounds. The advantage of prepruning is
that it avoids the computations associated with generating overly complex
subtrees that overfit the training data. However, one major drawback of this
method is that, even if no significant gain is obtained using one of the existing
splitting criterion, subsequent splitting may result in better subtrees. Such
subtrees would not be reached if prepruning is used because of the greedy
nature of decision tree induction.


In this approach, the decision tree is initially grown to its maximum size. This
is followed by a tree-pruning step, which proceeds to trim the fully grown tree
in a bottom-up fashion. Trimming can be done by replacing a subtree with (1)
a new leaf node whose class label is determined from the majority class of
instances affiliated with the subtree (approach known as subtree
replacement), or (2) the most frequently used branch of the subtree
(approach known as subtree raising). The tree-pruning step terminates when
no further improvement in the generalization error estimate is observed
beyond a certain threshold. Again, the estimates of generalization error rate
can be computed using any of the approaches presented in the previous three
subsections. Post-pruning tends to give better results than prepruning
because it makes pruning decisions based on a fully grown tree, unlike
prepruning, which can suffer from premature termination of the tree-growing
process. However, for post-pruning, the additional computations needed to
grow the full tree may be wasted when the subtree is pruned.

Figure 3.32 illustrates the simplified decision tree model for the web robot
detection example given in Section 3.3.5 . Notice that the subtree rooted at

has been replaced by one of its branches corresponding todepth=1

, and , using subtree raising. On the other hand,
the subtree corresponding to and has been replaced by
a leaf node assigned to class 0, using subtree replacement. The subtree for

and remains intact.

Figure 3.32.
Post-pruning of the decision tree for web robot detection.

breadth<=7, width>3 MultiP=1
depth>1 MultiAgent=0

depth>1 MultiAgent=1

3.6 Model Evaluation
The previous section discussed several approaches for model selection that
can be used to learn a classification model from a training set D.train. Here we
discuss methods for estimating its generalization performance, i.e. its
performance on unseen instances outside of D.train. This process is known as
model evaluation.

Note that model selection approaches discussed in Section 3.5 also
compute an estimate of the generalization performance using the training set
D.train. However, these estimates are biased indicators of the performance on
unseen instances, since they were used to guide the selection of classification
model. For example, if we use the validation error rate for model selection (as
described in Section 3.5.1 ), the resulting model would be deliberately
chosen to minimize the errors on the validation set. The validation error rate
may thus under-estimate the true generalization error rate, and hence cannot
be reliably used for model evaluation.

A correct approach for model evaluation would be to assess the performance
of a learned model on a labeled test set has not been used at any stage of
model selection. This can be achieved by partitioning the entire set of labeled
instances D, into two disjoint subsets, D.train, which is used for model
selection and D.test, which is used for computing the test error rate, . In
the following, we present two different approaches for partitioning D into
D.train and D.test, and computing the test error rate, .

3.6.1 Holdout Method



The most basic technique for partitioning a labeled data set is the holdout
method, where the labeled set D is randomly partitioned into two disjoint sets,
called the training set D.train and the test set D.test. A classification model is
then induced from D.train using the model selection approaches presented in
Section 3.5 , and its error rate on D.test, , is used as an estimate of
the generalization error rate. The proportion of data reserved for training and
for testing is typically at the discretion of the analysts, e.g., two-thirds for
training and one-third for testing.

Similar to the trade-off faced while partitioning D.train into D.tr and D.val in
Section 3.5.1 , choosing the right fraction of labeled data to be used for
training and testing is non-trivial. If the size of D.train is small, the learned
classification model may be improperly learned using an insufficient number of
training instances, resulting in a biased estimation of generalization
performance. On the other hand, if the size of D.test is small, may be
less reliable as it would be computed over a small number of test instances.
Moreover, can have a high variance as we change the random
partitioning of D into D.train and D.test.

The holdout method can be repeated several times to obtain a distribution of
the test error rates, an approach known as random subsampling or repeated
holdout method. This method produces a distribution of the error rates that
can be used to understand the variance of .

3.6.2 Cross-Validation

Cross-validation is a widely-used model evaluation method that aims to make
effective use of all labeled instances in D for both training and testing. To
illustrate this method, suppose that we are given a labeled set that we have





randomly partitioned into three equal-sized subsets, , and , as
shown in Figure 3.33 . For the first run, we train a model using subsets
and S (shown as empty blocks) and test the model on subset . The test
error rate on , denoted as , is thus computed in the first run.
Similarly, for the second run, we use and as the training set and as
the test set, to compute the test error rate, , on . Finally, we use
and for training in the third run, while is used for testing, thus resulting
in the test error rate for . The overall test error rate is obtained by
summing up the number of errors committed in each test subset across all
runs and dividing it by the total number of instances. This approach is called
three-fold cross-validation.

Figure 3.33.
Example demonstrating the technique of 3-fold cross-validation.

The k-fold cross-validation method generalizes this approach by segmenting
the labeled data D (of size N) into k equal-sized partitions (or folds). During
the i run, one of the partitions of D is chosen as D.test(i) for testing, while the
rest of the partitions are used as D.train(i) for training. A model m(i) is learned
using D.train(i) and applied on D.test(i) to obtain the sum of test errors,

S1, S2 S3

3 S1
S1 err(S1)

S1 S3 S2
err(S2) S2 S1

S3 S3
err(S3) S3


. This procedure is repeated k times. The total test error rate, ,
is then computed as

Every instance in the data is thus used for testing exactly once and for training
exactly times. Also, every run uses fraction of the data for
training and 1/k fraction for testing.

The right choice of k in k-fold cross-validation depends on a number of
characteristics of the problem. A small value of k will result in a smaller
training set at every run, which will result in a larger estimate of generalization
error rate than what is expected of a model trained over the entire labeled set.
On the other hand, a high value of k results in a larger training set at every
run, which reduces the bias in the estimate of generalization error rate. In the
extreme case, when , every run uses exactly one data instance for testing
and the remainder of the data for testing. This special case of k-fold cross-
validation is called the leave-one-out approach. This approach has the
advantage of utilizing as much data as possible for training. However, leave-
one-out can produce quite misleading results in some special scenarios, as
illustrated in Exercise 11. Furthermore, leave-one-out can be computationally
expensive for large data sets as the cross-validation procedure needs to be
repeated N times. For most practical applications, the choice of k between 5
and 10 provides a reasonable approach for estimating the generalization error
rate, because each fold is able to make use of 80% to 90% of the labeled data
for training.

The k-fold cross-validation method, as described above, produces a single
estimate of the generalization error rate, without providing any information
about the variance of the estimate. To obtain this information, we can run k-
fold cross-validation for every possible partitioning of the data into k partitions,

errsum(i) errtest

errtest=∑i=1kerrsum(i)N. (3.14)

(k−1) (k−1)/k


and obtain a distribution of test error rates computed for every such
partitioning. The average test error rate across all possible partitionings
serves as a more robust estimate of generalization error rate. This approach
of estimating the generalization error rate and its variance is known as the
complete cross-validation approach. Even though such an estimate is quite
robust, it is usually too expensive to consider all possible partitionings of a
large data set into k partitions. A more practical solution is to repeat the cross-
validation approach multiple times, using a different random partitioning of the
data into k partitions at every time, and use the average test error rate as the
estimate of generalization error rate. Note that since there is only one possible
partitioning for the leave-one-out approach, it is not possible to estimate the
variance of generalization error rate, which is another limitation of this method.

The k-fold cross-validation does not guarantee that the fraction of positive and
negative instances in every partition of the data is equal to the fraction
observed in the overall data. A simple solution to this problem is to perform a
stratified sampling of the positive and negative instances into k partitions, an
approach called stratified cross-validation.

In k-fold cross-validation, a different model is learned at every run and the
performance of every one of the k models on their respective test folds is then
aggregated to compute the overall test error rate, . Hence, does
not reflect the generalization error rate of any of the k models. Instead, it
reflects the expected generalization error rate of the model selection
approach, when applied on a training set of the same size as one of the
training folds . This is different than the computed in the
holdout method, which exactly corresponds to the specific model learned over
D.train. Hence, although effectively utilizing every data instance in D for
training and testing, the computed in the cross-validation method does
not represent the performance of a single model learned over a specific

errtest errtest

(N(k−1)/k) errtest


Nonetheless, in practice, is typically used as an estimate of the
generalization error of a model built on D. One motivation for this is that when
the size of the training folds is closer to the size of the overall data (when k is
large), then resembles the expected performance of a model learned
over a data set of the same size as D. For example, when k is 10, every
training fold is 90% of the overall data. The then should approach the
expected performance of a model learned over 90% of the overall data, which
will be close to the expected performance of a model learned over D.




3.7 Presence of Hyper-parameters
Hyper-parameters are parameters of learning algorithms that need to be
determined before learning the classification model. For instance, consider the
hyper-parameter that appeared in Equation 3.11 , which is repeated here
for convenience. This equation was used for estimating the generalization
error for a model selection approach that used an explicit representations of
model complexity. (See Section 3.5.2 .)

For other examples of hyper-parameters, see Chapter 4 .

Unlike regular model parameters, such as the test conditions in the internal
nodes of a decision tree, hyper-parameters such as do not appear in the
final classification model that is used to classify unlabeled instances.
However, the values of hyper-parameters need to be determined during model
selection—a process known as hyper-parameter selection—and must be
taken into account during model evaluation. Fortunately, both tasks can be
effectively accomplished via slight modifications of the cross-validation
approach described in the previous section.

3.7.1 Hyper-parameter Selection

In Section 3.5.2 , a validation set was used to select and this approach is
generally applicable for hyper-parameter section. Let p be the hyper-
parameter that needs to be selected from a finite range of values,


gen.error(m)=train.error(m, D.train)+α×complexity(M)




. Partition D.train into D.tr and D.val. For every choice of hyper-
parameter value , we can learn a model on D.tr, and apply this model on
D.val to obtain the validation error rate . Let be the hyper-
parameter value that provides the lowest validation error rate. We can then
use the model corresponding to as the final choice of classification

The above approach, although useful, uses only a subset of the data, D.train,
for training and a subset, D.val, for validation. The framework of cross-
validation, presented in Section 3.6.2 , addresses both of those issues,
albeit in the context of model evaluation. Here we indicate how to use a cross-
validation approach for hyper-parameter selection. To illustrate this approach,
let us partition D.train into three folds as shown in Figure 3.34 . At every
run, one of the folds is used as D.val for validation, and the remaining two
folds are used as D.tr for learning a model, for every choice of hyper-
parameter value . The overall validation error rate corresponding to each
is computed by summing the errors across all the three folds. We then select
the hyper-parameter value that provides the lowest validation error rate,
and use it to learn a model on the entire training set D.train.

Figure 3.34.
Example demonstrating the 3-fold cross-validation framework for hyper-
parameter selection using D.train.

{p1, p2, … pn }
pi mi

errval(pi) p*

m* p*

pi pi


Algorithm 3.2 generalizes the above approach using a k-fold cross-
validation framework for hyper-parameter selection. At the i run of cross-
validation, the data in the i fold is used as D.val(i) for validation (Step 4),
while the remainder of the data in D.train is used as D.tr(i) for training (Step
5). Then for every choice of hyper-parameter value , a model is learned on
D.tr(i) (Step 7), which is applied on D.val(i) to compute its validation error
(Step 8). This is used to compute the validation error rate corresponding to
models learning using over all the folds (Step 11). The hyper-parameter
value that provides the lowest validation error rate (Step 12) is now used to
learn the final model on the entire training set D.train (Step 13). Hence, at
the end of this algorithm, we obtain the best choice of the hyper-parameter
value as well as the final classification model (Step 14), both of which are
obtained by making an effective use of every data instance in D.train.

Algorithm 3.2 Procedure model-select(k, ,







3.7.2 Nested Cross-Validation

The approach of the previous section provides a way to effectively use all the
instances in D.train to learn a classification model when hyper-parameter
selection is required. This approach can be applied over the entire data set D
to learn the final classification model. However, applying Algorithm 3.2 on
D would only return the final classification model but not an estimate of its
generalization performance, . Recall that the validation error rates used
in Algorithm 3.2 cannot be used as estimates of generalization
performance, since they are used to guide the selection of the final model .
However, to compute , we can again use a cross-validation framework
for evaluating the performance on the entire data set D, as described
originally in Section 3.6.2 . In this approach, D is partitioned into D.train (for
training) and D.test (for testing) at every run of cross-validation. When hyper-
parameters are involved, we can use Algorithm 3.2 to train a model using
D.train at every run, thus “internally” using cross-validation for model
selection. This approach is called nested cross-validation or double cross-
validation. Algorithm 3.3 describes the complete approach for estimating

using nested cross-validation in the presence of hyper-parameters.

As an illustration of this approach, see Figure 3.35 where the labeled set D
is partitioned into D.train and D.test, using a 3-fold cross-validation method.




Figure 3.35.
Example demonstrating 3-fold nested cross-validation for computing .

At the i run of this method, one of the folds is used as the test set, D.test(i),
while the remaining two folds are used as the training set, D.train(i). This is
represented in Figure 3.35 as the i “outer” run. In order to select a model
using D.train(i), we again use an “inner” 3-fold cross-validation framework that
partitions D.train(i) into D.tr and D.val at every one of the three inner runs
(iterations). As described in Section 3.7 , we can use the inner cross-
validation framework to select the best hyper-parameter value as well as
its resulting classification model learned over D.train(i). We can then
apply on D.test(i) to obtain the test error at the i outer run. By repeating
this process for every outer run, we can compute the average test error rate,

, over the entire labeled set D. Note that in the above approach, the
inner cross-validation framework is being used for model selection while the
outer cross-validation framework is being used for model evaluation.

Algorithm 3.3 The nested cross-validation
approach for computing .





m*(i) th



3.8 Pitfalls of Model Selection and
Model selection and evaluation, when used effectively, serve as excellent
tools for learning classification models and assessing their generalization
performance. However, when using them effectively in practical settings, there
are several pitfalls that can result in improper and often misleading
conclusions. Some of these pitfalls are simple to understand and easy to
avoid, while others are quite subtle in nature and difficult to catch. In the
following, we present two of these pitfalls and discuss best practices to avoid

3.8.1 Overlap between Training and
Test Sets

One of the basic requirements of a clean model selection and evaluation
setup is that the data used for model selection (D.train) must be kept separate
from the data used for model evaluation (D.test). If there is any overlap
between the two, the test error rate computed over D.test cannot be
considered representative of the performance on unseen instances.
Comparing the effectiveness of classification models using can then be
quite misleading, as an overly complex model can show an inaccurately low
value of due to model overfitting (see Exercise 12 at the end of this




To illustrate the importance of ensuring no overlap between D.train and D.test,
consider a labeled data set where all the attributes are irrelevant, i.e. they
have no relationship with the class labels. Using such attributes, we should
expect no classification model to perform better than random guessing.
However, if the test set involves even a small number of data instances that
were used for training, there is a possibility for an overly complex model to
show better performance than random, even though the attributes are
completely irrelevant. As we will see later in Chapter 10 , this scenario can
actually be used as a criterion to detect overfitting due to improper setup of
experiment. If a model shows better performance than a random classifier
even when the attributes are irrelevant, it is an indication of a potential
feedback between the training and test sets.

3.8.2 Use of Validation Error as
Generalization Error

The validation error rate serves an important role during model
selection, as it provides “out-of-sample” error estimates of models on D.val,
which is not used for training the models. Hence, serves as a better
metric than the training error rate for selecting models and hyper-parameter
values, as described in Sections 3.5.1 and 3.7 , respectively. However,
once the validation set has been used for selecting a classification model

no longer reflects the performance of on unseen instances.

To realize the pitfall in using validation error rate as an estimate of
generalization performance, consider the problem of selecting a hyper-
parameter value p from a range of values using a validation set D.val. If the
number of possible values in is quite large and the size of D.val is small, it is



m*, errval m*


possible to select a hyper-parameter value that shows favorable
performance on D.val just by random chance. Notice the similarity of this
problem with the multiple comparisons problem discussed in Section 3.4.1 .
Even though the classification model learned using would show a low
validation error rate, it would lack generalizability on unseen test instances.

The correct approach for estimating the generalization error rate of a model
is to use an independently chosen test set D.test that hasn’t been used in

any way to influence the selection of . As a rule of thumb, the test set
should never be examined during model selection, to ensure the absence of
any form of overfitting. If the insights gained from any portion of a labeled data
set help in improving the classification model even in an indirect way, then that
portion of data must be discarded during testing.


m* p*


3.9 Model Comparison
One difficulty when comparing the performance of different classification
models is whether the observed difference in their performance is statistically
significant. For example, consider a pair of classification models, and .
Suppose achieves 85% accuracy when evaluated on a test set containing
30 instances, while achieves 75% accuracy on a different test set
containing 5000 instances. Based on this information, is a better model
than ? This example raises two key questions regarding the statistical
significance of a performance metric:

1. Although has a higher accuracy than , it was tested on a smaller
test set. How much confidence do we have that the accuracy for is
actually 85%?

2. Is it possible to explain the difference in accuracies between and
as a result of variations in the composition of their test sets?

The first question relates to the issue of estimating the confidence interval of
model accuracy. The second question relates to the issue of testing the
statistical significance of the observed deviation. These issues are
investigated in the remainder of this section.

3.9.1 Estimating the Confidence
Interval for Accuracy







To determine its confidence interval, we need to establish the probability
distribution for sample accuracy. This section describes an approach for
deriving the confidence interval by modeling the classification task as a
binomial random experiment. The following describes characteristics of such
an experiment:

1. The random experiment consists of N independent trials, where each
trial has two possible outcomes: success or failure.

2. The probability of success, p, in each trial is constant.

An example of a binomial experiment is counting the number of heads that
turn up when a coin is flipped N times. If X is the number of successes
observed in N trials, then the probability that X takes a particular value is
given by a binomial distribution with mean and variance :

For example, if the coin is fair and is flipped fifty times, then the
probability that the head shows up 20 times is

If the experiment is repeated many times, then the average number of heads
expected to show up is while its variance is

The task of predicting the class labels of test instances can also be
considered as a binomial experiment. Given a test set that contains N
instances, let X be the number of instances correctly predicted by a model
and p be the true accuracy of the model. If the prediction task is modeled as a
binomial experiment, then X has a binomial distribution with mean and
variance It can be shown that the empirical accuracy, also

Np Np(1−p)




50×0.5=25, 50×0.5×0.5=12.5.

Np(1−p). acc=X/N,

has a binomial distribution with mean p and variance (see Exercise
14). The binomial distribution can be approximated by a normal distribution
when N is sufficiently large. Based on the normal distribution, the confidence
interval for acc can be derived as follows:

where and are the upper and lower bounds obtained from a
standard normal distribution at confidence level Since a standard
normal distribution is symmetric around it follows that
Rearranging this inequality leads to the following confidence interval for p:

The following table shows the values of at different confidence levels:

0.99 0.98 0.95 0.9 0.8 0.7 0.5

2.58 2.33 1.96 1.65 1.28 1.04 0.67

3.11. Example Confidence Interval for Accuracy
Consider a model that has an accuracy of 80% when evaluated on 100
test instances. What is the confidence interval for its true accuracy at a
95% confidence level? The confidence level of 95% corresponds to

according to the table given above. Inserting this term into
Equation 3.16 yields a confidence interval between 71.1% and 86.7%.
The following table shows the confidence interval when the number of
instances, N, increases:

N 20 50 100 500 1000 5000


P(−Zα/2≤acc−pp(1−p)/N≤Z1−α/2)=1−α, (3.15)

Zα/2 Z1−α/2

Z=0, Zα/2=Z1−α/2.

2×N×acc×Zα/22±Zα/2Zα/22+4Nacc−4Nacc22(N+Zα/22). (3.16)





Confidence 0.584 0.670 0.711 0.763 0.774 0.789


Note that the confidence interval becomes tighter when N increases.

3.9.2 Comparing the Performance of
Two Models

Consider a pair of models, and which are evaluated on two
independent test sets, and Let denote the number of instances in

and denote the number of instances in In addition, suppose the
error rate for on is and the error rate for on is Our goal is
to test whether the observed difference between and is statistically

Assuming that and are sufficiently large, the error rates and can
be approximated using normal distributions. If the observed difference in the
error rate is denoted as then d is also normally distributed with
mean , its true difference, and variance, The variance of d can be
computed as follows:

where and are the variances of the error rates.
Finally, at the confidence level, it can be shown that the confidence
interval for the true difference dt is given by the following equation:

−0.919 −0.888 −0.867 −0.833 −0.824 −0.811

M1 M2,
D1 D2. n1

D1 n2 D2.
M1 D1 e1 M2 D2 e2.

e1 e2

n1 n2 e1 e2

dt σd2.

σd2≃σ^d2=e1(1−e1)n1+e2(1−e2)n2, (3.17)

e1(1−e1)/n1 e2(1−e1)/n2

3.12. Example Significance Testing
Consider the problem described at the beginning of this section. Model
has an error rate of when applied to test instances, while
model has an error rate of when applied to test
instances. The observed difference in their error rates is

. In this example, we are performing a two-sided test to
check whether or . The estimated variance of the observed
difference in error rates can be computed as follows:

or . Inserting this value into Equation 3.18 , we obtain the
following confidence interval for at 95% confidence level:

As the interval spans the value zero, we can conclude that the observed
difference is not statistically significant at a 95% confidence level.

At what confidence level can we reject the hypothesis that ? To do this,
we need to determine the value of such that the confidence interval for
does not span the value zero. We can reverse the preceding computation and
look for the value such that . Replacing the values of d and

gives . This value first occurs when (for a two-
sided test). The result suggests that the null hypothesis can be rejected at
confidence level of 93.6% or lower.

dt=d±zα/2σ^d. (3.18)

e1=0.15 N1=30

MB e2=0.25 N2=5000

dt=0 dt≠0




Zα/2 dt

Zα/2 d>Zσ/2σ^d
σ^d Zσ/2<1.527 (1−α)<~0.936

3.10 Bibliographic Notes
Early classification systems were developed to organize various collections of
objects, from living organisms to inanimate ones. Examples abound, from
Aristotle’s cataloguing of species to the Dewey Decimal and Library of
Congress classification systems for books. Such a task typically requires
considerable human efforts, both to identify properties of the objects to be
classified and to organize them into well distinguished categories.

With the development of statistics and computing, automated classification
has been a subject of intensive research. The study of classification in
classical statistics is sometimes known as discriminant analysis, where the
objective is to predict the group membership of an object based on its
corresponding features. A well-known classical method is Fisher’s linear
discriminant analysis [142], which seeks to find a linear projection of the data
that produces the best separation between objects from different classes.

Many pattern recognition problems also require the discrimination of objects
from different classes. Examples include speech recognition, handwritten
character identification, and image classification. Readers who are interested
in the application of classification techniques for pattern recognition may refer
to the survey articles by Jain et al. [150] and Kulkarni et al. [157] or classic
pattern recognition books by Bishop [125], Duda et al. [137], and Fukunaga
[143]. The subject of classification is also a major research topic in neural
networks, statistical learning, and machine learning. An in-depth treatment on
the topic of classification from the statistical and machine learning
perspectives can be found in the books by Bishop [126], Cherkassky and
Mulier [132], Hastie et al. [148], Michie et al. [162], Murphy [167], and Mitchell
[165]. Recent years have also seen the release of many publicly available

software packages for classification, which can be embedded in programming
languages such as Java (Weka [147]) and Python (scikit-learn [174]).

An overview of decision tree induction algorithms can be found in the survey
articles by Buntine [129], Moret [166], Murthy [168], and Safavian et al. [179].
Examples of some well-known decision tree algorithms include CART [127],
ID3 [175], C4.5 [177], and CHAID [153]. Both ID3 and C4.5 employ the
entropy measure as their splitting function. An in-depth discussion of the C4.5
decision tree algorithm is given by Quinlan [177]. The CART algorithm was
developed by Breiman et al. [127] and uses the Gini index as its splitting
function. CHAID [153] uses the statistical test to determine the best split
during the tree-growing process.

The decision tree algorithm presented in this chapter assumes that the
splitting condition at each internal node contains only one attribute. An oblique
decision tree can use multiple attributes to form the attribute test condition in a
single node [149, 187]. Breiman et al. [127] provide an option for using linear
combinations of attributes in their CART implementation. Other approaches
for inducing oblique decision trees were proposed by Heath et al. [149],
Murthy et al. [169], Cantú-Paz and Kamath [130], and Utgoff and Brodley
[187]. Although an oblique decision tree helps to improve the expressiveness
of the model representation, the tree induction process becomes
computationally challenging. Another way to improve the expressiveness of a
decision tree without using oblique decision trees is to apply a method known
as constructive induction [161]. This method simplifies the task of learning
complex splitting functions by creating compound features from the original

Besides the top-down approach, other strategies for growing a decision tree
include the bottom-up approach by Landeweerd et al. [159] and Pattipati and
Alexandridis [173], as well as the bidirectional approach by Kim and


Landgrebe [154]. Schuermann and Doster [181] and Wang and Suen [193]
proposed using a soft splitting criterion to address the data fragmentation
problem. In this approach, each instance is assigned to different branches of
the decision tree with different probabilities.

Model overfitting is an important issue that must be addressed to ensure that
a decision tree classifier performs equally well on previously unlabeled data
instances. The model overfitting problem has been investigated by many
authors including Breiman et al. [127], Schaffer [180], Mingers [164], and
Jensen and Cohen [151]. While the presence of noise is often regarded as
one of the primary reasons for overfitting [164, 170], Jensen and Cohen [151]
viewed overfitting as an artifact of failure to compensate for the multiple
comparisons problem.

Bishop [126] and Hastie et al. [148] provide an excellent discussion of model
overfitting, relating it to a well-known framework of theoretical analysis, known
as bias-variance decomposition [146]. In this framework, the prediction of a
learning algorithm is considered to be a function of the training set, which
varies as the training set is changed. The generalization error of a model is
then described in terms of its bias (the error of the average prediction
obtained using different training sets), its variance (how different are the
predictions obtained using different training sets), and noise (the irreducible
error inherent to the problem). An underfit model is considered to have high
bias but low variance, while an overfit model is considered to have low bias
but high variance. Although the bias-variance decomposition was originally
proposed for regression problems (where the target attribute is a continuous
variable), a unified analysis that is applicable for classification has been
proposed by Domingos [136]. The bias variance decomposition will be
discussed in more detail while introducing ensemble learning methods in
Chapter 4 .

Various learning principles, such as the Probably Approximately Correct
(PAC) learning framework [188], have been developed to provide a theoretical
framework for explaining the generalization performance of learning
algorithms. In the field of statistics, a number of performance estimation
methods have been proposed that make a trade-off between the goodness of
fit of a model and the model complexity. Most noteworthy among them are the
Akaike’s Information Criterion [120] and the Bayesian Information Criterion
[182]. They both apply corrective terms to the training error rate of a model, so
as to penalize more complex models. Another widely-used approach for
measuring the complexity of any general model is the VapnikChervonenkis
(VC) Dimension [190]. The VC dimension of a class of functions C is defined
as the maximum number of points that can be shattered (every point can be
distinguished from the rest) by functions belonging to C, for any possible
configuration of points. The VC dimension lays the foundation of the structural
risk minimization principle [189], which is extensively used in many learning
algorithms, e.g., support vector machines, which will be discussed in detail in
Chapter 4 .

The Occam’s razor principle is often attributed to the philosopher William of
Occam. Domingos [135] cautioned against the pitfall of misinterpreting
Occam’s razor as comparing models with similar training errors, instead of
generalization errors. A survey on decision tree-pruning methods to avoid
overfitting is given by Breslow and Aha [128] and Esposito et al. [141]. Some
of the typical pruning methods include reduced error pruning [176], pessimistic
error pruning [176], minimum error pruning [171], critical value pruning [163],
cost-complexity pruning [127], and error-based pruning [177]. Quinlan and
Rivest proposed using the minimum description length principle for decision
tree pruning in [178].

The discussions in this chapter on the significance of cross-validation error
estimates is inspired from Chapter 7 in Hastie et al. [148]. It is also an

excellent resource for understanding “the right and wrong ways to do cross-
validation”, which is similar to the discussion on pitfalls in Section 3.8 of
this chapter. A comprehensive discussion of some of the common pitfalls in
using cross-validation for model selection and evaluation is provided in
Krstajic et al. [156].

The original cross-validation method was proposed independently by Allen
[121], Stone [184], and Geisser [145] for model assessment (evaluation).
Even though cross-validation can be used for model selection [194], its usage
for model selection is quite different than when it is used for model evaluation,
as emphasized by Stone [184]. Over the years, the distinction between the
two usages has often been ignored, resulting in incorrect findings. One of the
common mistakes while using cross-validation is to perform pre-processing
operations (e.g., hyper-parameter tuning or feature selection) using the entire
data set and not “within” the training fold of every cross-validation run.
Ambroise et al., using a number of gene expression studies as examples,
[124] provide an extensive discussion of the selection bias that arises when
feature selection is performed outside cross-validation. Useful guidelines for
evaluating models on microarray data have also been provided by Allison et
al. [122].

The use of the cross-validation protocol for hyper-parameter tuning has been
described in detail by Dudoit and van der Laan [138]. This approach has been
called “grid-search cross-validation.” The correct approach in using cross-
validation for both hyper-parameter selection and model evaluation, as
discussed in Section 3.7 of this chapter, is extensively covered by Varma
and Simon [191]. This combined approach has been referred to as “nested
cross-validation” or “double cross-validation” in the existing literature.
Recently, Tibshirani and Tibshirani [185] have proposed a new approach for
hyper-parameter selection and model evaluation. Tsamardinos et al. [186]
compared this approach to nested cross-validation. The experiments they

performed found that, on average, both approaches provide conservative
estimates of model performance with the Tibshirani and Tibshirani approach
being more computationally efficient.

Kohavi [155] has performed an extensive empirical study to compare the
performance metrics obtained using different estimation methods such as
random subsampling and k-fold cross-validation. Their results suggest that the
best estimation method is ten-fold, stratified cross-validation.

An alternative approach for model evaluation is the bootstrap method, which
was presented by Efron in 1979 [139]. In this method, training instances are
sampled with replacement from the labeled set, i.e., an instance previously
selected to be part of the training set is equally likely to be drawn again. If the
original data has N instances, it can be shown that, on average, a bootstrap
sample of size N contains about 63.2% of the instances in the original data.
Instances that are not included in the bootstrap sample become part of the
test set. The bootstrap procedure for obtaining training and test sets is
repeated b times, resulting in a different error rate on the test set, err(i), at the
i run. To obtain the overall error rate, , the .632 bootstrap approach
combines err(i) with the error rate obtained from a training set containing all
the labeled examples, , as follows:

Efron and Tibshirani [140] provided a theoretical and empirical comparison
between cross-validation and a bootstrap method known as the rule.

While the .632 bootstrap method presented above provides a robust estimate
of the generalization performance with low variance in its estimate, it may
produce misleading results for highly complex models in certain conditions, as
demonstrated by Kohavi [155]. This is because the overall error rate is not

th errboot


errboot=1b∑i=1b(0.632)×err(i)+0.386×errs). (3.19)


truly an out-of-sample error estimate as it depends on the training error rate,
, which can be quite small if there is overfitting.

Current techniques such as C4.5 require that the entire training data set fit
into main memory. There has been considerable effort to develop parallel and
scalable versions of decision tree induction algorithms. Some of the proposed
algorithms include SLIQ by Mehta et al. [160], SPRINT by Shafer et al. [183],
CMP by Wang and Zaniolo [192], CLOUDS by Alsabti et al. [123], RainForest
by Gehrke et al. [144], and ScalParC by Joshi et al. [152]. A survey of parallel
algorithms for classification and other data mining tasks is given in [158]. More
recently, there has been extensive research to implement large-scale
classifiers on the compute unified device architecture (CUDA) [131, 134] and
MapReduce [133, 172] platforms.


[120] H. Akaike. Information theory and an extension of the maximum

likelihood principle. In Selected Papers of Hirotugu Akaike, pages 199–
213. Springer, 1998.

[121] D. M. Allen. The relationship between variable selection and data
agumentation and a method for prediction. Technometrics, 16(1):125–127,

[122] D. B. Allison, X. Cui, G. P. Page, and M. Sabripour. Microarray data
analysis: from disarray to consolidation and consensus. Nature reviews
genetics, 7(1):55–65, 2006.

[123] K. Alsabti, S. Ranka, and V. Singh. CLOUDS: A Decision Tree Classifier
for Large Datasets. In Proc. of the 4th Intl. Conf. on Knowledge Discovery
and Data Mining, pages 2–8, New York, NY, August 1998.

[124] C. Ambroise and G. J. McLachlan. Selection bias in gene extraction on
the basis of microarray gene-expression data. Proceedings of the national
academy of sciences, 99 (10):6562–6566, 2002.

[125] C. M. Bishop. Neural Networks for Pattern Recognition. Oxford
University Press, Oxford, U.K., 1995.

[126] C. M. Bishop. Pattern Recognition and Machine Learning. Springer,

[127] L. Breiman, J. H. Friedman, R. Olshen, and C. J. Stone. Classification
and Regression Trees. Chapman & Hall, New York, 1984.

[128] L. A. Breslow and D. W. Aha. Simplifying Decision Trees: A Survey.
Knowledge Engineering Review, 12(1):1–40, 1997.

[129] W. Buntine. Learning classification trees. In Artificial Intelligence
Frontiers in Statistics, pages 182–201. Chapman & Hall, London, 1993.

[130] E. Cantú-Paz and C. Kamath. Using evolutionary algorithms to induce
oblique decision trees. In Proc. of the Genetic and Evolutionary
Computation Conf., pages 1053–1060, San Francisco, CA, 2000.

[131] B. Catanzaro, N. Sundaram, and K. Keutzer. Fast support vector
machine training and classification on graphics processors. In Proceedings
of the 25th International Conference on Machine Learning, pages 104–
111, 2008.

[132] V. Cherkassky and F. M. Mulier. Learning from Data: Concepts, Theory,
and Methods. Wiley, 2nd edition, 2007.

[133] C. Chu, S. K. Kim, Y.-A. Lin, Y. Yu, G. Bradski, A. Y. Ng, and K.
Olukotun. Map-reduce for machine learning on multicore. Advances in
neural information processing systems, 19:281, 2007.

[134] A. Cotter, N. Srebro, and J. Keshet. A GPU-tailored Approach for
Training Kernelized SVMs. In Proceedings of the 17th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pages
805–813, San Diego, California, USA, 2011.

[135] P. Domingos. The Role of Occam’s Razor in Knowledge Discovery. Data
Mining and Knowledge Discovery, 3(4):409–425, 1999.

[136] P. Domingos. A unified bias-variance decomposition. In Proceedings of
17th International Conference on Machine Learning, pages 231–238,

[137] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John
Wiley & Sons, Inc., New York, 2nd edition, 2001.

[138] S. Dudoit and M. J. van der Laan. Asymptotics of cross-validated risk
estimation in estimator selection and performance assessment. Statistical
Methodology, 2(2):131–154, 2005.

[139] B. Efron. Bootstrap methods: another look at the jackknife. In
Breakthroughs in Statistics, pages 569–593. Springer, 1992.

[140] B. Efron and R. Tibshirani. Cross-validation and the Bootstrap:
Estimating the Error Rate of a Prediction Rule. Technical report, Stanford
University, 1995.

[141] F. Esposito, D. Malerba, and G. Semeraro. A Comparative Analysis of
Methods for Pruning Decision Trees. IEEE Trans. Pattern Analysis and
Machine Intelligence, 19(5):476–491, May 1997.

[142] R. A. Fisher. The use of multiple measurements in taxonomic problems.
Annals of Eugenics, 7:179–188, 1936.

[143] K. Fukunaga. Introduction to Statistical Pattern Recognition. Academic
Press, New York, 1990.

[144] J. Gehrke, R. Ramakrishnan, and V. Ganti. RainForest—A Framework
for Fast Decision Tree Construction of Large Datasets. Data Mining and
Knowledge Discovery, 4(2/3):127–162, 2000.

[145] S. Geisser. The predictive sample reuse method with applications.
Journal of the American Statistical Association, 70(350):320–328, 1975.

[146] S. Geman, E. Bienenstock, and R. Doursat. Neural networks and the
bias/variance dilemma. Neural computation, 4(1):1–58, 1992.

[147] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H.
Witten. The WEKA Data Mining Software: An Update. SIGKDD
Explorations, 11(1), 2009.

[148] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical
Learning: Data Mining, Inference, and Prediction. Springer, 2nd edition,

[149] D. Heath, S. Kasif, and S. Salzberg. Induction of Oblique Decision
Trees. In Proc. of the 13th Intl. Joint Conf. on Artificial Intelligence, pages
1002–1007, Chambery, France, August 1993.

[150] A. K. Jain, R. P. W. Duin, and J. Mao. Statistical Pattern Recognition: A
Review. IEEE Tran. Patt. Anal. and Mach. Intellig., 22(1):4–37, 2000.

[151] D. Jensen and P. R. Cohen. Multiple Comparisons in Induction
Algorithms. Machine Learning, 38(3):309–338, March 2000.

[152] M. V. Joshi, G. Karypis, and V. Kumar. ScalParC: A New Scalable and
Efficient Parallel Classification Algorithm for Mining Large Datasets. In
Proc. of 12th Intl. Parallel Processing Symp. (IPPS/SPDP), pages 573–
579, Orlando, FL, April 1998.

[153] G. V. Kass. An Exploratory Technique for Investigating Large Quantities
of Categorical Data. Applied Statistics, 29:119–127, 1980.

[154] B. Kim and D. Landgrebe. Hierarchical decision classifiers in high-
dimensional and large class data. IEEE Trans. on Geoscience and Remote
Sensing, 29(4):518–528, 1991.

[155] R. Kohavi. A Study on Cross-Validation and Bootstrap for Accuracy
Estimation and Model Selection. In Proc. of the 15th Intl. Joint Conf. on
Artificial Intelligence, pages 1137–1145, Montreal, Canada, August 1995.

[156] D. Krstajic, L. J. Buturovic, D. E. Leahy, and S. Thomas. Cross-
validation pitfalls when selecting and assessing regression and
classification models. Journal of cheminformatics, 6(1):1, 2014.

[157] S. R. Kulkarni, G. Lugosi, and S. S. Venkatesh. Learning Pattern
Classification—A Survey. IEEE Tran. Inf. Theory, 44(6):2178–2206, 1998.

[158] V. Kumar, M. V. Joshi, E.-H. Han, P. N. Tan, and M. Steinbach. High
Performance Data Mining. In High Performance Computing for
Computational Science (VECPAR 2002), pages 111–125. Springer, 2002.

[159] G. Landeweerd, T. Timmers, E. Gersema, M. Bins, and M. Halic. Binary
tree versus single level tree classification of white blood cells. Pattern
Recognition, 16:571–577, 1983.

[160] M. Mehta, R. Agrawal, and J. Rissanen. SLIQ: A Fast Scalable Classifier
for Data Mining. In Proc. of the 5th Intl. Conf. on Extending Database
Technology, pages 18–32, Avignon, France, March 1996.

[161] R. S. Michalski. A theory and methodology of inductive learning. Artificial
Intelligence, 20:111–116, 1983.

[162] D. Michie, D. J. Spiegelhalter, and C. C. Taylor. Machine Learning,
Neural and Statistical Classification. Ellis Horwood, Upper Saddle River,
NJ, 1994.

[163] J. Mingers. Expert Systems—Rule Induction with Statistical Data. J
Operational Research Society, 38:39–47, 1987.

[164] J. Mingers. An empirical comparison of pruning methods for decision
tree induction. Machine Learning, 4:227–243, 1989.

[165] T. Mitchell. Machine Learning. McGraw-Hill, Boston, MA, 1997.

[166] B. M. E. Moret. Decision Trees and Diagrams. Computing Surveys,
14(4):593–623, 1982.

[167] K. P. Murphy. Machine Learning: A Probabilistic Perspective. MIT Press,

[168] S. K. Murthy. Automatic Construction of Decision Trees from Data: A
Multi-Disciplinary Survey. Data Mining and Knowledge Discovery,
2(4):345–389, 1998.

[169] S. K. Murthy, S. Kasif, and S. Salzberg. A system for induction of oblique
decision trees. J of Artificial Intelligence Research, 2:1–33, 1994.

[170] T. Niblett. Constructing decision trees in noisy domains. In Proc. of the
2nd European Working Session on Learning, pages 67–78, Bled,
Yugoslavia, May 1987.

[171] T. Niblett and I. Bratko. Learning Decision Rules in Noisy Domains. In
Research and Development in Expert Systems III, Cambridge, 1986.

Cambridge University Press.

[172] I. Palit and C. K. Reddy. Scalable and parallel boosting with mapreduce.
IEEE Transactions on Knowledge and Data Engineering, 24(10):1904–
1916, 2012.

[173] K. R. Pattipati and M. G. Alexandridis. Application of heuristic search
and information theory to sequential fault diagnosis. IEEE Trans. on
Systems, Man, and Cybernetics, 20(4):872–887, 1990.

[174] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O.
Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas,
A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay.
Scikit-learn: Machine Learning in Python. Journal of Machine Learning
Research, 12:2825–2830, 2011.

[175] J. R. Quinlan. Discovering rules by induction from large collection of
examples. In D. Michie, editor, Expert Systems in the Micro Electronic Age.
Edinburgh University Press, Edinburgh, UK, 1979.

[176] J. R. Quinlan. Simplifying Decision Trees. Intl. J. Man-Machine Studies,
27:221–234, 1987.

[177] J. R. Quinlan. C4.5: Programs for Machine Learning. Morgan-Kaufmann
Publishers, San Mateo, CA, 1993.

[178] J. R. Quinlan and R. L. Rivest. Inferring Decision Trees Using the
Minimum Description Length Principle. Information and Computation,
80(3):227–248, 1989.

[179] S. R. Safavian and D. Landgrebe. A Survey of Decision Tree Classifier
Methodology. IEEE Trans. Systems, Man and Cybernetics, 22:660–674,
May/June 1998.

[180] C. Schaffer. Overfitting avoidence as bias. Machine Learning, 10:153–
178, 1993.

[181] J. Schuermann and W. Doster. A decision-theoretic approach in
hierarchical classifier design. Pattern Recognition, 17:359–369, 1984.

[182] G. Schwarz et al. Estimating the dimension of a model. The annals of
statistics, 6(2): 461–464, 1978.

[183] J. C. Shafer, R. Agrawal, and M. Mehta. SPRINT: A Scalable Parallel
Classifier for Data Mining. In Proc. of the 22nd VLDB Conf., pages 544–
555, Bombay, India, September 1996.

[184] M. Stone. Cross-validatory choice and assessment of statistical
predictions. Journal of the Royal Statistical Society. Series B
(Methodological), pages 111–147, 1974.

[185] R. J. Tibshirani and R. Tibshirani. A bias correction for the minimum
error rate in cross-validation. The Annals of Applied Statistics, pages 822–

829, 2009.

[186] I. Tsamardinos, A. Rakhshani, and V. Lagani. Performance-estimation
properties of cross-validation-based protocols with simultaneous hyper-
parameter optimization. In Hellenic Conference on Artificial Intelligence,
pages 1–14. Springer, 2014.

[187] P. E. Utgoff and C. E. Brodley. An incremental method for finding
multivariate splits for decision trees. In Proc. of the 7th Intl. Conf. on
Machine Learning, pages 58–65, Austin, TX, June 1990.

[188] L. Valiant. A theory of the learnable. Communications of the ACM,
27(11):1134–1142, 1984.

[189] V. N. Vapnik. Statistical Learning Theory. Wiley-Interscience, 1998.

[190] V. N. Vapnik and A. Y. Chervonenkis. On the uniform convergence of
relative frequencies of events to their probabilities. In Measures of
Complexity, pages 11–30. Springer, 2015.

[191] S. Varma and R. Simon. Bias in error estimation when using cross-
validation for model selection. BMC bioinformatics, 7(1):1, 2006.

[192] H. Wang and C. Zaniolo. CMP: A Fast Decision Tree Classifier Using
Multivariate Predictions. In Proc. of the 16th Intl. Conf. on Data
Engineering, pages 449–460, San Diego, CA, March 2000.

[193] Q. R. Wang and C. Y. Suen. Large tree classifier with heuristic search
and global training. IEEE Trans. on Pattern Analysis and Machine
Intelligence, 9(1):91–102, 1987.

[194] Y. Zhang and Y. Yang. Cross-validation for selecting a model selection
procedure. Journal of Econometrics, 187(1):95–112, 2015.

3.11 Exercises
1. Draw the full decision tree for the parity function of four Boolean attributes,
A, B, C, and D. Is it possible to simplify the tree?

2. Consider the training examples shown in Table 3.5 for a binary
classification problem.

Table 3.5. Data set for Exercise 2.

Customer ID Gender Car Type Shirt Size Class

1 M Family Small C0

2 M Sports Medium C0

3 M Sports Medium C0

4 M Sports Large C0

5 M Sports Extra Large C0

6 M Sports Extra Large C0

7 F Sports Small C0

8 F Sports Small C0

9 F Sports Medium C0

10 F Luxury Large C0

11 M Family Large C1

12 M Family Extra Large C1

13 M Family Medium C1

14 M Luxury Extra Large C1

15 F Luxury Small C1

16 F Luxury Small C1

17 F Luxury Medium C1

18 F Luxury Medium C1

19 F Luxury Medium C1

20 F Luxury Large C1

a. Compute the Gini index for the overall collection of training examples.

b. Compute the Gini index for the attribute.

c. Compute the Gini index for the attribute.

d. Compute the Gini index for the attribute using multiway split.

e. Compute the Gini index for the attribute using multiway split.

f. Which attribute is better, , , or ?

g. Explain why should not be used as the attribute test
condition even though it has the lowest Gini.

3. Consider the training examples shown in Table 3.6 for a binary
classification problem.

Table 3.6. Data set for Exercise 3.

Instance Target Classa1 a2 a3

1 T T 1.0 +

2 T T 6.0

3 T F 5.0

4 F F 4.0

5 F T 7.0

6 F T 3.0

7 F F 8.0

8 T F 7.0

9 F T 5.0

a. What is the entropy of this collection of training examples with respect to
the class attribute?

b. What are the information gains of and relative to these training

c. For , which is a continuous attribute, compute the information gain for
every possible split.

d. What is the best split (among , and ) according to the information

e. What is the best split (between and ) according to the
misclassification error rate?

f. What is the best split (between and ) according to the Gini index?




a1 a2


a1, a2 a3

a1 a2

a1 a2

4. Show that the entropy of a node never increases after splitting it into
smaller successor nodes.

5. Consider the following data set for a binary class problem.

A B Class Label











a. Calculate the information gain when splitting on A and B. Which attribute
would the decision tree induction algorithm choose?

b. Calculate the gain in the Gini index when splitting on A and B. Which
attribute would the decision tree induction algorithm choose?

c. Figure 3.11 shows that entropy and the Gini index are both
monotonically increasing on the range [0, 0.5] and they are both
monotonically decreasing on the range [0.5, 1]. Is it possible that





information gain and the gain in the Gini index favor different attributes?

6. Consider splitting a parent node P into two child nodes, and , using
some attribute test condition. The composition of labeled training instances at
every node is summarized in the Table below.


Class 0 7 3 4

Class 1 3 0 3

a. Calculate the Gini index and misclassification error rate of the parent
node P .

b. Calculate the weighted Gini index of the child nodes. Would you consider
this attribute test condition if Gini is used as the impurity measure?

c. Calculate the weighted misclassification rate of the child nodes. Would
you consider this attribute test condition if misclassification rate is used as
the impurity measure?

7. Consider the following set of training examples.

X Y Z No. of Class C1 Examples No. of Class C2 Examples

0 0 0 5 40

0 0 1 0 15

0 1 0 10 5

0 1 1 45 0

C1 C2

C1 C2

1 0 0 10 5

1 0 1 25 0

1 1 0 5 20

1 1 1 0 15

a. Compute a two-level decision tree using the greedy approach described
in this chapter. Use the classification error rate as the criterion for
splitting. What is the overall error rate of the induced tree?

b. Repeat part (a) using X as the first splitting attribute and then choose the
best remaining attribute for splitting at each of the two successor nodes.
What is the error rate of the induced tree?

c. Compare the results of parts (a) and (b). Comment on the suitability of
the greedy heuristic used for splitting attribute selection.

8. The following table summarizes a data set with three attributes A, B, C and
two class labels . Build a two-level decision tree.

A B C Number of Instances


T T T 5 0

F T T 0 20

T F T 20 0

F F T 0 5

T T F 0 0

+, −

F T F 25 0

T F F 0 0

F F F 0 25

a. According to the classification error rate, which attribute would be chosen
as the first splitting attribute? For each attribute, show the contingency
table and the gains in classification error rate.

b. Repeat for the two children of the root node.

c. How many instances are misclassified by the resulting decision tree?

d. Repeat parts (a), (b), and (c) using C as the splitting attribute.

e. Use the results in parts (c) and (d) to conclude about the greedy nature of
the decision tree induction algorithm.

9. Consider the decision tree shown in Figure 3.36 .

Figure 3.36.
Decision tree and data sets for Exercise 9.

a. Compute the generalization error rate of the tree using the optimistic

b. Compute the generalization error rate of the tree using the pessimistic
approach. (For simplicity, use the strategy of adding a factor of 0.5 to
each leaf node.)

c. Compute the generalization error rate of the tree using the validation set
shown above. This approach is known as reduced error pruning.

10. Consider the decision trees shown in Figure 3.37 . Assume they are
generated from a data set that contains 16 binary attributes and 3 classes,

, and .C1, C2 C3

Compute the total description length of each decision tree according to the
following formulation of the minimum description length principle.

The total description length of a tree is given by

Each internal node of the tree is encoded by the ID of the splitting
attribute. If there are m attributes, the cost of encoding each attribute is


Figure 3.37.
Decision trees for Exercise 10.

Each leaf is encoded using the ID of the class it is associated with. If
there are k classes, the cost of encoding a class is bits.

Cost(tree) is the cost of encoding all the nodes in the tree. To simplify the
computation, you can assume that the total cost of the tree is obtained by
adding up the costs of encoding each internal node and each leaf node.



log2 k

is encoded using the classification errors the tree commits
on the training set. Each error is encoded by bits, where n is the
total number of training instances.

Which decision tree is better, according to the MDL principle?

11. This exercise, inspired by the discussions in [155], highlights one of the
known limitations of the leave-one-out model evaluation procedure. Let us
consider a data set containing 50 positive and 50 negative instances, where
the attributes are purely random and contain no information about the class
labels. Hence, the generalization error rate of any classification model learned
over this data is expected to be 0.5. Let us consider a classifier that assigns
the majority class label of training instances (ties resolved by using the
positive label as the default class) to any test instance, irrespective of its
attribute values. We can call this approach as the majority inducer classifier.
Determine the error rate of this classifier using the following methods.

a. Leave-one-out.

b. 2-fold stratified cross-validation, where the proportion of class labels at
every fold is kept same as that of the overall data.

c. From the results above, which method provides a more reliable
evaluation of the classifier’s generalization error rate?

12. Consider a labeled data set containing 100 data instances, which is
randomly partitioned into two sets A and B, each containing 50 instances. We
use A as the training set to learn two decision trees, with 10 leaf nodes
and with 100 leaf nodes. The accuracies of the two decision trees on
data sets A and B are shown in Table 3.7 .

Table 3.7. Comparing the test accuracy of decision trees and .


log2 n


T10 T100

Data Set

A 0.86 0.97

B 0.84 0.77

a. Based on the accuracies shown in Table 3.7 , which classification
model would you expect to have better performance on unseen

b. Now, you tested and on the entire data set and found
that the classification accuracy of on data set is 0.85, whereas
the classification accuracy of on the data set is 0.87. Based
on this new information and your observations from Table 3.7 , which
classification model would you finally choose for classification?

13. Consider the following approach for testing whether a classifier A beats
another classifier B. Let N be the size of a given dataset, be the accuracy
of classifier A, be the accuracy of classifier B, and be the
average accuracy for both classifiers. To test whether classifier A is
significantly better than B, the following Z-statistic is used:

Classifier A is assumed to be better than classifier B if .

Table 3.8 compares the accuracies of three different classifiers, decision
tree classifiers, naïve Bayes classifiers, and support vector machines, on
various data sets. (The latter two classifiers are described in Chapter 4 .)

Summarize the performance of the classifiers given in Table 3.8 using the
following table:

win-loss-draw Decision tree Naïve Bayes Support vector machine

T10 T100

T10 T100 (A+B)
T10 (A+B)

T100 (A+B)

pB p=(pA+pB)/2




Decision tree 0 – 0 – 23

Naïve Bayes 0 – 0 – 23

Support vector machine 0 – 0 – 23

Table 3.8. Comparing the accuracy of various classification methods.

Data Set Size(N) Decision Tree

naïve Bayes

Support vector machine

Anneal 898 92.09 79.62 87.19

Australia 690 85.51 76.81 84.78

Auto 205 81.95 58.05 70.73

Breast 699 95.14 95.99 96.42

Cleve 303 76.24 83.50 84.49

Credit 690 85.80 77.54 85.07

Diabetes 768 72.40 75.91 76.82

German 1000 70.90 74.70 74.40

Glass 214 67.29 48.59 59.81

Heart 270 80.00 84.07 83.70

Hepatitis 155 81.94 83.23 87.10

Horse 368 85.33 78.80 82.61

Ionosphere 351 89.17 82.34 88.89

Iris 150 94.67 95.33 96.00

Labor 57 78.95 94.74 92.98

Led7 3200 73.34 73.16 73.56

Lymphography 148 77.03 83.11 86.49

Pima 768 74.35 76.04 76.95

Sonar 208 78.85 69.71 76.92

Tic-tac-toe 958 83.72 70.04 98.33

Vehicle 846 71.04 45.04 74.94

Wine 178 94.38 96.63 98.88

Zoo 101 93.07 93.07 96.04

Each cell in the table contains the number of wins, losses, and draws when
comparing the classifier in a given row to the classifier in a given column.

14. Let X be a binomial random variable with mean and variance .
Show that the ratio X/N also has a binomial distribution with mean p and
variance .

Np Np(1−p)


4 Classification: Alternative

The previous chapter introduced the classification
problem and presented a technique known as the
decision tree classifier. Issues such as model overfitting
and model evaluation were also discussed. This
chapter presents alternative techniques for building
classification models—from simple techniques such as
rule-based and nearest neighbor classifiers to more
sophisticated techniques such as artificial neural
networks, deep learning, support vector machines, and
ensemble methods. Other practical issues such as the
class imbalance and multiclass problems are also
discussed at the end of the chapter.

4.1 Types of Classifiers
Before presenting specific techniques, we first categorize the different types of
classifiers available. One way to distinguish classifiers is by considering the
characteristics of their output.

Binary versus Multiclass

Binary classifiers assign each data instance to one of two possible labels,
typically denoted as and . The positive class usually refers to the
category we are more interested in predicting correctly compared to the
negative class (e.g., the category in email classification problems). If
there are more than two possible labels available, then the technique is known
as a multiclass classifier. As some classifiers were designed for binary classes
only, they must be adapted to deal with multiclass problems. Techniques for
transforming binary classifiers into multiclass classifiers are described in
Section 4.12 .

Deterministic versus Probabilistic

A deterministic classifier produces a discrete-valued label to each data
instance it classifies whereas a probabilistic classifier assigns a continuous
score between 0 and 1 to indicate how likely it is that an instance belongs to a
particular class, where the probability scores for all the classes sum up to 1.
Some examples of probabilistic classifiers include the naïve Bayes classifier,
Bayesian networks, and logistic regression. Probabilistic classifiers provide
additional information about the confidence in assigning an instance to a class
than deterministic classifiers. A data instance is typically assigned to the class

+1 −1

with the highest probability score, except when the cost of misclassifying the
class with lower probability is significantly higher. We will discuss the topic of
cost-sensitive classification with probabilistic outputs in Section 4.11.2 .

Another way to distinguish the different types of classifiers is based on their
technique for discriminating instances from different classes.

Linear versus Nonlinear

A linear classifier uses a linear separating hyperplane to discriminate
instances from different classes whereas a nonlinear classifier enables the
construction of more complex, nonlinear decision surfaces. We illustrate an
example of a linear classifier (perceptron) and its nonlinear counterpart (multi-
layer neural network) in Section 4.7 . Although the linearity assumption
makes the model less flexible in terms of fitting complex data, linear classifiers
are thus less susceptible to model overfitting compared to nonlinear
classifiers. Furthermore, one can transform the original set of attributes,

, into a more complex feature set, e.g.,
, before applying the linear classifier. Such feature

transformation allows the linear classifier to fit data sets with nonlinear
decision surfaces (see Section 4.9.4 ).

Global versus Local

A global classifier fits a single model to the entire data set. Unless the model
is highly nonlinear, this one-size-fits-all strategy may not be effective when the
relationship between the attributes and the class labels varies over the input
space. In contrast, a local classifier partitions the input space into smaller
regions and fits a distinct model to training instances in each region. The k-
nearest neighbor classifier (see Section 4.3 ) is a classic example of local
classifiers. While local classifiers are more flexible in terms of fitting complex

(x1, x2, ⋯ ,xd) Φ(x)=
(x1, x2, x1x2, x12, x22, ⋯)

decision boundaries, they are also more susceptible to the model overfitting
problem, especially when the local regions contain few training examples.

Generative versus Discriminative

Given a data instance , the primary objective of any classifier is to predict
the class label, y, of the data instance. However, apart from predicting the
class label, we may also be interested in describing the underlying
mechanism that generates the instances belonging to every class label. For
example, in the process of classifying spam email messages, it may be useful
to understand the typical characteristics of email messages that are labeled
as spam, e.g., specific usage of keywords in the subject or the body of the
email. Classifiers that learn a generative model of every class in the process
of predicting class labels are known as generative classifiers. Some examples
of generative classifiers include the naïve Bayes classifier and Bayesian
networks. In contrast, discriminative classifiers directly predict the class labels
without explicitly describing the distribution of every class label. They solve a
simpler problem than generative models since they do not have the onus of
deriving insights about the generative mechanism of data instances. They are
thus sometimes preferred over generative models, especially when it is not
crucial to obtain information about the properties of every class. Some
examples of discriminative classifiers include decision trees, rule-based
classifier, nearest neighbor classifier, artificial neural networks, and support
vector machines.

4.2 Rule-Based Classifier
A rule-based classifier uses a collection of “if …then…” rules (also known as a
rule set) to classify data instances. Table 4.1 shows an example of a rule
set generated for the vertebrate classification problem described in the
previous chapter. Each classification rule in the rule set can be expressed in
the following way:

The left-hand side of the rule is called the rule antecedent or precondition. It
contains a conjunction of attribute test conditions:

where is an attribute-value pair and op is a comparison operator
chosen from the set . Each attribute test is also
known as a conjunct. The right-hand side of the rule is called the rule
consequent, which contains the predicted class .

A rule r covers a data instance x if the precondition of r matches the attributes
of x. r is also said to be fired or triggered whenever it covers a given instance.
For an illustration, consider the rule given in Table 4.1 and the following
attributes for two vertebrates: hawk and grizzly bear.

Table 4.1. Example of a rule set for the vertebrate classification problem.

ri:(Conditioni)→yi. (4.1)

Conditioni=(A1 op v1)∧(A2 op v2)…(Ak op vk), (4.2)

(Aj, vj)
{=, ≠, <, >, ≤, ≥} (Aj op vj)



r1:(Gives Birth=no)∧(Aerial Creature=yes)→Birdsr2:(Gives Birth=no)∧(Aquatic Creature=yes)→Fishesr3:(Gives Birth=yes)∧(Body Temperature=warm-
blooded)→Mammalsr4:(Gives Birth=no)∧(Aerial Creature=no)→Reptilesr5:(Aquatic Creature=semi)→Amphibians

Name Body







hawk warm-

feather no no yes yes no



fur yes no no yes yes

covers the first vertebrate because its precondition is satisfied by the
hawk’s attributes. The rule does not cover the second vertebrate because
grizzly bears give birth to their young and cannot fly, thus violating the
precondition of .

The quality of a classification rule can be evaluated using measures such as
coverage and accuracy. Given a data set D and a classification rule r : ,
the coverage of the rule is the fraction of instances in D that trigger the rule r.
On the other hand, its accuracy or confidence factor is the fraction of
instances triggered by r whose class labels are equal to y. The formal
definitions of these measures are

where is the number of instances that satisfy the rule antecedent, is
the number of instances that satisfy both the antecedent and consequent, and

is the total number of instances.

Example 4.1.
Consider the data set shown in Table 4.2 . The rule




Coverage(r)=| A || D |Coverage(r)=|A∩y || A |, (4.3)

|A| |A∩y|


(Gives Birth=yes)∧(Body Temperature=warm-blooded)→Mammals

has a coverage of 33% since five of the fifteen instances support the rule
antecedent. The rule accuracy is 100% because all five vertebrates
covered by the rule are mammals.

Table 4.2. The vertebrate data set.
Name Body






Hibernates Class

human warm-

hair yes no no yes no Mammals

python cold-blooded scales no no no no yes Reptiles

salmon cold-blooded scales no yes no no no Fishes

whale warm-

hair yes yes no no no Mammals

frog cold-blooded none no semi no yes yes Amphibians


cold-blooded scales no no no yes no Reptiles

bat warm-

hair yes no yes yes yes Mammals

pigeon warm-

feathers no no yes yes no Birds

cat warm-

fur yes no no yes no Mammals

guppy cold-blooded scales yes yes no no no Fishes

alligator cold-blooded scales no semi no yes no Reptiles

penguin warm-


feathers no semi no yes no Birds

porcupine warm-

quills yes no no yes yes Mammals

eel cold-blooded scales no yes no no no Fishes

4.2.1 How a Rule-Based Classifier

A rule-based classifier classifies a test instance based on the rule triggered by
the instance. To illustrate how a rule-based classifier works, consider the rule
set shown in Table 4.1 and the following vertebrates:

Name Body







lemur warm-

fur yes no no yes yes

turtle cold-blooded scales no semi no yes no


cold-blooded scales yes yes no no no

The first vertebrate, which is a lemur, is warm-blooded and gives birth to its
young. It triggers the rule , and thus, is classified as a mammal.
The second vertebrate, which is a turtle, triggers the rules and . Since
the classes predicted by the rules are contradictory (reptiles versus
amphibians), their conflicting classes must be resolved.
None of the rules are applicable to a dogfish shark. In this case, we need
to determine what class to assign to such a test instance.

eel cold-blooded scales no yes no no no Fishes

salamander cold-blooded none no semi no yes yes Amphibians

r4 r5

4.2.2 Properties of a Rule Set

The rule set generated by a rule-based classifier can be characterized by the
following two properties.

Definition 4.1 (Mutually Exclusive Rule
The rules in a rule set R are mutually exclusive if no two rules in
R are triggered by the same instance. This property ensures that
every instance is covered by at most one rule in R.

Definition 4.2 (Exhaustive Rule Set).
A rule set R has exhaustive coverage if there is a rule for each
combination of attribute values. This property ensures that every
instance is covered by at least one rule in R.

Table 4.3. Example of a mutually exclusive and exhaustive rule set.

r1: (Body Temperature=cold-blooded)→Non-mammalsr2: (Body Temperature=warm-
blooded)∧(Gives Birth=yes)→Mammalsr3: (Body Temperature=warm-

Together, these two properties ensure that every instance is covered by
exactly one rule. An example of a mutually exclusive and exhaustive rule set
is shown in Table 4.3 . Unfortunately, many rule-based classifiers, including
the one shown in Table 4.1 , do not have such properties. If the rule set is
not exhaustive, then a default rule, , must be added to cover the
remaining cases. A default rule has an empty antecedent and is triggered
when all other rules have failed. is known as the default class and is
typically assigned to the majority class of training instances not covered by the
existing rules. If the rule set is not mutually exclusive, then an instance can be
covered by more than one rule, some of which may predict conflicting classes.

Definition 4.3 (Ordered Rule Set).
The rules in an ordered rule set R are ranked in decreasing
order of their priority. An ordered rule set is also known as a
decision list.

The rank of a rule can be defined in many ways, e.g., based on its accuracy or
total description length. When a test instance is presented, it will be classified
by the highest-ranked rule that covers the instance. This avoids the problem
of having conflicting classes predicted by multiple classification rules if the rule
set is not mutually exclusive.

blooded)∧(Gives Birth=no)→Non-mammals

rd: ()→yd


An alternative way to handle a non-mutually exclusive rule set without
ordering the rules is to consider the consequent of each rule triggered by a
test instance as a vote for a particular class. The votes are then tallied to
determine the class label of the test instance. The instance is usually
assigned to the class that receives the highest number of votes. The vote may
also be weighted by the rule’s accuracy. Using unordered rules to build a rule-
based classifier has both advantages and disadvantages. Unordered rules are
less susceptible to errors caused by the wrong rule being selected to classify
a test instance unlike classifiers based on ordered rules, which are sensitive
to the choice of rule-ordering criteria. Model building is also less expensive
because the rules do not need to be kept in sorted order. Nevertheless,
classifying a test instance can be quite expensive because the attributes of
the test instance must be compared against the precondition of every rule in
the rule set.

In the next two sections, we present techniques for extracting an ordered rule
set from data. A rule-based classifier can be constructed using (1) direct
methods, which extract classification rules directly from data, and (2) indirect
methods, which extract classification rules from more complex classification
models, such as decision trees and neural networks. Detailed discussions of
these methods are presented in Sections 4.2.3 and 4.2.4 , respectively.

4.2.3 Direct Methods for Rule

To illustrate the direct method, we consider a widely-used rule induction
algorithm called RIPPER. This algorithm scales almost linearly with the
number of training instances and is particularly suited for building models from

data sets with imbalanced class distributions. RIPPER also works well with
noisy data because it uses a validation set to prevent model overfitting.

RIPPER uses the sequential covering algorithm to extract rules directly from
data. Rules are grown in a greedy fashion one class at a time. For binary
class problems, RIPPER chooses the majority class as its default class and
learns the rules to detect instances from the minority class. For multiclass
problems, the classes are ordered according to their prevalence in the training
set. Let be the ordered list of classes, where is the least
prevalent class and is the most prevalent class. All training instances that
belong to are initially labeled as positive examples, while those that belong
to other classes are labeled as negative examples. The sequential covering
algorithm learns a set of rules to discriminate the positive from negative
examples. Next, all training instances from are labeled as positive, while
those from classes are labeled as negative. The sequential
covering algorithm would learn the next set of rules to distinguish from
other remaining classes. This process is repeated until we are left with only
one class, , which is designated as the default class.

Example 4.1. Sequential covering algorithm.

(y1, y2, … ,yc) y1


y3, y4, ⋯, yc



A summary of the sequential covering algorithm is shown in Algorithm 4.1 .
The algorithm starts with an empty decision list, R, and extracts rules for each
class based on the ordering specified by the class prevalence. It iteratively
extracts the rules for a given class y using the Learn-One-Rule function. Once
such a rule is found, all the training instances covered by the rule are
eliminated. The new rule is added to the bottom of the decision list R. This
procedure is repeated until the stopping criterion is met. The algorithm then
proceeds to generate rules for the next class.

Figure 4.1 demonstrates how the sequential covering algorithm works for a
data set that contains a collection of positive and negative examples. The rule
R1, whose coverage is shown in Figure 4.1(b) , is extracted first because it
covers the largest fraction of positive examples. All the training instances
covered by R1 are subsequently removed and the algorithm proceeds to look
for the next best rule, which is R2.

Learn-One-Rule Function
Finding an optimal rule is computationally expensive due to the exponential
search space to explore. The Learn-One-Rule function addresses this
problem by growing the rules in a greedy fashion. It generates an initial rule

, where the left-hand side is an empty set and the right-hand side
corresponds to the positive class. It then refines the rule until a certain
stopping criterion is met. The accuracy of the initial rule may be poor because
some of the training instances covered by the rule belong to the negative

r: {}→+

class. A new conjunct must be added to the rule antecedent to improve its

Figure 4.1.
An example of the sequential covering algorithm.

RIPPER uses the FOIL’s information gain measure to choose the best
conjunct to be added into the rule antecedent. The measure takes into
consideration both the gain in accuracy and support of a candidate rule,
where support is defined as the number of positive examples covered by the
rule. For example, suppose the rule initially covers positive
examples and negative examples. After adding a new conjunct B, the
extended rule covers positive examples and negative

r: A→+ p0
r′: A∧B→+ p1 n1

examples. The FOIL’s information gain of the extended rule is computed as

RIPPER chooses the conjunct with highest FOIL’s information gain to extend
the rule, as illustrated in the next example.

Example 4.2. [Foil’s Information Gain]
Consider the training set for the vertebrate classification problem shown in
Table 4.2 . Suppose the target class for the Learn-One-Rule function is
mammals. Initially, the antecedent of the rule covers 5
positive and 10 negative examples. Thus, the accuracy of the rule is only
0.333. Next, consider the following three candidate conjuncts to be added
to the left-hand side of the rule: ,
and . The number of positive and negative examples covered
by the rule after adding each conjunct along with their respective accuracy
and FOIL’s information gain are shown in the following table.

Candidate rule Accuracy Info Gain

3 0 1.000 4.755

5 1 0.714 5.498

2 4 0.200

Although has the highest accuracy among the three
candidates, the conjunct has the highest FOIL’s
information gain. Thus, it is chosen to extend the rule (see Figure 4.2 ).

FOIL’s information gain=p1×(log2p1p1+n1−log2p0p0+n0). (4.4)


Skin cover=hair, Body temperature=warm
Has legs=No

p1 n1

{Skin Cover=hair}→mammals

{Body temperature=wam}→mammals

{Has legs=No}→mammals −0.737

Skin cover=hair
Body temperature=warm

This process continues until adding new conjuncts no longer improves the
information gain measure.

Rule Pruning

The rules generated by the Learn-One-Rule function can be pruned to
improve their generalization errors. RIPPER prunes the rules based on their
performance on the validation set. The following metric is computed to
determine whether pruning is needed: , where p(n) is the number
of positive (negative) examples in the validation set covered by the rule. This
metric is monotonically related to the rule’s accuracy on the validation set. If
the metric improves after pruning, then the conjunct is removed. Pruning is
done starting from the last conjunct added to the rule. For example, given a
rule , RIPPER checks whether D should be pruned first, followed by
CD, BCD, etc. While the original rule covers only positive examples, the
pruned rule may cover some of the negative examples in the training set.

Building the Rule Set

After generating a rule, all the positive and negative examples covered by the
rule are eliminated. The rule is then added into the rule set as long as it does
not violate the stopping condition, which is based on the minimum description
length principle. If the new rule increases the total description length of the
rule set by at least d bits, then RIPPER stops adding rules into its rule set (by
default, d is chosen to be 64 bits). Another stopping condition used by
RIPPER is that the error rate of the rule on the validation set must not exceed



Figure 4.2.
General-to-specific and specific-to-general rule-growing strategies.

RIPPER also performs additional optimization steps to determine whether
some of the existing rules in the rule set can be replaced by better alternative
rules. Readers who are interested in the details of the optimization method
may refer to the reference cited at the end of this chapter.

Instance Elimination

After a rule is extracted, RIPPER eliminates the positive and negative
examples covered by the rule. The rationale for doing this is illustrated in the
next example.

Figure 4.3 shows three possible rules, R1, R2, and R3, extracted from a
training set that contains 29 positive examples and 21 negative examples.
The accuracies of R1, R2, and R3 are 12/15 (80%), 7/10 (70%), and 8/12
(66.7%), respectively. R1 is generated first because it has the highest
accuracy. After generating R1, the algorithm must remove the examples
covered by the rule so that the next rule generated by the algorithm is different
than R1. The question is, should it remove the positive examples only,
negative examples only, or both? To answer this, suppose the algorithm must
choose between generating R2 or R3 after R1. Even though R2 has a higher
accuracy than R3 (70% versus 66.7%), observe that the region covered by R2
is disjoint from R1, while the region covered by R3 overlaps with R1. As a
result, R1 and R3 together cover 18 positive and 5 negative examples
(resulting in an overall accuracy of 78.3%), whereas R1 and R2 together
cover 19 positive and 6 negative examples (resulting in a lower overall
accuracy of 76%). If the positive examples covered by R1 are not removed,
then we may overestimate the effective accuracy of R3. If the negative
examples covered by R1 are not removed, then we may underestimate the
accuracy of R3. In the latter case, we might end up preferring R2 over R3
even though half of the false positive errors committed by R3 have already
been accounted for by the preceding rule, R1. This example shows that the
effective accuracy after adding R2 or R3 to the rule set becomes evident only
when both positive and negative examples covered by R1 are removed.

Figure 4.3.
Elimination of training instances by the sequential covering algorithm. R1, R2,
and R3 represent regions covered by three different rules.

4.2.4 Indirect Methods for Rule

This section presents a method for generating a rule set from a decision tree.
In principle, every path from the root node to the leaf node of a decision tree
can be expressed as a classification rule. The test conditions encountered
along the path form the conjuncts of the rule antecedent, while the class label
at the leaf node is assigned to the rule consequent. Figure 4.4 shows an
example of a rule set generated from a decision tree. Notice that the rule set
is exhaustive and contains mutually exclusive rules. However, some of the
rules can be simplified as shown in the next example.

Figure 4.4.
Converting a decision tree into classification rules.

Example 4.3.
Consider the following three rules from Figure 4.4 :

Observe that the rule set always predicts a positive class when the value
of Q is Yes. Therefore, we may simplify the rules as follows:

is retained to cover the remaining instances of the positive class.
Although the rules obtained after simplification are no longer mutually
exclusive, they are less complex and are easier to interpret.

In the following, we describe an approach used by the C4.5rules algorithm to
generate a rule set from a decision tree. Figure 4.5 shows the decision tree




and resulting classification rules obtained for the data set given in Table
4.2 .

Rule Generation

Classification rules are extracted for every path from the root to one of the leaf
nodes in the decision tree. Given a classification rule , we consider a
simplified rule, where is obtained by removing one of the conjuncts
in A. The simplified rule with the lowest pessimistic error rate is retained
provided its error rate is less than that of the original rule. The rule-pruning
step is repeated until the pessimistic error of the rule cannot be improved
further. Because some of the rules may become identical after pruning, the
duplicate rules are discarded.

Figure 4.5.

r′:A′→y A′

Classification rules extracted from a decision tree for the vertebrate
classification problem.

Rule Ordering

After generating the rule set, C4.5rules uses the class-based ordering scheme
to order the extracted rules. Rules that predict the same class are grouped
together into the same subset. The total description length for each subset is
computed, and the classes are arranged in increasing order of their total
description length. The class that has the smallest description length is given
the highest priority because it is expected to contain the best set of rules. The
total description length for a class is given by , where

is the number of bits needed to encode the misclassified
examples, Lmodel is the number of bits needed to encode the model, and g is
a tuning parameter whose default value is 0.5. The tuning parameter depends
on the number of redundant attributes present in the model. The value of the
tuning parameter is small if the model contains many redundant attributes.

4.2.5 Characteristics of Rule-Based

1. Rule-based classifiers have very similar characteristics as decision
trees. The expressiveness of a rule set is almost equivalent to that of a
decision tree because a decision tree can be represented by a set of
mutually exclusive and exhaustive rules. Both rule-based and decision
tree classifiers create rectilinear partitions of the attribute space and
assign a class to each partition. However, a rule-based classifier can


allow multiple rules to be triggered for a given instance, thus enabling
the learning of more complex models than decision trees.

2. Like decision trees, rule-based classifiers can handle varying types of
categorical and continuous attributes and can easily work in multiclass
classification scenarios. Rule-based classifiers are generally used to
produce descriptive models that are easier to interpret but give
comparable performance to the decision tree classifier.

3. Rule-based classifiers can easily handle the presence of redundant
attributes that are highly correlated with one other. This is because
once an attribute has been used as a conjunct in a rule antecedent, the
remaining redundant attributes would show little to no FOIL’s
information gain and would thus be ignored.

4. Since irrelevant attributes show poor information gain, rule-based
classifiers can avoid selecting irrelevant attributes if there are other
relevant attributes that show better information gain. However, if the
problem is complex and there are interacting attributes that can
collectively distinguish between the classes but individually show poor
information gain, it is likely for an irrelevant attribute to be accidentally
favored over a relevant attribute just by random chance. Hence, rule-
based classifiers can show poor performance in the presence of
interacting attributes, when the number of irrelevant attributes is large.

5. The class-based ordering strategy adopted by RIPPER, which
emphasizes giving higher priority to rare classes, is well suited for
handling training data sets with imbalanced class distributions.

6. Rule-based classifiers are not well-suited for handling missing values in
the test set. This is because the position of rules in a rule set follows a
certain ordering strategy and even if a test instance is covered by
multiple rules, they can assign different class labels depending on their
position in the rule set. Hence, if a certain rule involves an attribute that
is missing in a test instance, it is difficult to ignore the rule and proceed

to the subsequent rules in the rule set, as such a strategy can result in
incorrect class assignments.

4.3 Nearest Neighbor Classifiers
The classification framework shown in Figure 3.3 involves a two-step

(1) an inductive step for constructing a classification model from data, and

(2) a deductive step for applying the model to test examples. Decision tree
and rule-based classifiers are examples of eager learners because they are
designed to learn a model that maps the input attributes to the class label as
soon as the training data becomes available. An opposite strategy would be to
delay the process of modeling the training data until it is needed to classify the
test instances. Techniques that employ this strategy are known as lazy
learners. An example of a lazy learner is the Rote classifier, which
memorizes the entire training data and performs classification only if the
attributes of a test instance match one of the training examples exactly. An
obvious drawback of this approach is that some test instances may not be
classified because they do not match any training example.

One way to make this approach more flexible is to find all the training
examples that are relatively similar to the attributes of the test instances.
These examples, which are known as nearest neighbors, can be used to
determine the class label of the test instance. The justification for using
nearest neighbors is best exemplified by the following saying: “If it walks like a
duck, quacks like a duck, and looks like a duck, then it’s probably a duck.” A
nearest neighbor classifier represents each example as a data point in a d-
dimensional space, where d is the number of attributes. Given a test instance,
we compute its proximity to the training instances according to one of the
proximity measures described in Section 2.4 on page 71. The k-nearest

neighbors of a given test instance z refer to the k training examples that are
closest to z.

Figure 4.6 illustrates the 1-, 2-, and 3-nearest neighbors of a test instance
located at the center of each circle. The instance is classified based on the
class labels of its neighbors. In the case where the neighbors have more than
one label, the test instance is assigned to the majority class of its nearest
neighbors. In Figure 4.6(a) , the 1-nearest neighbor of the instance is a
negative example. Therefore the instance is assigned to the negative class. If
the number of nearest neighbors is three, as shown in Figure 4.6(c) , then
the neighborhood contains two positive examples and one negative example.
Using the majority voting scheme, the instance is assigned to the positive
class. In the case where there is a tie between the classes (see Figure
4.6(b) ), we may randomly choose one of them to classify the data point.

Figure 4.6.
The 1-, 2-, and 3-nearest neighbors of an instance.

The preceding discussion underscores the importance of choosing the right
value for k. If k is too small, then the nearest neighbor classifier may be
susceptible to overfitting due to noise, i.e., mislabeled examples in the training

data. On the other hand, if k is too large, the nearest neighbor classifier may
misclassify the test instance because its list of nearest neighbors includes
training examples that are located far away from its neighborhood (see Figure
4.7 ).

Figure 4.7.
k-nearest neighbor classification with large k.

4.3.1 Algorithm

A high-level summary of the nearest neighbor classification method is given in
Algorithm 4.2 . The algorithm computes the distance (or similarity) between
each test instance and all the training examples to
determine its nearest neighbor list, . Such computation can be costly if the
number of training examples is large. However, efficient indexing techniques
are available to reduce the computation needed to find the nearest neighbors
of a test instance.

z=(x′, y′) (x, y)∈D

Algorithm 4.2 The k-nearest neighbor classifier.

′ ′

′ ∑ ∈

Once the nearest neighbor list is obtained, the test instance is classified
based on the majority class of its nearest neighbors:

where v is a class label, is the class label for one of the nearest neighbors,
and is an indicator function that returns the value 1 if its argument is true
and 0 otherwise.

In the majority voting approach, every neighbor has the same impact on the
classification. This makes the algorithm sensitive to the choice of k, as shown
in Figure 4.6 . One way to reduce the impact of k is to weight the influence
of each nearest neighbor according to its distance: . As a
result, training examples that are located far away from z have a weaker
impact on the classification compared to those that are located close to z.
Using the distance-weighted voting scheme, the class label can be
determined as follows:

Majority Voting: y′=argmaxv∑(xi, yi)∈DzI(v=yi), (4.5)


xi wi=1/d(x′, xi)2

Distance-Weighted Voting: y′=argmaxv∑(xi, yi)∈Dzwi×I(v=yi). (4.6)

4.3.2 Characteristics of Nearest
Neighbor Classifiers

1. Nearest neighbor classification is part of a more general technique
known as instance-based learning, which does not build a global
model, but rather uses the training examples to make predictions for a
test instance. (Thus, such classifiers are often said to be “model free.”)
Such algorithms require a proximity measure to determine the similarity
or distance between instances and a classification function that returns
the predicted class of a test instance based on its proximity to other

2. Although lazy learners, such as nearest neighbor classifiers, do not
require model building, classifying a test instance can be quite
expensive because we need to compute the proximity values
individually between the test and training examples. In contrast, eager
learners often spend the bulk of their computing resources for model
building. Once a model has been built, classifying a test instance is
extremely fast.

3. Nearest neighbor classifiers make their predictions based on local
information. (This is equivalent to building a local model for each test
instance.) By contrast, decision tree and rule-based classifiers attempt
to find a global model that fits the entire input space. Because the
classification decisions are made locally, nearest neighbor classifiers
(with small values of k) are quite susceptible to noise.

4. Nearest neighbor classifiers can produce decision boundaries of
arbitrary shape. Such boundaries provide a more flexible model
representation compared to decision tree and rule-based classifiers
that are often constrained to rectilinear decision boundaries. The
decision boundaries of nearest neighbor classifiers also have high

variability because they depend on the composition of training
examples in the local neighborhood. Increasing the number of nearest
neighbors may reduce such variability.

5. Nearest neighbor classifiers have difficulty handling missing values in
both the training and test sets since proximity computations normally
require the presence of all attributes. Although, the subset of attributes
present in two instances can be used to compute a proximity, such an
approach may not produce good results since the proximity measures
may be different for each pair of instances and thus hard to compare.

6. Nearest neighbor classifiers can handle the presence of interacting
attributes, i.e., attributes that have more predictive power taken in
combination then by themselves, by using appropriate proximity
measures that can incorporate the effects of multiple attributes

7. The presence of irrelevant attributes can distort commonly used
proximity measures, especially when the number of irrelevant attributes
is large. Furthermore, if there are a large number of redundant
attributes that are highly correlated with each other, then the proximity
measure can be overly biased toward such attributes, resulting in
improper estimates of distance. Hence, the presence of irrelevant and
redundant attributes can adversely affect the performance of nearest
neighbor classifiers.

8. Nearest neighbor classifiers can produce wrong predictions unless the
appropriate proximity measure and data preprocessing steps are taken.
For example, suppose we want to classify a group of people based on
attributes such as height (measured in meters) and weight (measured
in pounds). The height attribute has a low variability, ranging from 1.5
m to 1.85 m, whereas the weight attribute may vary from 90 lb. to 250
lb. If the scale of the attributes are not taken into consideration, the
proximity measure may be dominated by differences in the weights of a

4.4 Naïve Bayes Classifier
Many classification problems involve uncertainty. First, the observed attributes
and class labels may be unreliable due to imperfections in the measurement
process, e.g., due to the limited preciseness of sensor devices. Second, the
set of attributes chosen for classification may not be fully representative of the
target class, resulting in uncertain predictions. To illustrate this, consider the
problem of predicting a person’s risk for heart disease based on a model that
uses their diet and workout frequency as attributes. Although most people
who eat healthily and exercise regularly have less chance of developing heart
disease, they may still be at risk due to other latent factors, such as heredity,
excessive smoking, and alcohol abuse, that are not captured in the model.
Third, a classification model learned over a finite training set may not be able
to fully capture the true relationships in the overall data, as discussed in the
context of model overfitting in the previous chapter. Finally, uncertainty in
predictions may arise due to the inherent random nature of real-world
systems, such as those encountered in weather forecasting problems.

In the presence of uncertainty, there is a need to not only make predictions of
class labels but also provide a measure of confidence associated with every
prediction. Probability theory offers a systematic way for quantifying and
manipulating uncertainty in data, and thus, is an appealing framework for
assessing the confidence of predictions. Classification models that make use
of probability theory to represent the relationship between attributes and class
labels are known as probabilistic classification models. In this section, we
present the naïve Bayes classifier, which is one of the simplest and most
widely-used probabilistic classification models.

4.4.1 Basics of Probability Theory

Before we discuss how the naïve Bayes classifier works, we first introduce
some basics of probability theory that will be useful in understanding the
probabilistic classification models presented in this chapter. This involves
defining the notion of probability and introducing some common approaches
for manipulating probability values.

Consider a variable X, which can take any discrete value from the set
. When we have multiple observations of that variable, such as in a

data set where the variable describes some characteristic of data objects,
then we can compute the relative frequency with which each value occurs.
Specifically, suppose that X has the value for data objects. The relative
frequency with which we observe the event is then , where N
denotes the total number of occurrences ( ). These relative
frequencies characterize the uncertainty that we have with respect to what
value X may take for an unseen observation and motivates the notion of

More formally, the probability of an event e, e.g., , measures how
likely it is for the event e to occur. The most traditional view of probability is
based on relative frequency of events (frequentist), while the Bayesian
viewpoint (described later) takes a more flexible view of probabilities. In either
case, a probability is always a number between 0 and 1. Further, the sum of
probability values of all possible events, e.g., outcomes of a variable X is
equal to 1. Variables that have probabilities associated with each possible
outcome (values) are known as random variables.

Now, let us consider two random variables, X and Y , that can each take k
discrete values. Let be the number of times we observe and , out

{x1, …, xk}

xi ni
X=xi ni/N



nij X=xi Y=yj

of a total number of N occurrences. The joint probability of observing
and together can be estimated as

(This is an estimate since we typically have only a finite subset of all possible
observations.) Joint probabilities can be used to answer questions such as
“what is the probability that there will be a surprise quiz today I will be late
for the class.” Joint probabilities are symmetric, i.e.,

. For joint probabilities, it is to useful to consider
their sum with respect to one of the random variables, as described in the
following equation:

where is the total number of times we observe irrespective of the value
of Y. Notice that is essentially the probability of observing . Hence,
by summing out the joint probabilities with respect to a random variable Y , we
obtain the probability of observing the remaining variable X. This operation is
called marginalization and the probability value obtained by
marginalizing out Y is sometimes called the marginal probability of X. As we
will see later, joint probability and marginal probability form the basic building
blocks of a number of probabilistic classification models discussed in this

Notice that in the previous discussions, we used to denote the
probability of a particular outcome of a random variable X. This notation can
easily become cumbersome when a number of random variables are involved.
Hence, in the remainder of this section, we will use P(X) to denote the
probability of any generic outcome of the random variable X, while will be
used to represent the probability of the specific outcome .


P(X=xi, Y=yi)=nijN. (4.7)

P(X=x, Y=y)=P(Y=y, X=x)

∑j=1kP(X=xi,Y=yj)=∑j=1knijN=niN=P(X=xi), (4.8)

ni X=xi
ni/N X=xi




Bayes Theorem
Suppose you have invited two of your friends Alex and Martha to a dinner party. You know that Alex

attends 40% of the parties he is invited to. Further, if Alex is going to a party, there is an 80% chance

of Martha coming along. On the other hand, if Alex is not going to the party, the chance of Martha

coming to the party is reduced to 30%. If Martha has responded that she will be coming to your party,

what is the probability that Alex will also be coming?

Bayes theorem presents the statistical principle for answering questions like
the previous one, where evidence from multiple sources has to be combined
with prior beliefs to arrive at predictions. Bayes theorem can be briefly
described as follows.

Let denotethe conditional probability of observing the random
variable Y whenever the random variable X takes a particular value. is
often read as the probability of observing Y conditioned on the outcome of X.
Conditional probabilities can be used for answering questions such as “given
that it is going to rain today, what will be the probability that I will go to the
class.” Conditional probabilities of X and Y are related to their joint

probability in the following way:

Rearranging the last two expressions in Equation 4.10 leads to Equation
4.11 , which is known as Bayes theorem:


P(Y|X)=P(X, Y)P(X), which implies (4.9)

P(X, Y)=P(Y|X)×P(X)=P(X|Y)×P(Y). (4.10)

P(Y|X)=P(X|Y)P(Y)P(X). (4.11)

Bayes theorem provides a relationship between the conditional probabilities
and . Note that the denominator in Equation 4.11 involves the

marginal probability of X, which can also be represented as

Using the previous expression for P(X), we can obtain the following equation
for solely in terms of and P(Y):

Example 4.4. [Bayes Theorem]
Bayes theorem can be used to solve a number of inferential questions
about random variables. For example, consider the problem stated at the
beginning on inferring whether Alex will come to the party. Let
denote the probability of Alex going to a party, while denotes the
probability of him not going to a party. We know that

Further, let denote the conditional probability of Martha going to
a party conditioned on whether Alex is going to the party. takes
the following values:

We can use the above values of and P(A) to compute the
probability of Alex going to the party given Martha is going to the party,

, as follows:

P(Y|X) P(X|Y)

P(X)=∑i=1kP(X, yi)=∑i=1kP(X|yi)×P(yi).

P(Y|X) P(X|Y)

P(Y|X)=P(X|Y)P(Y)∑i−1kP(X|yi)P(yi). (4.12)







Notice that even though the prior probability P(A) of Alex going to the party
is low, the observation that Martha is going, , affects the conditional
probability . This shows the value of Bayes theorem in
combining prior assumptions with observed outcomes to make predictions.
Since , it is more likely for Alex to join if Martha is going to
the party.

Using Bayes Theorem for Classification
For the purpose of classification, we are interested in computing the
probability of observing a class label y for a data instance given its set of
attribute values . This can be represented as , which is known as the
posterior probability of the target class. Using the Bayes Theorem, we can
represent the posterior probability as

Note that the numerator of the previous equation involves two terms,
and P(y), both of which contribute to the posterior probability . We
describe both of these terms in the following.

The first term is known as the class-conditional probability of the
attributes given the class label. measures the likelihood of observing
from the distribution of instances belonging to y. If indeed belongs to class
y, then we should expect to be high. From this point of view, the use of
class-conditional probabilities attempts to capture the process from which the
data instances were generated. Because of this interpretation, probabilistic
classification models that involve computing class-conditional probabilities are





P(y|x)=P(x|y)P(y)P(x) (4.14)




known as generative classification models. Apart from their use in
computing posterior probabilities and making predictions, class-conditional
probabilities also provide insights about the underlying mechanism behind the
generation of attribute values.

The second term in the numerator of Equation 4.14 is the prior probability
P(y). The prior probability captures our prior beliefs about the distribution of
class labels, independent of the observed attribute values. (This is the
Bayesian viewpoint.) For example, we may have a prior belief that the
likelihood of any person to suffer from a heart disease is , irrespective of their
diagnosis reports. The prior probability can either be obtained using expert
knowledge, or inferred from historical distribution of class labels.

The denominator in Equation 4.14 involves the probability of evidence, P
( ). Note that this term does not depend on the class label and thus can be
treated as a normalization constant in the computation of posterior
probabilities. Further, the value of P( ) can be calculated as


Bayes theorem provides a convenient way to combine our prior beliefs with
the likelihood of obtaining the observed attribute values. During the training
phase, we are required to learn the parameters for P(y) and . The prior
probability P(y) can be easily estimated from the training set by computing the
fraction of training instances that belong to each class. To compute the class-
conditional probabilities, one approach is to consider the fraction of training
instances of a given class for every possible combination of attribute values.
For example, suppose that there are two attributes and that can each
take a discrete value from to . Let denote the number of training
instances belonging to class 0, out of which number of training instances
have and . The class-conditional probability can then be given as




X1 X2
c1 ck n0

X1=ci X2=cj

This approach can easily become computationally prohibitive as the number
of attributes increase, due to the exponential growth in the number of attribute
value combinations. For example, if every attribute can take k discrete values,
then the number of attribute value combinations is equal to , where d is the
number of attributes. The large number of attribute value combinations can
also result in poor estimates of class-conditional probabilities, since every
combination will have fewer training instances when the size of training set is

In the following, we present the naïve Bayes classifier, which makes a
simplifying assumption about the class-conditional probabilities, known as the
naïve Bayes assumption. The use of this assumption significantly helps in
obtaining reliable estimates of class-conditional probabilities, even when the
number of attributes are large.

4.4.2 Naïve Bayes Assumption

The naïve Bayes classifier assumes that the class-conditional probability of all
attributes can be factored as a product of class-conditional probabilities of
every attribute , as described in the following equation:

where every data instance consists of d attributes, . The
basic assumption behind the previous equation is that the attribute values
are conditionally independent of each other, given the class label y. This
means that the attributes are influenced only by the target class and if we

P(X1=ci, X2=cj|Y=0)=nij0n0.



P(x|y)=∏i=1dP(xi|y), (4.15)

{x1, x2, …, xd}

know the class label, then we can consider the attributes to be independent of
each other. The concept of conditional independence can be formally stated
as follows.

Conditional Independence
Let , and Y denote three sets of random variables. The variables in
are said to be conditionally independent of , given Y, if the following
condition holds:

This means that conditioned on Y, the distribution of is not influenced by
the outcomes of , and hence is conditionally independent of . To illustrate
the notion of conditional independence, consider the relationship between a
person’s arm length and his or her reading skills . One might observe
that people with longer arms tend to have higher levels of reading skills, and
thus consider and to be related to each other. However, this
relationship can be explained by another factor, which is the age of the person
(Y). A young child tends to have short arms and lacks the reading skills of an
adult. If the age of a person is fixed, then the observed relationship between
arm length and reading skills disappears. Thus, we can conclude that arm
length and reading skills are not directly related to each other and are
conditionally independent when the age variable is fixed.

Another way of describing conditional independence is to consider the joint
conditional probability, , as follows:

X1, X2, X1

P(X1|X2, Y)=P(X1|Y). (4.16)

X2 X2

(X1) (X2)

X1 X2

P(X1, X2|Y)

P(X1, X2|Y)=P(X1, X2, Y)P(Y)=P(X1, X2, Y)P(X2, Y)×P(X2, Y)P(Y)=P(X1|X2, Y(4.17)

where Equation 4.16 was used to obtain the last line of Equation 4.17 .
The previous description of conditional independence is quite useful from an
operational perspective. It states that the joint conditional probability of and

given Y can be factored as the product of conditional probabilities of
and considered separately. This forms the basis of the naïve Bayes
assumption stated in Equation 4.15 .

How a Naïve Bayes Classifier Works
Using the naïve Bayes assumption, we only need to estimate the conditional
probability of each given Y separately, instead of computing the class-
conditional probability for every combination of attribute values. For example,
if and denote the number of training instances belonging to class 0
with and , respectively, then the class-conditional probability can
be estimated as

In the previous equation, we only need to count the number of training
instances for every one of the k values of an attribute X, irrespective of the
values of other attributes. Hence, the number of parameters needed to learn
class-conditional probabilities is reduced from to dk. This greatly simplifies
the expression for the class-conditional probability and makes it more
amenable to learning parameters and making predictions, even in high-
dimensional settings.

The naïve Bayes classifier computes the posterior probability for a test
instance by using the following equation:

X2 X1



ni0 nj0
X1=ci X2=cj

P(X1=ci, X2=xj|Y=0)=ni0n0×nj0n0.


P(y|x)=P(y)∏i=1dP(xi|y)P(x) (4.18)

Since P ( ) is fixed for every y and only acts as a normalizing constant to
ensure that , we can write

Hence, it is sufficient to choose the class that maximizes .

One of the useful properties of the naïve Bayes classifier is that it can easily
work with incomplete information about data instances, when only a subset of
attributes are observed at every instance. For example, if we only observe p
out of d attributes at a data instance, then we can still compute

using those p attributes and choose the class with the
maximum value. The naïve Bayes classifier can thus naturally handle missing
values in test instances. In fact, in the extreme case where no attributes are
observed, we can still use the prior probability P(y) as an estimate of the
posterior probability. As we observe more attributes, we can keep refining the
posterior probability to better reflect the likelihood of observing the data

In the next two subsections, we describe several approaches for estimating
the conditional probabilities for categorical and continuous attributes
from the training set.

Estimating Conditional Probabilities for
Categorical Attributes
For a categorical attribute , the conditional probability is estimated
according to the fraction of training instances in class y where takes on a
particular categorical value c.

P(y|x)∈[0, 1]





Xi P(Xi=c|y)

where n is the number of training instances belonging to class y, out of which
number of instances have . For example, in the training set given in

Figure 4.8 , seven people have the class label , out
of which three people have while the remaining four have

. As a result, the conditional probability for
is equal to 3/7. Similarly, the

conditional probability for defaulted borrowers with is
given by . Note that the
sum of conditional probabilities over all possible outcomes of is equal to
one, i.e., .

Figure 4.8.
Training set for predicting the loan default problem.


nc Xi=c
Defaulted Borrower=No

Home Owner=Yes
Home Owner=No
P(Home Owner=Yes|Defaulted Borrower=No)

Marital Status=Single
P(Marital Status=Single|Defaulted Borrower=Yes)=2/3


Estimating Conditional Probabilities for
Continuous Attributes
There are two ways to estimate the class-conditional probabilities for
continuous attributes:

1. We can discretize each continuous attribute and then replace the
continuous values with their corresponding discrete intervals. This
approach transforms the continuous attributes into ordinal attributes,
and the simple method described previously for computing the
conditional probabilities of categorical attributes can be employed. Note
that the estimation error of this method depends on the discretization
strategy (as described in Section 2.3.6 on page 63), as well as the
number of discrete intervals. If the number of intervals is too large,
every interval may have an insufficient number of training instances to
provide a reliable estimate of . On the other hand, if the number
of intervals is too small, then the discretization process may loose
information about the true distribution of continuous values, and thus
result in poor predictions.

2. We can assume a certain form of probability distribution for the
continuous variable and estimate the parameters of the distribution
using the training data. For example, we can use a Gaussian
distribution to represent the conditional probability of continuous
attributes. The Gaussian distribution is characterized by two
parameters, the mean, , and the variance, . For each class , the
class-conditional probability
for attribute is


μ σ2 yj

P(Xi=xi|Y=yj)=12πσijexp[−(xi−μij)22σij2 ]. (4.19)

The parameter can be estimated using the sample mean of
for all training instances that belong to . Similarly, can be
estimated from the sample variance of such training instances. For
example, consider the annual income attribute shown in Figure 4.8 .
The sample mean and variance for this attribute with respect to the
class are

Given a test instance with taxable income equal to $120K, we can use
the following value as its conditional probability given class :

Example 4.5. [Naïve Bayes Classifier]
Consider the data set shown in Figure 4.9(a) , where the target class is
Defaulted Borrower, which can take two values Yes and No. We can
compute the class-conditional probability for each categorical attribute and
the sample mean and variance for the continuous attribute, as summarized
in Figure 4.9(b) .

We are interested in predicting the class label of a test instance
. To do

this, we first compute the prior probabilities by counting the number of
training instances belonging to every class. We thus obtain and

. Next, we can compute the class-conditional probability as

μij Xi(x¯)
yj σij2




(Home Owner=No, Marital Status=Married, Annual Income=$120K)


Figure 4.9.
The naïve Bayes classifier for the loan classification problem.

Notice that the class-conditional probability for class has become 0
because there are no instances belonging to class with

in the training set. Using these class-conditional
probabilities, we can estimate the posterior probabilities as

where is a normalizing constant. Since , the
instance is classified as .

P(x|NO)=P(Home Owner=No|No)×P(Status=Married|No)×P(Annual Income



α=1/P(x) P(No|x)>P(Yes|x)

Handling Zero Conditional Probabilities
The preceding example illustrates a potential problem with using the naïve
Bayes assumption in estimating class-conditional probabilities. If the
conditional probability for any of the attributes is zero, then the entire
expression for the class-conditional probability becomes zero. Note that zero
conditional probabilities arise when the number of training instances is small
and the number of possible values of an attribute is large. In such cases, it
may happen that a combination of attribute values and class labels are never
observed, resulting in a zero conditional probability.

In a more extreme case, if the training instances do not cover some
combinations of attribute values and class labels, then we may not be able to
even classify some of the test instances. For example, if

is zero instead of 1/7, then a data instance
with attribute set

has the
following class-conditional probabilities:

Since both the class-conditional probabilities are 0, the naïve Bayes classifier
will not be able to classify the instance. To address this problem, it is important
to adjust the conditional probability estimates so that they are not as brittle as
simply using fractions of training instances. This can be achieved by using the
following alternate estimates of conditional probability:

P(Marital Status=Divorced|No)

(Home Owner=Yes, Marital Status=Divorced, Income=$120K)


Laplace estimate:P(Xi=c|y)=nc+1n+v, (4.20)

m-estimate:P(Xi=c|y)=nc+mpn+m, (4.21)

where n is the number of training instances belonging to class y, is the
number of training instances with and , v is the total number of
attribute values that can take, p is some initial estimate of that is
known a priori, and m is a hyper-parameter that indicates our confidence in
using p when the fraction of training instances is too brittle. Note that even if

, both Laplace and m-estimate provide non-zero values of conditional
probabilities. Hence, they avoid the problem of vanishing class-conditional
probabilities and thus generally provide more robust estimates of posterior

Characteristics of Naïve Bayes Classifiers
1. Naïve Bayes classifiers are probabilistic classification models that are

able to quantify the uncertainty in predictions by providing posterior
probability estimates. They are also generative classification models as
they treat the target class as the causative factor for generating the
data instances. Hence, apart from computing posterior probabilities,
naïve Bayes classifiers also attempt to capture the underlying
mechanism behind the generation of data instances belonging to every
class. They are thus useful for gaining predictive as well as descriptive

2. By using the naïve Bayes assumption, they can easily compute class-
conditional probabilities even in high-dimensional settings, provided
that the attributes are conditionally independent of each other given the
class labels. This property makes naïve Bayes classifier a simple and
effective classification technique that is commonly used in diverse
application problems, such as text classification.

3. Naïve Bayes classifiers are robust to isolated noise points because
such points are not able to significantly impact the conditional
probability estimates, as they are often averaged out during training.

Xi=c Y=y

Xi P(Xi=c|y)


4. Naïve Bayes classifiers can handle missing values in the training set by
ignoring the missing values of every attribute while computing its
conditional probability estimates. Further, naïve Bayes classifiers can
effectively handle missing values in a test instance, by using only the
non-missing attribute values while computing posterior probabilities. If
the frequency of missing values for a particular attribute value depends
on class label, then this approach will not accurately estimate posterior

5. Naïve Bayes classifiers are robust to irrelevant attributes. If is an
irrelevant attribute, then becomes almost uniformly distributed
for every class y. The class-conditional probabilities for every class
thus receive similar contributions of , resulting in negligible
impact on the posterior probability estimates.

6. Correlated attributes can degrade the performance of naïve Bayes
classifiers because the naïve Bayes assumption of conditional
independence no longer holds for such attributes. For example,
consider the following probabilities:

where A is a binary attribute and Y is a binary class variable. Suppose
there is another binary attribute B that is perfectly correlated with A
when , but is independent of A when . For simplicity, assume
that the conditional probabilities for B are the same as for A. Given an
instance with attributes , and assuming conditional
independence, we can compute its posterior probabilities as follows:

If , then the naïve Bayes classifier would assign the instance
to class 1. However, the truth is,




Y=0 Y=1

A=0, B=0

P(Y=0|A=0, B=0)=P(A=0|Y=0)P(B=0|Y=0)P(Y=0)P(A=0, B=0)=0.16×P(Y


P(A=0, B=0|Y=0)=P(A=0|Y=0)=0.4,

because A and B are perfectly correlated when . As a result, the posterior
probability for is

which is larger than that for . The instance should have been classified as
class 0. Hence, the naïve Bayes classifier can produce incorrect results when
the attributes are not conditionally independent given the class labels. Naïve
Bayes classifiers are thus not well-suited for handling redundant or interacting


P(Y=0|A=0, B=0)=P(A=0, B=0|Y=0)P(Y=0)P(A=0, B=0)=0.4×P(Y=0)P(A=0, B=


4.5 Bayesian Networks
The conditional independence assumption made by naïve Bayes classifiers
may seem too rigid, especially for classification problems where the attributes
are dependent on each other even after conditioning on the class labels. We
thus need an approach to relax the naïve Bayes assumption so that we can
capture more generic representations of conditional independence among

In this section, we present a flexible framework for modeling probabilistic
relationships between attributes and class labels, known as Bayesian
Networks. By building on concepts from probability theory and graph theory,
Bayesian networks are able to capture more generic forms of conditional
independence using simple schematic representations. They also provide the
necessary computational structure to perform inferences over random
variables in an efficient way. In the following, we first describe the basic
representation of a Bayesian network, and then discuss methods for
performing inference and learning model parameters in the context of

4.5.1 Graphical Representation

Bayesian networks belong to a broader family of models for capturing
probabilistic relationships among random variables, known as probabilistic
graphical models. The basic concept behind these models is to use
graphical representations where the nodes of the graph correspond to random
variables and the edges between the nodes express probabilistic

relationships. Figures 4.10(a) and 4.10(b) show examples of
probabilistic graphical models using directed edges (with arrows) and
undirected edges (without arrows), respectively. Directed graphical models
are also known as Bayesian networks while undirected graphical models are
known as Markov random fields. The two approaches use different
semantics for expressing relationships among random variables and are thus
useful in different contexts. In the following, we briefly describe Bayesian
networks that are useful in the context of classification.

A Bayesian network (also referred to as a belief network) involves directed
edges between nodes, where every edge represents a direction of influence
among random variables. For example, Figure 4.10(a) shows a Bayesian
network where variable C depends upon the values of variables A and B, as
indicated by the arrows pointing toward C from A and B. Consequently, the
variable C influences the values of variables D and E. Every edge in a
Bayesian network thus encodes a dependence relationship between random
variables with a particular directionality.

Figure 4.10.
Illustrations of two basic types of graphical models.

Bayesian networks are directed acyclic graphs (DAG) because they do not
contain any directed cycles such that the influence of a node loops back to the
same node. Figure 4.11 shows some examples of Bayesian networks that
capture different types of dependence structures among random variables. In
a directed acyclic graph, if there is a directed edge from X to Y ,then X is
called the parent of Y and Y is called the child of X. Note that a node can
have multiple parents in a Bayesian network, e.g., node D has two parent
nodes, B and C, in Figure 4.11(a) . Furthermore, if there is a directed path
in the network from X to Z, then X is an ancestor of Z, while Z is a
descendant of X. For example, in the diagram shown in Figure 4.11(b) , A
is a descendant of D and D is an ancestor of B. Note that there can be
multiple directed paths between two nodes of a directed acyclic graph, as is
the case for nodes A and D in Figure 4.11(a) .

Figure 4.11.
Examples of Bayesian networks.

Conditional Independence
An important property of a Bayesian network is its ability to represent varying
forms of conditional independence among random variables. There are
several ways of describing the conditional independence assumptions
captured by Bayesian networks. One of the most generic ways of expressing
conditional independence is the concept of d-separation, which can be used
to determine if any two sets of nodes A and B are conditionally independent
given another set of nodes C. Another useful concept is that of the Markov
blanket of a node Y , which denotes the minimal set of nodes X that makes Y
independent of the other nodes in the graph, when conditioned on X. (See
Bibliographic Notes for more details on d-separation and Markov blanket.)
However, for the purpose of classification, it is sufficient to describe a simpler
expression of conditional independence in Bayesian networks, known as the
local Markov property.

Property 1 (Local Markov Property).
A node in a Bayesian network is conditionally independent of its
non-descendants, if its parents are known.

To illustrate the local Markov property, consider the Bayes network shown in
Figure 4.11(b) . We can state that A is conditionally independent of both B
and D given C, because C is the parent of A and nodes B and D are non-
descendants of A. The local Markov property helps in interpreting parent-child
relationships in Bayesian networks as representations of conditional
probabilities. Since a node is conditionally independent of its non-descendants

given it parents, the conditional independence assumptions imposed by a
Bayesian network is often sparse in structure. Nonetheless, Bayesian
networks are able to express a richer class of conditional independence
statements among attributes and class labels than the naïve Bayes classifier.
In fact, the naïve Bayes classifier can be viewed as a special type of Bayesian
network, where the target class Y is at the root of a tree and every attribute
is connected to the root node by a directed edge, as shown in Figure
4.12(a) .

Figure 4.12.
Comparing the graphical representation of a naïve Bayes classifier with that of
a generic Bayesian network.

Note that in a naïve Bayes classifier, every directed edge points from the
target class to the observed attributes, suggesting that the class label is a
factor behind the generation of attributes. Inferring the class label can thus be
viewed as diagnosing the root cause behind the observed attributes. On the
other hand, Bayesian networks provide a more generic structure of
probabilistic relationships, since the target class is not required to be at the
root of a tree but can appear anywhere in the graph, as shown in Figure


4.12(b) . In this diagram, inferring Y not only helps in diagnosing the factors
influencing and , but also helps in predicting the influence of and .

Joint Probability
The local Markov property can be used to succinctly express the joint
probability of the set of random variables involved in a Bayesian network. To
realize this, let us first consider a Bayesian network consisting of d nodes,
to , where the nodes have been numbered in such a way that is an
ancestor of only if . The joint probability of can be
generically factorized using the chain rule of probability as

By the way we have constructed the graph, note that the set
contains only non-descendants of . Hence, by using the local Markov
property, we can write as , where denotes
the parents of . The joint probability can then be represented as

It is thus sufficient to represent the probability of every node in terms of its
parent nodes, , for computing P( ). This is achieved with the help of
probability tables that associate every node to its parent nodes as follows:

1. The probability table for node contains the conditional probability
values for every combination of values in and .

2. If has no parents , then the table contains only the prior
probability .

X3 X4 X1 X2

Xd Xi

Xj i<j X={X1, …, Xd}

P(X)=P(X1)P(X2|X1)P(X3|X1, X2) … P(Xd|X1, … Xd−1)=∏i=1dP(Xi|X1, … Xi


{X1, … Xi−1 }

P(Xi|X1, … Xi−1) P(Xi|pa(Xi)) pa(Xi)

P(X)=∏i=1dP(Xi|pa(Xi)) (4.23)


P(Xi|pa(Xi)) Xi pa(Xi)

Xi (pa(Xi)=ϕ)

Example 4.6. [Probability Tables]
Figure 4.13 shows an example of a Bayesian network for modeling the
relationships between a patient’s symptoms and risk factors. The
probability tables are shown at the side of every node in the figure. The
probability tables associated with the risk factors (Exercise and Diet)
contain only the prior probabilities, whereas the tables for heart disease,
heartburn, blood pressure, and chest pain, contain the conditional

Figure 4.13.
A Bayesian network for detecting heart disease and heartburn in patients.

Use of Hidden Variables

A Bayesian network typically involves two types of variables: observed
variables that are clamped to specific observed values, and unobserved
variables, whose values are not known and need to be inferred from the
network. To distinguish between these two types of variables, observed
variables are generally represented using shaded nodes while unobserved
variables are represented using empty nodes. Figure 4.14 shows an
example of a Bayesian network with observed variables (A, B, and E ) and
unobserved variables (C and D).

Figure 4.14.
Observed and unobserved variables are represented using unshaded and
shaded circles, respectively.

In the context of classification, the observed variables correspond to the set of
attributes X, while the target class is represented using an unobserved
variable Y that needs to be inferred during testing. However, note that a
generic Bayesian network may contain many other unobserved variables
apart from the target class, as represented in Figure 4.15 as the set of
variables H. These unobserved variables represent hidden or confounding
factors that affect the probabilities of attributes and class labels, although they
are never directly observed. The use of hidden variables enhances the
expressive power of Bayesian networks in representing complex probabilistic

relationships between attributes and class labels. This is one of the key
distinguishing properties of Bayesian networks as compared to naïve Bayes

4.5.2 Inference and Learning

Given the probability tables corresponding to every node in a Bayesian
network, the problem of inference corresponds to computing the probabilities
of different sets of random variables. In the context of classification, one of the
key inference problems is to compute the probability of a target class Y taking
on a specific value y, given the set of observed attributes at a data instance,
. This can be represented using the following conditional probability:

The previous equation involves marginal probabilities of the form P(y, ).
They can be computed by marginalizing out the hidden variables H from the
joint probability as follows:

where the joint probability P(y, , H) can be obtained by using the
factorization described in Equation 4.23 . To understand the nature of
computations involved in estimating P(y, ), consider the example Bayesian
network shown in Figure 4.15 , which involves a target class, Y , three
observed attributes, to , and four hidden variables, to . For this
network, we can express P(y, ) as

P(Y=y|x)=(y, x)P(x)=(y, x)∑y′P(y′, x) (4.24)

P(y, x)=∑HP(y, x, H), (4.25)

X1 X3 H1 H4

Figure 4.15.
An example of a Bayesian network with four hidden variables, to , three
observed attributes, to , and one target class Y .

where f is a factor that depends on the values of to . In the previous
simplistic expression of P(y, ), a different summand is considered for every
combination of values, to , in the hidden variables, to . If we
assume that every variable in the network can take k discrete values, then the
summation has to be carried out for a total number of times. The
computational complexity of this approach is thus . Moreover, the
number of computations grows exponentially with the number of hidden
variables, making it difficult to use this approach with networks that have a
large number of hidden variables. In the following, we present different
computational techniques for efficiently performing inferences in Bayesian

H1 H4
X1 X3

P(y, x)=∑h1∑h2∑h3∑h4P(y, x1, x2, h1, h2, h3, h4),=∑h1∑h2∑h3∑h4
[P(h1)P(h2)P(x2)P(h4)P(x1|h1, h2) ×P(h3|x2, h2)P(y|x1, h3)P(x3|h3, h4) ],


=∑h1∑h2∑h3∑h4f(h1, h2, h3, h4), (4.27)

h1 h4

h1 h4 H1 H4


Variable Elimination
To reduce the number of computations involved in estimating P(y, ), let us
closely examine the expressions in Equations 4.26 and 4.27 . Notice that
although depends on the values of all four hidden variables, it
can be decomposed as a product of several smaller factors, where every
factor involves only a small number of hidden variables. For example, the
factor depends only on the value of , and thus acts as a constant
multiplicative term when summations are performed over , or .
Hence, if we place outside the summations of to , we can save
some repeated multiplications occurring inside every summand.

In general, we can push every summation as far inside as possible, so that
the factors that do not depend on the summing variable are placed outside the
summation. This will help reduce the number of wasteful computations by
using smaller factors at every summation. To illustrate this process, consider
the following sequence of steps for computing P(y, ), by rearranging the

of summations in Equation 4.26 .

where represents the intermediate factor term obtained by summing out .
To check if the previous rearrangements provide any improvements in

f(h1, h2, h3, h4)

P(h4) h4
h1, h2 h3

P(h4) h1 h3

P(y, x)=P(x2)∑h4P(h4)∑h3P(y|x1, h3)P(x3|h3, h4)×∑h2P(h2)P(h3|x2, h2)∑h1P(4.28)

=P(x2)∑h4P(h4)∑h3P(y|x1, h3)P(x3|h3, h4)×∑h2P(h2)P(h3|x2, h2)f1(h2)(4.29)

=P(x2)∑h4P(h4)∑h3P(y|x1, h3)P(x3|h3, h4)f2(h3) (4.30)

=P(x2)∑h4P(h4)f3(h4) (4.31)

fi hi

computational efficiency, let us count the number of computations occurring at
every step of the process. At the first step (Equation 4.28 ), we perform a
summation over using factors that depend on and . This requires
considering every pair of values in and , resulting in computations.
Similarly, the second step (Equation 4.29 ) involves summing out using
factors of and , leading to computations. The third step (Equation
4.30 ) again requires computations as it involves summing out
over factors depending on and . Finally, the fourth step (Equation
4.31 ) involves summing out using factors depending on , resulting in
O(k) computations.

The overall complexity of the previous approach is thus , which is
considerably smaller than the complexity of the basic approach. Hence,
by merely rearranging summations and using algebraic manipulations, we are
able to improve the computational efficiency in computing P(y, ). This
procedure is known as variable elimination.

The basic concept that variable elimination exploits to reduce the number of
computations is the distributive nature of multiplication over addition
operations. For example, consider the following multiplication and addition

Notice that the right-hand side of the previous equation involves three
multiplications and three additions, while the left-hand side involves only one
multiplication and three additions, thus saving on two arithmetic operations.
This property is utilized by variable elimination in pushing out constant terms
outside the summation, such that they are multiplied only once.

h1 h1 h2
h1 h2 O(k2)

h2 h3 O(k2)

O(k2) h3
h3 h4

h4 h4



Note that the efficiency of variable elimination depends on the order of hidden
variables used for performing summations. Hence, we would ideally like to
find the optimal order of variables that result in the smallest number of
computations. Unfortunately, finding the optimal order of summations for a
generic Bayesian network is an NP-Hard problem, i.e., there does not exist an
efficient algorithm for finding the optimal ordering that can run in polynomial
time. However, there exists efficient techniques for handling special types of
Bayesian networks, e.g., those involving tree-like graphs, as described in the

Sum-Product Algorithm for Trees
Note that in Equations 4.28 and 4.29 , whenever a variable is
eliminated during marginalization, it results in the creation of a factor that
depends on the neighboring nodes of . is then absorbed in the factors of
neighboring variables and the process is repeated until all unobserved
variables have marginalized. This phenomena of variable elimination can be
viewed as transmitting a local message from the variable being marginalized
to its neighboring nodes. This idea of message passing utilizes the structure
of the graph for performing computations, thus making it possible to use
graph-theoretic approaches for making effective inferences. The sum-
product algorithm builds on the concept of message passing for computing
marginal and conditional probabilities on tree-based graphs.

Figure 4.16 shows an example of a tree involving five variables, to .
A key characteristic of a tree is that every node in the tree has exactly one
parent, and there is only one directed edge between any two nodes in the
tree. For the purpose of illustration, let us consider the problem of estimating
the marginal probability of . This can be obtained by marginalizing
out every variable in the graph except and rearranging the summations to
obtain the following expression:


hi fi

X1 X5

X2, P(X2)

Figure 4.16.
An example of a Bayesian network with a tree structure.

where has been conveniently chosen to represent the factor of that is
obtained by summing out . We can view as a local message passed
from node to node , as shown using arrows in Figure 4.17(a) . These
local messages capture the influence of eliminating nodes on the marginal
probabilities of neighboring nodes.

Before we formally describe the formula for computing and , we
first define a potential function that is associated every node and edge of
the graph. We can define the potential of a node as


mij(xj) xj
xi mij(xj)

xi xj

mij(xj) P(xj)


ψ(Xi)={P(Xi),if Xi is the root node.1,otherwise. (4.32)

Figure 4.17.
Illustration of message passing in the sum-product algorithm.

Similarly, we can define the potential of an edge between nodes and
(where is the parent of ) as

Using and , we can represent using the following

where N(i) represents the set of neighbors of node . The message that is
transmitted from to can thus be recursively computed using the

Xi Xj
Xi Xj

ψ(Xi, Xj)=P(Xj|Xi).

ψ(Xi) ψ(Xi, Xj) mij(xj)

mij(xj)=∑xi(ψ(xi)ψ(xi, xj)∏k∈N(i)imki(xi)), (4.33)

Xi mij
Xi Xj

messages incident on from its neighboring nodes excluding . Note that
the formula for involves taking a sum over all possible values of , after
multiplying the factors obtained from the neighbors of . This approach of
message passing is thus called the “sum-product” algorithm. Further, since
represents a notion of “belief” propagated from to , this algorithm is also
known as belief propagation. The marginal probability of a node

is then given as

A useful property of the sum-product algorithm is that it allows the messages
to be reused for computing a different marginal probability in the future. For
example, if we had to compute the marginal probability for node , we would
require the following messages from its neighboring nodes: ,
and . However, note that , and have already been
computed in the process of computing the marginal probability of and thus
can be reused.

Notice that the basic operations of the sum-product algorithm resemble a
message passing protocol over the edges of the network. A node sends out a
message to all its neighboring nodes only after it has received incoming
messages from all its neighbors. Hence, we can initialize the message
passing protocol from the leaf nodes, and transmit messages till we reach the
root node. We can then run a second pass of messages from the root node
back to the leaf nodes. In this way, we can compute the messages for every
edge in both directions, using just operations, where is the number
of edges. Once we have transmitted all possible messages as shown in
Figure 4.17(b) , we can easily compute the marginal probability of every
node in the graph using Equation 4.34 .

Xi Xi
mij Xj


Xi Xj

P(xi)=ψ(xi)∏j∈N(i)mji(xi). (4.34)

m23(x3), m43(x3)

m53(x3) m43(x3) m53(x3)

O(2|E|) |E|

In the context of classification, the sum-product algorithm can be easily
modified for computing the conditional probability of the class label y given the
set of observed attributes , i.e., . This basically amounts to
computing in Equation 4.24 , where X is clamped to the
observed values . To handle the scenario where some of the random
variables are fixed and do not need to be normalized, we consider the
following modification.

If is a random variable that is fixed to a specific value , then we can
simply modify and as follows:

We can run the sum-product algorithm using these modified values for every
observed variable and thus compute .

x^ P(y|x^)
P(y, X=x^)


Xi x^i
ψ(Xi) ψ(Xi, Xj)

ψ(Xi)={1,if Xi=x^i.0,otherwise. (4.35)

ψ(Xi, Xj)={P(Xi|x^i),if Xi=x^i.0,otherwise. (4.36)

P(y, X=x^)

Figure 4.18.
Example of a poly-tree and its corresponding factor graph.

Generalizations for Non-Tree Graphs
The sum-product algorithm is guaranteed to optimally converge in the case of
trees using a single run of message passing in both directions of every edge.
This is because any two nodes in a tree have a unique path for the
transmission of messages. Furthermore, since every node in a tree has a
single parent, the joint probability involves only factors of at most two
variables. Hence, it is sufficient to consider potentials over edges and not
other generic substructures in the graph.

Both of the previous properties are violated in graphs that are not trees, thus
making it difficult to directly apply the sum-product algorithm for making
inferences. However, a number of variants of the sum-product algorithm have
been devised to perform inferences on a broader family of graphs than trees.
Many of these variants transform the original graph into an alternative tree-
based representation, and then apply the sum-product algorithm on the
transformed tree. In this section, we briefly discuss one such transformations
known as factor graphs.

Factor graphs are useful for making inferences over graphs that violate the
condition that every node has a single parent. Nonetheless, they still require
the absence of multiple paths between any two nodes, to guarantee
convergence. Such graphs are known as poly-trees. An example of a poly-
tree is shown in Figure 4.18(a) .

A poly-tree can be transformed into a tree-based representation with the help
of factor graphs. These graphs consist of two types of nodes, variables nodes
(that are represented using circles) and factor nodes (that are represented

using squares). The factor nodes represent conditional independence
relationships among the variables of the poly-tree. In particular, every
probability table can be represented as a factor node. The edges in a factor
graph are undirected in nature and relate a variable node to a factor node if
the variable is involved in the probability table corresponding to the factor
node. Figure 4.18(b) presents the factor graph representation of the poly-
tree shown in Figure 4.18(a) .

Note that the factor graph of a poly-tree always forms a tree-like structure,
where there is a unique path of influence between any two nodes in the factor
graph. Hence, we can apply a modified form of sum-product algorithm to
transmit messages between variable nodes and factor nodes, which is
guaranteed to converge to optimal values.

Learning Model Parameters
In all our previous discussions on Bayesian networks, we had assumed that
the topology of the Bayesian network and the values in the probability tables
of every node were already known. In this section, we discuss approaches for
learning both the topology and the probability table values of a Bayesian
network from the training data.

Let us first consider the case where the topology of the network is known and
we are only required to compute the probability tables. If there are no
unobserved variables in the training data, then we can easily compute the
probability table for , by counting the fraction of training instances
for every value of and every combination of values in . However, if
there are unobserved variables in or , then computing the fraction of
training instances for such variables is non-trivial and requires the use of
advances techniques such as the Expectation-Maximization algorithm
(described later in Chapter 8 ).

Xi pa(Xi)

Xi pa(Xi)

Learning the structure of the Bayesian network is a much more challenging
task than learning the probability tables. Although there are some scoring
approaches that attempt to find a graph structure that maximizes the training
likelihood, they are often computationally infeasible when the graph is large.
Hence, a common approach for constructing Bayesian networks is to use the
subjective knowledge of domain experts.

4.5.3 Characteristics of Bayesian

1. Bayesian networks provide a powerful approach for representing
probabilistic relationships between attributes and class labels with the
help of graphical models. They are able to capture complex forms of
dependencies among variables. Apart from encoding prior beliefs, they
are also able to model the presence of latent (unobserved) factors as
hidden variables in the graph. Bayesian networks are thus quite
expressive and provide predictive as well as descriptive insights about
the behavior of attributes and class labels.

2. Bayesian networks can easily handle the presence of correlated or
redundant attributes, as opposed to the naïve Bayes classifier. This is
because Bayesian networks do not use the naïve Bayes assumption
about conditional independence, but instead are able to express richer
forms of conditional independence.

3. Similar to the naïve Bayes classifier, Bayesian networks are also quite
robust to the presence of noise in the training data. Further, they can
handle missing values during training as well as testing. If a test
instance contains an attribute with a missing value, then a Bayesian
network can perform inference by treating as an unobserved node


and marginalizing out its effect on the target class. Hence, Bayesian
networks are well-suited for handling incompleteness in the data, and
can work with partial information. However, unless the pattern with
which missing values occurs is completely random, then their presence
will likely introduce some degree of error and/or bias into the analysis.

4. Bayesian networks are robust to irrelevant attributes that contain no
discriminatory information about the class labels. Such attributes show
no impact on the conditional probability of the target class, and are thus
rightfully ignored.

5. Learning the structure of a Bayesian network is a cumbersome task
that often requires assistance from expert knowledge. However, once
the structure has been decided, learning the parameters of the network
can be quite straightforward, especially if all the variables in the
network are observed.

6. Due to its additional ability of representing complex forms of
relationships, Bayesian networks are more susceptible to overfitting as
compared to the naïve Bayes classifier. Furthermore, Bayesian
networks typically require more training instances for effectively
learning the probability tables than the naïve Bayes classifier.

7. Although the sum-product algorithm provides computationally efficient
techniques for performing inference over tree-like graphs, the
complexity of the approach increase significantly when dealing with
generic graphs of large sizes. In situations where exact inference is
computationally infeasible, it is quite common to use approximate
inference techniques.

4.6 Logistic Regression
The naïve Bayes and the Bayesian network classifiers described in the
previous sections provide different ways of estimating the conditional
probability of an instance given class y, . Such models are known as
probabilistic generative models. Note that the conditional probability
essentially describes the behavior of instances in the attribute space that are
generated from class y. However, for the purpose of making predictions, we
are finally interested in computing the posterior probability . For
example, computing the following ratio of posterior probabilities is sufficient for
inferring class labels in a binary classification problem:

This ratio is known as the odds. If this ratio is greater than 1, then is
classified as . Otherwise, it is assigned to class . Hence, one may
simply learn a model of the odds based on the attribute values of training
instances, without having to compute as an intermediate quantity in the
Bayes theorem.

Classification models that directly assign class labels without computing class-
conditional probabilities are called discriminative models. In this section, we
present a probabilistic discriminative model known as logistic regression,
which directly estimates the odds of a data instance using its attribute
values. The basic idea of logistic regression is to use a linear predictor,

, for representing the odds of as follows:




y=1 y=0



P(y=1|x)P(y=0|x)=ez=ewTx+b, (4.37)

where and b are the parameters of the model and denotes the transpose
of a vector . Note that if , then belongs to class 1 since its odds
is greater than 1. Otherwise, belongs to class 0.

Figure 4.19.
Plot of sigmoid (logistic) function, .

Since , we can re-write Equation 4.37 as

This can be further simplified to express as a function of z.

where the function is known as the logistic or sigmoid function. Figure
4.19 shows the behavior of the sigmoid function as we vary z. We can see
that only when . We can also derive using as






P(y=1|x)=11+e−z=σ(z), (4.38)


σ(z)≥0.5 z≥0 P(y=0|x) σ(z)

Hence, if we have learned a suitable value of parameters and b, we can
use Equations 4.38 and 4.39 to estimate the posterior probabilities of
any data instance and determine its class label.

4.6.1 Logistic Regression as a
Generalized Linear Model

Since the posterior probabilities are real-valued, their estimation using the
previous equations can be viewed as solving a regression problem. In fact,
logistic regression belongs to a broader family of statistical regression models,
known as generalized linear models (GLM). In these models, the target
variable y is considered to be generated from a probability distribution ,
whose mean can be estimated using a link function as follows:

For binary classification using logistic regression, y follows a Bernoulli
distribution (y can either be 0 or 1) and is equal to . The link
function of logistic regression, called the logit function, can thus be
represented as

Depending on the choice of link function and the form of probability
distribution , GLMs are able to represent a broad family of regression
models, such as linear regression and Poisson regression. They require

P(y=0|x)=1−σ(z)=11+e−z (4.39)

μ g(⋅)

g(μ)=z=wT x + b. (4.40)

μ P(y=1|x)



different approaches for estimating their model parameters, ( , ). In this
chapter, we will only discuss approaches for estimating the model parameters
of logistic regression, although methods for estimating parameters of other
types of GLMs are often similar (and sometimes even simpler). (See
Bibliographic Notes for more details on GLMs.)

Note that even though logistic regression has relationships with regression
models, it is a classification model since the computed posterior probabilities
are eventually used to determine the class label of a data instance.

4.6.2 Learning Model Parameters

The parameters of logistic regression, ( , ), are estimated during training
using a statistical approach known as the maximum likelihood estimation
(MLE) method. This method involves computing the likelihood of observing
the training data given ( , ), and then determining the model parameters

that yield maximum likelihood.

Let denote a set of n training
instances, where is a binary variable (0 or 1). For a given training instance
, we can compute its posterior probabilities using Equations 4.38 and

4.39 . We can then express the likelihood of observing given , , and b

where is the sigmoid function as described above, Equation 4.41
basically means that the likelihood is equal to when

(w*, b*)

D.train={(x1, y1), (x2, y2), … , (xn, yn)}

yi xi

P(yi|xi, w, b)=P(y=1|xi)yi×P(y=0|xi)1−yi,=(σ(zi))yi×(1−σ(zi))1−yi,=


P(yi|xi, w, b) P(y=1|xi)

, and equal to when . The likelihood of all training instances,
, can then be computed by taking the product of individual likelihoods

(assuming independence among training instances) as follows:

The previous equation involves multiplying a large number of probability
values, each of which are smaller than or equal to 1. Since this naïve
computation can easily become numerically unstable when n is large, a more
practical approach is to consider the negative logarithm (to base e) of the
likelihood function, also known as the cross entropy function:

The cross entropy is a loss function that measures how unlikely it is for the
training data to be generated from the logistic regression model with
parameters ( , ). Intuitively, we would like to find model parameters
that result in the lowest cross entropy, .

where is the loss function. It is worth emphasizing that
E( , ) is a convex function, i.e., any minima of E( , ) will be a global
minima. Hence, we can use any of the standard convex optimization
techniques to solve Equation 4.43 , which are mentioned in Appendix E.
Here, we briefly describe the Newton-Raphson method that is commonly used
for estimating the parameters of logistic regression. For ease of
representation, we will use a single vector to describe , which is of
size one greater than . Similarly, we will consider the concatenated feature
vector , such that the linear predictor can be succinctly

yi=1 P(y=0|xi) yi=0
L(w, b)

L(w, b)=∏i=1nP(yi|xi, w, b)=∏i=1nP(y=1|xi)yi×P(y=0|xi)1−yi. (4.42)

−logL(w, b)=−∑i=1nyilog(P(y=1|xi))+(1−yi)log(P(y=0|xi)).=

(w*, b*)
−logL(w*, b*)

(w*, b*)=argmin(w, b)E(w, b)=argmin(w, b)−logL(w, b) (4.43)

E(w, b)=−logL(w, b)

w˜=(wT b)T

x˜=(xT 1)T z=wTx+b

written as . Also, the concatenation of all training labels, to , will
be represented as y, the set consisting of to will be represented as
, and the concatenation of to will be represented as .

The Newton-Raphson is an iterative method for finding that uses the
following equation to update the model parameters at every iteration:

where and H are the first- and second-order derivatives of the loss
function with respect to , respectively. The key intuition behind
Equation 4.44 is to move the model parameters in the direction of
maximum gradient, such that takes larger steps when is large.
When arrives at a minima after some number of iterations, then
would become equal to 0 and thus result in convergence. Hence, we start with
some initial values of (either randomly assigned or set to 0) and use
Equation 4.44 to iteratively update till there are no significant changes in
its value (beyond a certain threshold).

The first-order derivative of is given by

where we have used the fact that . Using , we
can compute the second-order derivative of as

where R is a diagonal matrix whose i diagonal element . We can
now use the first- and second-order derivatives of in Equation 4.44 to

z=w˜Tx˜ y1 yn
σ(z1) σ(zn)

σ x˜1 x˜n X˜


w˜(new)=w˜(old)−H−1∇E(w˜), (4.44)

E( w˜) w˜

w˜ ∇E(w˜)
w˜ ∇E(w˜)




dσ(z)/dz=σ(z)(1−σ(z)) ∇E(w˜)

H=∇∇E(w˜)=∑i=1nσ(w˜Tx˜i)(1−σ(w˜Tx˜i)x˜ix˜iT)=X˜TRX˜, (4.46)

th Rii=σi(1−σi)


obtain the following update equation at the k iteration:

where the subscript k under and refers to using to compute both

4.6.3 Characteristics of Logistic

1. Logistic Regression is a discriminative model for classification that
directly computes the poster probabilities without making any
assumption about the class conditional probabilities. Hence, it is quite
generic and can be applied in diverse applications. It can also be easily
extended to multiclass classification, where it is known as multinomial
logistic regression. However, its expressive power is limited to
learning only linear decision boundaries.

2. Because there are different weights (parameters) for every attribute,
the learned parameters of logistic regression can be analyzed to
understand the relationships between attributes and class labels.

3. Because logistic regression does not involve computing densities and
distances in the attribute space, it can work more robustly even in high-
dimensional settings than distance-based methods such as nearest
neighbor classifiers. However, the objective function of logistic
regression does not involve any term relating to the complexity of the
model. Hence, logistic regression does not provide a way to make a
trade-off between model complexity and training performance, as
compared to other classification models such as support vector


w˜(k+1)=w˜(k)−(X˜TRkX˜)−1X˜T(σk−y) (4.47)

Rk σk w˜(k)

machines. Nevertheless, variants of logistic regression can easily be
developed to account for model complexity, by including appropriate
terms in the objective function along with the cross entropy function.

4. Logistic regression can handle irrelevant attributes by learning weight
parameters close to 0 for attributes that do not provide any gain in
performance during training. It can also handle interacting attributes
since the learning of model parameters is achieved in a joint fashion by
considering the effects of all attributes together. Furthermore, if there
are redundant attributes that are duplicates of each other, then logistic
regression can learn equal weights for every redundant attribute,
without degrading classification performance. However, the presence of
a large number of irrelevant or redundant attributes in high-dimensional
settings can make logistic regression susceptible to model overfitting.

5. Logistic regression cannot handle data instances with missing values,
since the posterior probabilities are only computed by taking a
weighted sum of all the attributes. If there are missing values in a
training instance, it can be discarded from the training set. However, if
there are missing values in a test instance, then logistic regression
would fail to predict its class label.

4.7 Artificial Neural Network (ANN)
Artificial neural networks (ANN) are powerful classification models that are
able to learn highly complex and nonlinear decision boundaries purely from
the data. They have gained widespread acceptance in several applications
such as vision, speech, and language processing, where they have been
repeatedly shown to outperform other classification models (and in some
cases even human performance). Historically, the study of artificial neural
networks was inspired by attempts to emulate biological neural systems. The
human brain consists primarily of nerve cells called neurons, linked together
with other neurons via strands of fiber called axons. Whenever a neuron is
stimulated (e.g., in response to a stimuli), it transmits nerve activations via
axons to other neurons. The receptor neurons collect these nerve activations
using structures called dendrites, which are extensions from the cell body of
the neuron. The strength of the contact point between a dendrite and an axon,
known as a synapse, determines the connectivity between neurons.
Neuroscientists have discovered that the human brain learns by changing the
strength of the synaptic connection between neurons upon repeated
stimulation by the same impulse.

The human brain consists of approximately 100 billion neurons that are inter-
connected in complex ways, making it possible for us to learn new tasks and
perform regular activities. Note that a single neuron only performs a simple
modular function, which is to respond to the nerve activations coming from
sender neurons connected at its dendrite, and transmit its activation to
receptor neurons via axons. However, it is the composition of these simple
functions that together is able to express complex functions. This idea is at the
basis of constructing artificial neural networks.

Analogous to the structure of a human brain, an artificial neural network is
composed of a number of processing units, called nodes, that are connected
with each other via directed links. The nodes correspond to neurons that
perform the basic units of computation, while the directed links correspond to
connections between neurons, consisting of axons and dendrites. Further, the
weight of a directed link between two neurons represents the strength of the
synaptic connection between neurons. As in biological neural systems, the
primary objective of ANN is to adapt the weights of the links until they fit the
input-output relationships of the underlying data.

The basic motivation behind using an ANN model is to extract useful features
from the original attributes that are most relevant for classification.
Traditionally, feature extraction has been achieved by using dimensionality
reduction techniques such as PCA (introduced in Chapter 2), which show
limited success in extracting nonlinear features, or by using hand-crafted
features provided by domain experts. By using a complex combination of
inter-connected nodes, ANN models are able to extract much richer sets of
features, resulting in good classification performance. Moreover, ANN models
provide a natural way of representing features at multiple levels of abstraction,
where complex features are seen as compositions of simpler features. In
many classification problems, modeling such a hierarchy of features turns out
to be very useful. For example, in order to detect a human face in an image,
we can first identify low-level features such as sharp edges with different
gradients and orientations. These features can then be combined to identify
facial parts such as eyes, nose, ears, and lips. Finally, an appropriate
arrangement of facial parts can be used to correctly identify a human face.
ANN models provide a powerful architecture to represent a hierarchical
abstraction of features, from lower levels of abstraction (e.g., edges) to higher
levels (e.g., facial parts).

Artificial neural networks have had a long history of developments spanning
over five decades of research. Although classical models of ANN suffered
from several challenges that hindered progress for a long time, they have re-
emerged with widespread popularity because of a number of recent
developments in the last decade, collectively known as deep learning. In this
section, we examine classical approaches for learning ANN models, starting
from the simplest model called perceptrons to more complex architectures
called multi-layer neural networks. In the next section, we discuss some of
the recent advancements in the area of ANN that have made it possible to
effectively learn modern ANN models with deep architectures.

4.7.1 Perceptron

A perceptron is a basic type of ANN model that involves two types of nodes:
input nodes, which are used to represent the input attributes, and an output
node, which is used to represent the model output. Figure 4.20 illustrates
the basic architecture of a perceptron that takes three input attributes, ,
and , and produces a binary output y. The input node corresponding to an
attribute is connected via a weighted link to the output node. The
weighted link is used to emulate the strength of a synaptic connection
between neurons.

x1, x2

xi wi

Figure 4.20.
Basic architecture of a perceptron.

The output node is a mathematical device that computes a weighted sum of
its inputs, adds a bias factor b to the sum, and then examines the sign of the
result to produce the output as follows:

To simplify notations, and b can be concatenated to form , while
can be appended with 1 at the end to form . The output of the

perceptron can then be written:

where the sign function acts as an activation function by providing an output
value of if the argument is positive and if its argument is negative.

Learning the Perceptron
Given a training set, we are interested in learning parameters such that
closely resembles the true y of training instances. This is achieved by using
the perceptron learning algorithm given in Algorithm 4.3 . The key
computation for this algorithm is the iterative weight update formula given in
Step 8 of the algorithm:

where is the weight parameter associated with the i input link after the
k iteration, is a parameter known as the learning rate, and is the value


3^y={1,if wTx+b>0.−1,otherwise. (4.48)

w˜=(wT b)T
x˜=(xT 1)T



+1 −1

w˜ y^

wj(k+1)=wj(k)+λ(yi−yi^(k))xij, (4.49)

w(k) th

th λ xij


of the j attribute of the training example . The justification for Equation
4.49 is rather intuitive. Note that captures the discrepancy between
and , such that its value is 0 only when the true label and the predicted

output match. Assume is positive. If and , then is increased at
the next iteration so that can become positive. On the other hand, if

and , then is decreased so that can become negative.
Hence, the weights are modified at every iteration to reduce the discrepancies
between and y across all training instances. The learning rate , a
parameter whose value is between 0 and 1, can be used to control the
amount of adjustments made in each iteration. The algorithm halts when the
average number of discrepancies are smaller than a threshold .

Algorithm 4.3 Perceptron learning algorithm.


∑ γ

The perceptron is a simple classification model that is designed to learn linear
decision boundaries in the attribute space. Figure 4.21 shows the decision

th xi

yi y^i
xij y^=0 y=1 wj

y^=1 y=0 wj w˜Txi

y^ λ


boundary obtained by applying the perceptron learning algorithm to the data
set provided on the left of the figure. However, note that there can be multiple
decision boundaries that can separate the two classes, and the perceptron
arbitrarily learns one of these boundaries depending on the random initial
values of parameters. (The selection of the optimal decision boundary is a
problem that will be revisited in the context of support vector machines in
Section 4.9 .) Further, the perceptron learning algorithm is only guaranteed
to converge when the classes are linearly separable. However, if the classes
are not linearly separable, the algorithm fails to converge. Figure 4.22
shows an example of a nonlinearly separable data given by the XOR function.
The perceptron cannot find the right solution for this data because there is no
linear decision boundary that can perfectly separate the training instances.
Thus, the stopping condition at line 12 of Algorithm 4.3 would never be
met and hence, the perceptron learning algorithm would fail to converge. This
is a major limitation of perceptrons since real-world classification problems
often involve nonlinearly separable classes.

Figure 4.21.
Perceptron decision boundary for the data given on the left ( represents a
positively labeled instance while o represents a negatively labeled instance.


Figure 4.22.
XOR classification problem. No linear hyperplane can separate the two

4.7.2 Multi-layer Neural Network

A multi-layer neural network generalizes the basic concept of a perceptron to
more complex architectures of nodes that are capable of learning nonlinear
decision boundaries. A generic architecture of a multi-layer neural network is
shown in Figure 4.23 where the nodes are arranged in groups called
layers. These layers are commonly organized in the form of a chain such that
every layer operates on the outputs of its preceding layer. In this way, the
layers represent different levels of abstraction that are applied on the input
features in a sequential manner. The composition of these abstractions
generates the final output at the last layer, which is used for making
predictions. In the following, we briefly describe the three types of layers used
in multi-layer neural networks.

Figure 4.23.
Example of a multi-layer artificial neural network (ANN).

The first layer of the network, called the input layer, is used for representing
inputs from attributes. Every numerical or binary attribute is typically
represented using a single node on this layer, while a categorical attribute is
either represented using a different node for each categorical value, or by
encoding the k-ary attribute using input nodes. These inputs are fed
into intermediary layers known as hidden layers, which are made up of
processing units known as hidden nodes. Every hidden node operates on
signals received from the input nodes or hidden nodes at the preceding layer,
and produces an activation value that is transmitted to the next layer. The final
layer is called the output layer and processes the activation values from its
preceding layer to produce predictions of output variables. For binary
classification, the output layer contains a single node representing the binary
class label. In this architecture, since the signals are propagated only in the
forward direction from the input layer to the output layer, they are also called
feedforward neural networks.

⌈log2k ⌉

A major difference between multi-layer neural networks and perceptrons is the
inclusion of hidden layers, which dramatically improves their ability to
represent arbitrarily complex decision boundaries. For example, consider the
XOR problem described in the previous section. The instances can be
classified using two hyperplanes that partition the input space into their
respective classes, as shown in Figure 4.24(a) . Because a perceptron can
create only one hyperplane, it cannot find the optimal solution. However, this
problem can be addressed by using a hidden layer consisting of two nodes,
as shown in Figure 4.24(b) . Intuitively, we can think of each hidden node
as a perceptron that tries to construct one of the two hyperplanes, while the
output node simply combines the results of the perceptrons to yield the
decision boundary shown in Figure 4.24(a) .

Figure 4.24.
A two-layer neural network for the XOR problem.

The hidden nodes can be viewed as learning latent representations or
features that are useful for distinguishing between the classes. While the first
hidden layer directly operates on the input attributes and thus captures
simpler features, the subsequent hidden layers are able to combine them and

construct more complex features. From this perspective, multi-layer neural
networks learn a hierarchy of features at different levels of abstraction that are
finally combined at the output nodes to make predictions. Further, there are
combinatorially many ways we can combine the features learned at the
hidden layers of ANN, making them highly expressive. This property chiefly
distinguishes ANN from other classification models such as decision trees,
which can learn partitions in the attribute space but are unable to combine
them in exponential ways.

Figure 4.25.
Schematic illustration of the parameters of an ANN model with hidden

To understand the nature of computations happening at the hidden and output
nodes of ANN, consider the i node at the l layer of the network , where
the layers are numbered from 0 (input layer) to L (output layer), as shown in
Figure 4.25 . The activation value generated at this node, , can be
represented as a function of the inputs received from nodes at the preceding
layer. Let represent the weight of the connection from the j node at layer


th th (l>0)


wijl th


to the i node at layer l. Similarly, let us denote the bias term at this node
as . The activation value can then be expressed as

where z is called the linear predictor and is the activation function that
converts z to a. Further, note that, by definition, at the input layer and

at the output node.

There are a number of alternate activation functions apart from the sign
function that can be used in multi-layer neural networks. Some examples
include linear, sigmoid (logistic), and hyperbolic tangent functions, as shown in
Figure 4.26 . These functions are able to produce real-valued and nonlinear
activation values. Among these activation functions, the sigmoid has been
widely used in many ANN models, although the use of other types of
activation functions in the context of deep learning will be discussed in
Section 4.8 . We can thus represent as

(l−1) th

bjl ail






Figure 4.26.
Types of activation functions used in multi-layer neural networks.

Learning Model Parameters
The weights and bias terms ( , b) of the ANN model are learned during
training so that the predictions on training instances match the true labels.
This is achieved by using a loss function

ail=σ(zil)=11+e−zil. (4.50)

E(w, b)=∑k=1nLoss (yk, y^k) (4.51)

where is the true label of the kth training instance and is equal to ,
produced by using . A typical choice of the loss function is the squared loss

Note that E( , b) is a function of the model parameters ( , b) because the
output activation value depends on the weights and bias terms. We are
interested in choosing ( , b) that minimizes the training loss E( , b).
Unfortunately, because of the use of hidden nodes with nonlinear activation
functions, E( , b) is not a convex function of and b, which means that E( ,
b) can have local minima that are not globally optimal. However, we can still
apply standard optimization techniques such as the gradient descent
method to arrive at a locally optimal solution. In particular, the weight
parameter and the bias term can be iteratively updated using the
following equations:

where is a hyper-parameter known as the learning rate. The intuition behind
this equation is to move the weights in a direction that reduces the training
loss. If we arrive at a minima using this procedure, the gradient of the training
loss will be close to 0, eliminating the second term and resulting in the
convergence of weights. The weights are commonly initialized with values
drawn randomly from a Gaussian or a uniform distribution.

A necessary tool for updating weights in Equation 4.53 is to compute the
partial derivative of E with respect to . This computation is nontrivial
especially at hidden layers , since does not directly affect (and

yk y^k aL

Loss (yk, y^k)=(yk, y^k)2. (4.52)


wijl bil

wijl←wijl−λ∂E∂wijl, (4.53)

bil←bil−λ∂E∂bil, (4.54)


(l<L) wijl y^=aL

hence the training loss), but has complex chains of influences via activation
values at subsequent layers. To address this problem, a technique known as
backpropagation was developed, which propagates the derivatives
backward from the output layer to the hidden layers. This technique can be
described as follows.

Recall that the training loss E is simply the sum of individual losses at training
instances. Hence the partial derivative of E can be decomposed as a sum of
partial derivatives of individual losses.

To simplify discussions, we will consider only the derivatives of the loss at the
k training instance, which will be generically represented as . By
using the chain rule of differentiation, we can represent the partial derivatives
of the loss with respect to as

The last term of the previous equation can be written as

Also, if we use the sigmoid activation function, then

Equation 4.55 can thus be simplified as

∂E∂wjl=∑k=1n∂ Loss (yk, y^k)∂wjl.

th Loss(y, aL)


∂ Loss∂wijl=∂ Loss∂ail×∂ail∂zil×∂zil∂wijl. (4.55)


∂ail∂zil=∂ σ(zil)∂zil=ail(1−ai1).

∂ Loss∂wijl=δil×ail(1−ai1)×ajl−1,where δil=∂ Loss∂ail. (4.56)

A similar formula for the partial derivatives with respect to the bias terms is
given by

Hence, to compute the partial derivatives, we only need to determine .
Using a squared loss function, we can easily write at the output node as

However, the approach for computing at hidden nodes is more
involved. Notice that affects the activation values of all nodes at the
next layer, which in turn influences the loss. Hence, again using the chain rule
of differentiation, can be represented as

The previous equation provides a concise representation of the values at
layer l in terms of the values computed at layer . Hence, proceeding
backward from the output layer L to the hidden layers, we can recursively
apply Equation 4.59 to compute at every hidden node. can then be
used in Equations 4.56 and 4.57 to compute the partial derivatives of
the loss with respect to and , respectively. Algorithm 4.4 summarizes
the complete approach for learning the model parameters of ANN using
backpropagation and gradient descent method.

Algorithm 4.4 Learning ANN using
backpropagation and gradient descent.


∂ Loss∂bil=δil×ail(1−ai1). (4.57)


δL=∂ Loss∂aL=∂ (y−aL)2∂aL=2(aL−y). (4.58)

δjl (l<L)
ajl ail+1


δjl=∂ Loss∂ajl=∑i(∂ Loss∂ail+1×∂ail+1∂ajl).=∑i(∂ Loss∂ail+1×∂ail+1∂zil+1×∂zil+1(4.59)

δjl+1 l+1

δil δil

wijl bil

∂ ∂ ∂ ∂

∂ ∂ ∑ ∂ ∂

∂ ∂ ∑ ∂ ∂

4.7.3 Characteristics of ANN

1. Multi-layer neural networks with at least one hidden layer are universal
approximators; i.e., they can be used to approximate any target
function. They are thus highly expressive and can be used to learn
complex decision boundaries in diverse applications. ANN can also be
used for multiclass classification and regression problems, by

appropriately modifying the output layer. However, the high model
complexity of classical ANN models makes it susceptible to overfitting,
which can be overcome to some extent by using deep learning
techniques discussed in Section 4.8.3 .

2. ANN provides a natural way to represent a hierarchy of features at
multiple levels of abstraction. The outputs at the final hidden layer of
the ANN model thus represent features at the highest level of
abstraction that are most useful for classification. These features can
also be used as inputs in other supervised classification models, e.g.,
by replacing the output node of the ANN by any generic classifier.

3. ANN represents complex high-level features as compositions of simpler
lower-level features that are easier to learn. This provides ANN the
ability to gradually increase the complexity of representations, by
adding more hidden layers to the architecture. Further, since simpler
features can be combined in combinatorial ways, the number of
complex features learned by ANN is much larger than traditional
classification models. This is one of the main reasons behind the high
expressive power of deep neural networks.

4. ANN can easily handle irrelevant attributes, by using zero weights for
attributes that do not help in improving the training loss. Also,
redundant attributes receive similar weights and do not degrade the
quality of the classifier. However, if the number of irrelevant or
redundant attributes is large, the learning of the ANN model may suffer
from overfitting, leading to poor generalization performance.

5. Since the learning of ANN model involves minimizing a non-convex
function, the solutions obtained by gradient descent are not guaranteed
to be globally optimal. For this reason, ANN has a tendency to get
stuck in local minima, a challenge that can be addressed by using deep
learning techniques discussed in Section 4.8.4 .

6. Training an ANN is a time consuming process, especially when the
number of hidden nodes is large. Nevertheless, test examples can be

classified rapidly.
7. Just like logistic regression, ANN can learn in the presence of

interacting variables, since the model parameters are jointly learned
over all variables together. In addition, ANN cannot handle instances
with missing values in the training or testing phase.

4.8 Deep Learning
As described above, the use of hidden layers in ANN is based on the general
belief that complex high-level features can be constructed by combining
simpler lower-level features. Typically, the greater the number of hidden
layers, the deeper the hierarchy of features learned by the network. This
motivates the learning of ANN models with long chains of hidden layers,
known as deep neural networks. In contrast to “shallow” neural networks
that involve only a small number of hidden layers, deep neural networks are
able to represent features at multiple levels of abstraction and often require far
fewer nodes per layer to achieve generalization performance similar to
shallow networks.

Despite the huge potential in learning deep neural networks, it has remained
challenging to learn ANN models with a large number of hidden layers using
classical approaches. Apart from reasons related to limited computational
resources and hardware architectures, there have been a number of
algorithmic challenges in learning deep neural networks. First, learning a deep
neural network with low training error has been a daunting task because of the
saturation of sigmoid activation functions, resulting in slow convergence of
gradient descent. This problem becomes even more serious as we move
away from the output node to the hidden layers, because of the compounded
effects of saturation at multiple layers, known as the vanishing gradient
problem. Because of this reason, classical ANN models have suffered from
slow and ineffective learning, leading to poor training and test performance.
Second, the learning of deep neural networks is quite sensitive to the initial
values of model parameters, chiefly because of the non-convex nature of the
optimization function and the slow convergence of gradient descent. Third,
deep neural networks with a large number of hidden layers have high model

complexity, making them susceptible to overfitting. Hence, even if a deep
neural network has been trained to show low training error, it can still suffer
from poor generalization performance.

These challenges have deterred progress in building deep neural networks for
several decades and it is only recently that we have started to unlock their
immense potential with the help of a number of advances being made in the
area of deep learning. Although some of these advances have been around
for some time, they have only gained mainstream attention in the last decade,
with deep neural networks continually beating records in various competitions
and solving problems that were too difficult for other classification approaches.

There are two factors that have played a major role in the emergence of deep
learning techniques. First, the availability of larger labeled data sets, e.g., the
ImageNet data set contains more than 10 million labeled images, has made it
possible to learn more complex ANN models than ever before, without falling
easily into the traps of model overfitting. Second, advances in computational
abilities and hardware infrastructures, such as the use of graphical processing
units (GPU) for distributed computing, have greatly helped in experimenting
with deep neural networks with larger architectures that would not have been
feasible with traditional resources.

In addition to the previous two factors, there have been a number of
algorithmic advancements to overcome the challenges faced by classical
methods in learning deep neural networks. Some examples include the use of
more responsive combinations of loss functions and activation functions,
better initialization of model parameters, novel regularization techniques, more
agile architecture designs, and better techniques for model learning and
hyper-parameter selection. In the following, we describe some of the deep
learning advances made to address the challenges in learning deep neural

networks. Further details on recent developments in deep learning can be
obtained from the Bibliographic Notes.

4.8.1 Using Synergistic Loss Functions

One of the major realizations leading to deep learning has been the
importance of choosing appropriate combinations of activation and loss
functions. Classical ANN models commonly made use of the sigmoid
activation function at the output layer, because of its ability to produce real-
valued outputs between 0 and 1, which was combined with a squared loss
objective to perform gradient descent. It was soon noticed that this particular
combination of activation and loss function resulted in the saturation of output
activation values, which can be described as follows.

Saturation of Outputs
Although the sigmoid has been widely-used as an activation function, it easily
saturates at high and low values of inputs that are far away from 0. Observe
from Figure 4.27(a) that shows variance in its values only when z is
close to 0. For this reason, is non-zero for only a small range of z
around 0, as shown in Figure 4.27(b) . Since is one of the
components in the gradient of loss (see Equation 4.55 ), we get a
diminishing gradient value when the activation values are far from 0.



Figure 4.27.
Plots of sigmoid function and its derivative.

To illustrate the effect of saturation on the learning of model parameters at the
output node, consider the partial derivative of loss with respect to the weight

at the output node. Using the squared loss function, we can write this as

In the previous equation, notice that when is highly negative, (and
hence the gradient) is close to 0. On the other hand, when is highly
positive, becomes close to 0, nullifying the value of the gradient.
Hence, irrespective of whether the prediction matches the true label y or
not, the gradient of the loss with respect to the weights is close to 0 whenever
is highly positive or negative. This causes an unnecessarily slow

convergence of the model parameters of the ANN model, often resulting in
poor learning.

Note that it is the combination of the squared loss function and the sigmoid
activation function at the output node that together results in diminishing


∂ Loss∂wjL=2(aL−y)×σ(zL)(1−σ(zL))×ajL−1. (4.60)

zL σ(zL)



gradients (and thus poor learning) upon saturation of outputs. It is thus
important to choose a synergistic combination of loss function and activation
function that does not suffer from the saturation of outputs.

Cross entropy loss function
The cross entropy loss function, which was described in the context of logistic
regression in Section 4.6.2 , can significantly avoid the problem of
saturating outputs when used in combination with the sigmoid activation
function. The cross entropy loss function of a real-valued prediction
on a data instance with binary label can be defined as

where log represents the natural logarithm (to base e) and for
convenience. The cross entropy function has foundations in information theory
and measures the amount of disagreement between y and . The partial
derivative of this loss function with respect to can be given as

Using this value of in Equation 4.56 , we can obtain the partial derivative
of the loss with respect to the weight at the output node as

Notice the simplicity of the previous formula using the cross entropy loss
function. The partial derivatives of the loss with respect to the weights at the
output node depend only on the difference between the prediction and the
true label y. In contrast to Equation 4.60 , it does not involve terms such as

that can be impacted by saturation of . Hence, the gradients

y^∈(0, 1)
y∈{0, 1}

Loss(y, y^)=−ylog(y^)−(1−y)log(1−y^), (4.61)

0 log(0)=0


δL=∂ Loss∂aL=−yaL+(1−y)(1−aL).=(aL−y)aL(1−aL). (4.62)


∂ Loss∂wjL=(aL−y)aL(1−aL)×aL(1−aL)×ajL−1.=(aL−y)×ajL−1. (4.63)


σ(zL)(1−σ(zL)) zL

are high whenever is large, promoting effective learning of the model
parameters at the output node. This has been a major breakthrough in the
learning of modern ANN models and it is now a common practice to use the
cross entropy loss function with sigmoid activations at the output node.

4.8.2 Using Responsive Activation

Even though the cross entropy loss function helps in overcoming the problem
of saturating outputs, it still does not solve the problem of saturation at hidden
layers, arising due to the use of sigmoid activation functions at hidden nodes.
In fact, the effect of saturation on the learning of model parameters is even
more aggravated at hidden layers, a problem known as the vanishing gradient
problem. In the following, we describe the vanishing gradient problem and the
use of a more responsive activation function, called the rectified linear
output unit (ReLU), to overcome this problem.

Vanishing Gradient Problem
The impact of saturating activation values on the learning of model
parameters increases at deeper hidden layers that are farther away from the
output node. Even if the activation in the output layer does not saturate, the
repeated multiplications performed as we backpropagate the gradients from
the output layer to the hidden layers may lead to decreasing gradients in the
hidden layers. This is called the vanishing gradient problem, which has been
one of the major hindrances in learning deep neural networks.


To illustrate the vanishing gradient problem, consider an ANN model that
consists of a single node at every hidden layer of the network, as shown in
Figure 4.28 . This simplified architecture involves a single chain of hidden
nodes where a single weighted link connects the node at layer to the
node at layer l. Using Equations 4.56 and 4.59 , we can represent the
partial derivative of the loss with respect to as

Notice that if any of the linear predictors saturates at subsequent layers,
then the term becomes close to 0, thus diminishing the overall
gradient. The saturation of activations thus gets compounded and has
multiplicative effects on the gradients at hidden layers, making them highly
unstable and thus, unsuitable for use with gradient descent. Even though the
previous discussion only pertains to the simplified architecture involving a
single chain of hidden nodes, a similar argument can be made for any generic
ANN architecture involving multiple chains of hidden nodes. Note that the
vanishing gradient problem primarily arises because of the use of sigmoid
activation function at hidden nodes, which is known to easily saturate
especially after repeated multiplications.

Figure 4.28.
An example of an ANN model with only one node at every hidden layer.

wl l−1


∂ Loss∂wl=δl×al(1−al)×al−1,where δl=2(aL−y)×∏r=lL−1(ar+1(1−ar+1)×wr+1).(4.64)


Figure 4.29.
Plot of the rectified linear unit (ReLU) activation function.

Rectified Linear Units (ReLU)
To overcome the vanishing gradient problem, it is important to use an
activation function f(z)at the hidden nodes that provides a stable and
significant value of the gradient whenever a hidden node is active, i.e., .
This is achieved by using rectified linear units (ReLU) as activation functions
at hidden nodes, which can be defined as

The idea of ReLU has been inspired from biological neurons, which are either
in an inactive state or show an activation value proportional to the
input. Figure 4.29 shows a plot of the ReLU function. We can see that it is
linear with respect to z when . Hence, the gradient of the activation value
with respect to z can be written as


a=f(z)={z,if z>0.0,otherwise. (4.65)



Although f(z)is not differentiable at 0, it is common practice to use
when . Since the gradient of the ReLU activation function is equal to 1
whenever , it avoids the problem of saturation at hidden nodes, even after
repeated multiplications. Using ReLU, the partial derivatives of the loss with
respect to the weight and bias parameters can be given by

Notice that ReLU shows a linear behavior in the activation values whenever a
node is active, as compared to the nonlinear properties of the sigmoid
function. This linearity promotes better flows of gradients during
backpropagation, and thus simplifies the learning of ANN model parameters.
The ReLU is also highly responsive at large values of z away from 0, as
opposed to the sigmoid activation function, making it more suitable for
gradient descent. These differences give ReLU a major advantage over the
sigmoid function. Indeed, ReLU is used as the preferred choice of activation
function at hidden layers in most modern ANN models.

4.8.3 Regularization

A major challenge in learning deep neural networks is the high model
complexity of ANN models, which grows with the addition of hidden layers in
the network. This can become a serious concern, especially when the training
set is small, due to the phenomena of model overfitting. To overcome this

∂a∂z={1,if z>0.0,if z<0. (4.66)



∂ Loss∂wijl=δil×I(zil)×ajl−1, (4.67)

∂ Loss∂bil=δil×I(zil),where δil=∑i=1n(δil+1×I(zil+1)×wijl+1),and I(z)=
{1,if z>0.0,otherwise.


challenge, it is important to use techniques that can help in reducing the
complexity of the learned model, known as regularization techniques.
Classical approaches for learning ANN models did not have an effective way
to promote regularization of the learned model parameters. Hence, they had
often been sidelined by other classification methods, such as support vector
machines (SVM), which have in-built regularization mechanisms. (SVMs will
be discussed in more detail in Section 4.9 ).

One of the major advancements in deep learning has been the development
of novel regularization techniques for ANN models that are able to offer
significant improvements in generalization performance. In the following, we
discuss one of the regularization techniques for ANN, known as the dropout
method, that have gained a lot of attention in several applications.

The main objective of dropout is to avoid the learning of spurious features at
hidden nodes, occurring due to model overfitting. It uses the basic intuition
that spurious features often “co-adapt” themselves such that they show good
training performance only when used in highly selective combinations. On the
other hand, relevant features can be used in a diversity of feature
combinations and hence are quite resilient to the removal or modification of
other features. The dropout method uses this intuition to break complex “co-
adaptations” in the learned features by randomly dropping input and hidden
nodes in the network during training.

Dropout belongs to a family of regularization techniques that uses the criteria
of resilience to random perturbations as a measure of the robustness (and
hence, simplicity) of a model. For example, one approach to regularization is
to inject noise in the input attributes of the training set and learn a model with
the noisy training instances. If a feature learned from the training data is

indeed generalizable, it should not be affected by the addition of noise.
Dropout can be viewed as a similar regularization approach that perturbs the
information content of the training set not only at the level of attributes but also
at multiple levels of abstractions, by dropping input and hidden nodes.

The dropout method draws inspiration from the biological process of gene
swapping in sexual reproduction, where half of the genes from both parents
are combined together to create the genes of the offspring. This favors the
selection of parent genes that are not only useful but can also inter-mingle
with diverse combinations of genes coming from the other parent. On the
other hand, co-adapted genes that function only in highly selective
combinations are soon eliminated in the process of evolution. This idea is
used in the dropout method for eliminating spurious co-adapted features. A
simplified description of the dropout method is provided in the rest of this

Figure 4.30.
Examples of sub-networks generated in the dropout method using .

Let represent the model parameters of the ANN model at the k
iteration of the gradient descent method. At every iteration, we randomly
select a fraction of input and hidden nodes to be dropped from the network,
where is a hyper-parameter that is typically chosen to be 0.5. The
weighted links and bias terms involving the dropped nodes are then
eliminated, resulting in a “thinned” sub-network of smaller size. The model
parameters of the sub-network are then updated by computing
activation values and performing backpropagation on this smaller sub-
network. These updated values are then added back in the original network to


(wk, bk) th

γ∈(0, 1)

(wsk, bsk)

obtain the updated model parameters, , to be used in the next

Figure 4.30 shows some examples of sub-networks that can be generated
at different iterations of the dropout method, by randomly dropping input and
hidden nodes. Since every sub-network has a different architecture, it is
difficult to learn complex co-adaptations in the features that can result in
overfitting. Instead, the features at the hidden nodes are learned to be more
agile to random modifications in the network structure, thus improving their
generalization ability. The model parameters are updated using a different
random sub-network at every iteration, till the gradient descent method

Let denote the model parameters at the last iteration
of the gradient descent method. These parameters are finally scaled down by
a factor of , to produce the weights and bias terms of the final ANN
model, as follows:

We can now use the complete neural network with model parameters
for testing. The dropout method has been shown to provide significant
improvements in the generalization performance of ANN models in a number
of applications. It is computationally cheap and can be applied in combination
with any of the other deep learning techniques. It also has a number of
similarities with a widely-used ensemble learning method known as bagging,
which learns multiple models using random subsets of the training set, and
then uses the average output of all the models to make predictions. (Bagging
will be presented in more detail later in Section 4.10.4 ). In a similar vein, it
can be shown that the predictions of the final network learned using dropout
approximates the average output of all possible sub-networks that can be

(wk+1, bk+1)

(wkmax, bkmax) kmax


(w*, b*)=((1−γ)×wkmax, (1−γ)×bkmax)

(w*, b*)


formed using n nodes. This is one of the reasons behind the superior
regularization abilities of dropout.

4.8.4 Initialization of Model Parameters

Because of the non-convex nature of the loss function used by ANN models, it
is possible to get stuck in locally optimal but globally inferior solutions. Hence,
the initial choice of model parameter values plays a significant role in the
learning of ANN by gradient descent. The impact of poor initialization is even
more aggravated when the model is complex, the network architecture is
deep, or the classification task is difficult. In such cases, it is often advisable to
first learn a simpler model for the problem, e.g., using a single hidden layer,
and then incrementally increase the complexity of the model, e.g., by adding
more hidden layers. An alternate approach is to train the model for a simpler
task and then use the learned model parameters as initial parameter choices
in the learning of the original task. The process of initializing ANN model
parameters before the actual training process is known as pretraining.

Pretraining helps in initializing the model to a suitable region in the parameter
space that would otherwise be inaccessible by random initialization.
Pretraining also reduces the variance in the model parameters by fixing the
starting point of gradient descent, thus reducing the chances of overfitting due
to multiple comparisons. The models learned by pretraining are thus more
consistent and provide better generalization performance.

Supervised Pretraining
A common approach for pretraining is to incrementally train the ANN model in
a layer-wise manner, by adding one hidden layer at a time. This approach,

known as supervised pretraining, ensures that the parameters learned at
every layer are obtained by solving a simpler problem, rather than learning all
model parameters together. These parameter values thus provide a good
choice for initializing the ANN model. The approach for supervised pretraining
can be briefly described as follows.

We start the supervised pretraining process by considering a reduced ANN
model with only a single hidden layer. By applying gradient descent on this
simple model, we are able to learn the model parameters of the first hidden
layer. At the next run, we add another hidden layer to the model and apply
gradient descent to learn the parameters of the newly added hidden layer,
while keeping the parameters of the first layer fixed. This procedure is
recursively applied such that while learning the parameters of the l hidden
layer, we consider a reduced model with only l hidden layers, whose first
hidden layers are not updated on the l run but are instead fixed using
pretrained values from previous runs. In this way, we are able to learn the
model parameters of all hidden layers. These pretrained values are
used to initialize the hidden layers of the final ANN model, which is fine-tuned
by applying a final round of gradient descent over all the layers.

Unsupervised Pretraining
Supervised pretraining provides a powerful way to initialize model parameters,
by gradually growing the model complexity from shallower to deeper
networks. However, supervised pretraining requires a sufficient number of
labeled training instances for effective initialization of the ANN model. An
alternate pretraining approach is unsupervised pretraining, which initializes
model parameters by using unlabeled instances that are often abundantly
available. The basic idea of unsupervised pretraining is to initialize the ANN




model in such a way that the learned features capture the latent structure in
the unlabeled data.

Figure 4.31.
The basic architecture of a single-layer autoencoder.

Unsupervised pretraining relies on the assumption that learning the
distribution of the input data can indirectly help in learning the classification
model. It is most helpful when the number of labeled examples is small and
the features for the supervised problem bear resemblance to the factors
generating the input data. Unsupervised pretraining can be viewed as a
different form of regularization, where the focus is not explicitly toward finding
simpler features but instead toward finding features that can best explain the
input data. Historically, unsupervised pretraining has played an important role
in reviving the area of deep learning, by making it possible to train any generic
deep neural network without requiring specialized architectures.

Use of Autoencoders

One simple and commonly used approach for unsupervised pretraining is to
use an unsupervised ANN model known as an autoencoder. The basic
architecture of an autoencoder is shown in Figure 4.31 . An autoencoder
attempts to learn a reconstruction of the input data by mapping the attributes
to latent features , and then re-projecting back to the original attribute

space to create the reconstruction . The latent features are represented
using a hidden layer of nodes, while the input and output layers represent the
attributes and contain the same number of nodes. During training, the goal is
to learn an autoencoder model that provides the lowest reconstruction error,

, on all input data instances. A typical choice of the reconstruction
error is the squared loss function:

The model parameters of the autoencoder can be learned by using a similar
gradient descent method as the one used for learning supervised ANN
models for classification. The key difference is the use of the reconstruction
error on all training instances as the training loss. Autoencoders that have
multiple layers of hidden layers are known as stacked autoencoders.

Autoencoders are able to capture complex representations of the input data
by the use of hidden nodes. However, if the number of hidden nodes is large,
it is possible for an autoencoder to learn the identity relationship, where the
input is just copied and returned as the output , resulting in a trivial
solution. For example, if we use as many hidden nodes as the number of
attributes, then it is possible for every hidden node to copy an attribute and
simply pass it along to an output node, without extracting any useful
information. To avoid this problem, it is common practice to keep the number
of hidden nodes smaller than the number of input attributes. This forces the
autoencoder to learn a compact and useful encoding of the input data, similar
to a dimensionality reduction technique. An alternate approach is to corrupt


RE(x, x^)

RE(x, x^)=ǁx−x^ ǁ2.


the input instances by adding random noise, and then learn the autoencoder
to reconstruct the original instance from the noisy input. This approach is
known as the denoising autoencoder, which offers strong regularization
capabilities and is often used to learn complex features even in the presence
of a large number of hidden nodes.

To use an autoencoder for unsupervised pretraining, we can follow a similar
layer-wise approach like supervised pretraining. In particular, to pretrain the
model parameters of the l hidden layer, we can construct a reduced ANN
model with only l hidden layers and an output layer containing the same
number of nodes as the attributes and is used for reconstruction. The
parameters of the l hidden layer of this network are then learned using a
gradient descent method to minimize the reconstruction error. The use of
unlabeled data can be viewed as providing hints to the learning of parameters
at every layer that aid in generalization. The final model parameters of the
ANN model are then learned by applying gradient descent over all the layers,
using the initial values of parameters obtained from pretraining.

Hybrid Pretraining
Unsupervised pretraining can also be combined with supervised pretraining by
using two output layers at every run of pretraining, one for reconstruction and
the other for supervised classification. The parameters of the l hidden layer
are then learned by jointly minimizing the losses on both output layers, usually
weighted by a trade-off hyper-parameter . Such a combined approach often
shows better generalization performance than either of the approaches, since
it provides a way to balance between the competing objectives of representing
the input data and improving classification performance.





4.8.5 Characteristics of Deep Learning

Apart from the basic characteristics of ANN discussed in Section 4.7.3 , the
use of deep learning techniques provides the following additional

1. An ANN model trained for some task can be easily re-used for a
different task that involves the same attributes, by using pretraining
strategies. For example, we can use the learned parameters of the
original task as initial parameter choices for the target task. In this way,
ANN promotes re-usability of learning, which can be quite useful when
the target application has a smaller number of labeled training

2. Deep learning techniques for regularization, such as the dropout
method, help in reducing the model complexity of ANN and thus
promoting good generalization performance. The use of regularization
techniques is especially useful in high-dimensional settings, where the
number of training labels is small but the classification problem is
inherently difficult.

3. The use of an autoencoder for pretraining can help eliminate irrelevant
attributes that are not related to other attributes. Further, it can help
reduce the impact of redundant attributes by representing them as
copies of the same attribute.

4. Although the learning of an ANN model can succumb to finding inferior
and locally optimal solutions, there are a number of deep learning
techniques that have been proposed to ensure adequate learning of an
ANN. Apart from the methods discussed in this section, some other
techniques involve novel architecture designs such as skip connections
between the output layer and lower layers, which aids the easy flow of
gradients during backpropagation.

5. A number of specialized ANN architectures have been designed to
handle a variety of input data sets. Some examples include
convolutional neural networks (CNN) for two-dimensional gridded
objects such as images, and recurrent neural network (RNN) for
sequences. While CNNs have been extensively used in the area of
computer vision, RNNs have found applications in processing speech
and language.

4.9 Support Vector Machine (SVM)
A support vector machine (SVM) is a discriminative classification model that
learns linear or nonlinear decision boundaries in the attribute space to
separate the classes. Apart from maximizing the separability of the two
classes, SVM offers strong regularization capabilities, i.e., it is able to control
the complexity of the model in order to ensure good generalization
performance. Due to its unique ability to innately regularize its learning, SVM
is able to learn highly expressive models without suffering from overfitting. It
has thus received considerable attention in the machine learning community
and is commonly used in several practical applications, ranging from
handwritten digit recognition to text categorization. SVM has strong roots in
statistical learning theory and is based on the principle of structural risk
minimization. Another unique aspect of SVM is that it represents the decision
boundary using only a subset of the training examples that are most difficult to
classify, known as the support vectors. Hence, it is a discriminative model
that is impacted only by training instances near the boundary of the two
classes, in contrast to learning the generative distribution of every class.

To illustrate the basic idea behind SVM, we first introduce the concept of the
margin of a separating hyperplane and the rationale for choosing such a
hyperplane with maximum margin. We then describe how a linear SVM can
be trained to explicitly look for this type of hyperplane. We conclude by
showing how the SVM methodology can be extended to learn nonlinear
decision boundaries by using kernel functions.

4.9.1 Margin of a Separating


The generic equation of a separating hyperplane can be written as

where represents the attributes and ( , ) represent the parameters of the
hyperplane. A data instance can belong to either side of the hyperplane
depending on the sign of . For the purpose of binary classification, we
are interested in finding a hyperplane that places instances of both classes on
opposite sides of the hyperplane, thus resulting in a separation of the two
classes. If there exists a hyperplane that can perfectly separate the classes in
the data set, we say that the data set is linearly separable. Figure 4.32
shows an example of linearly separable data involving two classes, squares
and circles. Note that there can be infinitely many hyperplanes that can
separate the classes, two of which are shown in Figure 4.32 as lines
and . Even though every such hyperplane will have zero training error, they
can provide different results on previously unseen instances. Which
separating hyperplane should we thus finally choose to obtain the best
generalization performance? Ideally, we would like to choose a simple
hyperplane that is robust to small perturbations. This can be achieved by
using the concept of the margin of a separating hyperplane, which can be
briefly described as follows.




Figure 4.32.
Margin of a hyperplane in a two-dimensional data set.

For every separating hyperplane , let us associate a pair of parallel
hyperplanes, and , such that they touch the closest instances of both
classes, respectively. For example, if we move parallel to its direction, we
can touch the first square using and the first circle using . and
are known as the margin hyperplanes of and the distance between them
is known as the margin of the separating hyperplane . From the diagram
shown in Figure 4.32 , notice that the margin for is considerably larger
than that for . In this example, turns out to be the separating hyperplane
with the maximum margin, known as the maximum margin hyperplane.

Rationale for Maximum Margin

bi1 bi2

b11 b12 bi1 bi2


B2 b1

Hyperplanes with large margins tend to have better generalization
performance than those with small margins. Intuitively, if the margin is small,
then any slight perturbation in the hyperplane or the training instances located
at the boundary can have quite an impact on the classification performance.
Small margin hyperplanes are thus more susceptible to overfitting, as they are
barely able to separate the classes with a very narrow room to allow
perturbations. On the other hand, a hyperplane that is farther away from
training instances of both classes has sufficient leeway to be robust to minor
modifications in the data, and thus shows superior generalization

The idea of choosing the maximum margin separating hyperplane also has
strong foundations in statistical learning theory. It can be shown that the
margin of such a hyperplane is inversely related to the VC-dimension of the
classifier, which is a commonly used measure of the complexity of a model.
As discussed in Section 3.4 of the last chapter, a simpler model should be
preferred over a more complex model if they both show similar training
performance. Hence, maximizing the margin results in the selection of a
separating hyperplane with the lowest model complexity, which is expected to
show better generalization performance.

4.9.2 Linear SVM

A linear SVM is a classifier that searches for a separating hyperplane with the
largest margin, which is why it is often known as a maximal margin
classifier. The basic idea of SVM can be described as follows.

Consider a binary classification problem consisting of n training instances,
where every training instance is associated with a binary label .xi yi∈{−1, 1}

Let be the equation of a separating hyperplane that separates the
two classes by placing them on opposite sides. This means that

The distance of any point from the hyperplane is then given by

where denotes the absolute value and denotes the length of a vector.
Let the distance of the closest point from the hyperplane with be .
Similarly, let denote the distance of the closest point from class .

This can be represented using the following constraints:

The previous equations can be succinctly represented by using the product of
and as

where M is a parameter related to the margin of the hyperplane, i.e., if
, then margin . In order to find the maximum margin

hyperplane that adheres to the previous constraints, we can consider the
following optimization problem:

To find the solution to the previous problem, note that if and b satisfy the
constraints of the previous problem, then any scaled version of and b would


wTxi+b>0if yi=1,wTxi+b<0if yi=−1.

D(x)=|wTx+b |ǁ w ǁ

|⋅| ǁ ⋅ ǁ
y=1 k+>0

k−>0 −1

wTxi+bǁ w ǁ≥k+if yi=1,wTxi+bǁ w ǁ≤−k−if yi=−1, (4.69)

yi (wTxi+b)

yi(wTxi+b)≥Mǁwǁ (4.70)

−=M =k+−k−=2M

maxw, bMsubject toyi(wTxi+b)≥Mǁ w ǁ. (4.71)

satisfy them too. Hence, we can conveniently choose to simplify the
right-hand side of the inequalities. Furthermore, maximizing M amounts to
minimizing . Hence, the optimization problem of SVM is commonly
represented in the following form:

Learning Model Parameters
Equation 4.72 represents a constrained optimization problem with linear
inequalities. Since the objective function is convex and quadratic with respect
to , it is known as a quadratic programming problem (QPP), which can be
solved using standard optimization techniques, as described in Appendix E. In
the following, we present a brief sketch of the main ideas for learning the
model parameters of SVM.

First, we rewrite the objective function in a form that takes into account the
constraints imposed on its solutions. The new objective function is known as
the Lagrangian primal problem, which can be represented as follows,

where the parameters correspond to the constraints and are called the
Lagrange multipliers. Next, to minimize the Lagrangian, we take the
derivative of with respect to and b and set them equal to zero:



minw, bǁ w ǁ22subject toyi(wTxi+b)≥1. (4.72)

LP=12ǁ w ǁ2−∑i=1nλi(yi(wTxi+b)−1), (4.73)



∂LP∂w=0⇒w=∑i=1nλiyixi, (4.74)

∂LP∂b=0⇒∑i=1nλiyi=0. (4.75)

Note that using Equation 4.74 , we can represent completely in terms of
the Lagrange multipliers. There is another relationship between ( , b) and
that is derived from the Karush-Kuhn-Tucker (KKT) conditions, a commonly
used technique for solving QPP. This relationship can be described as

Equation 4.76 is known as the complementary slackness condition,
which sheds light on a valuable property of SVM. It states that the Lagrange
multiplier is strictly greater than 0 only when satisfies the equation

, which means that lies exactly on a margin hyperplane.
However, if is farther away from the margin hyperplanes such that

, then is necessarily 0. Hence, for only a small number of
instances that are closest to the separating hyperplane, which are known as
support vectors. Figure 4.33 shows the support vectors of a hyperplane
as filled circles and squares. Further, if we look at Equation 4.74 , we will
observe that training instances with do not contribute to the weight
parameter . This suggests that can be concisely represented only in terms
of the support vectors in the training data, which are quite fewer than the
overall number of training instances. This ability to represent the decision
function only in terms of the support vectors is what gives this classifier the
name support vector machines.


λi[yi(wTxi+b)−1]=0. (4.76)

λi xi
yi(w⋅xi+b)=1 xi

yi(w⋅xi+b)>1 λi λi>0


Figure 4.33.
Support vectors of a hyperplane shown as filled circles and squares.

Using equations 4.74 , 4.75 , and 4.76 in Equation 4.73 , we obtain
the following optimization problem in terms of the Lagrange multipliers :

The previous optimization problem is called the dual optimization problem.
Maximizing the dual problem with respect to is equivalent to minimizing the
primal problem with respect to and b.

The key differences between the dual and primal problems are as follows:


maxλi∑i=1nλi−12∑i=1n∑j=1nλiλjyiyjxiTxjsubject to∑i=1nλiyi=0,λi≥0. (4.77)


1. Solving the dual problem helps us identify the support vectors in the
data that have non-zero values of . Further, the solution of the dual
problem is influenced only by the support vectors that are closest to the
decision boundary of SVM. This helps in summarizing the learning of
SVM solely in terms of its support vectors, which are easier to manage
computationally. Further, it represents a unique ability of SVM to be
dependent only on the instances closest to the boundary, which are
harder to classify, rather than the distribution of instances farther away
from the boundary.

2. The objective of the dual problem involves only terms of the form ,
which are basically inner products in the attribute space. As we will see
later in Section 4.9.4 , this property will prove to be quite useful in
learning nonlinear decision boundaries using SVM.

Because of these differences, it is useful to solve the dual optimization
problem using any of the standard solvers for QPP. Having found an optimal
solution for , we can use Equation 4.74 to solve for . We can then use
Equation 4.76 on the support vectors to solve for b as follows:

where S represents the set of support vectors and is the
number of support vectors. The maximum margin hyperplane can then be
expressed as

Using this separating hyperplane, a test instance can be assigned a class
label using the sign of f( ).




b=1nS∑i∈S1−yiwTxiyi (4.78)

(S={i|λi>0}) nS

f(x)=(∑i=1nλiyixiTx)+b=0. (4.79)

Example 4.7.
Consider the two-dimensional data set shown in Figure 4.34 , which
contains eight training instances. Using quadratic programming, we can
solve the optimization problem stated in Equation 4.77 to obtain the
Lagrange multiplier for each training instance. The Lagrange multipliers
are depicted in the last column of the table. Notice that only the first two
instances have non-zero Lagrange multipliers. These instances
correspond to the support vectors for this data set.

Let and b denote the parameters of the decision boundary.
Using Equation 4.74 , we can solve for and in the following way:


w=(w1, w2)
w1 w2


Figure 4.34.
Example of a linearly separable data set.

The bias term b can be computed using Equation 4.76 for each support

Averaging these values, we obtain . The decision boundary
corresponding to these parameters is shown in Figure 4.34 .

4.9.3 Soft-margin SVM

Figure 4.35 shows a data set that is similar to Figure 4.32 , except it has
two new examples, P and Q. Although the decision boundary misclassifies
the new examples, while classifies them correctly, this does not mean that

is a better decision boundary than because the new examples may
correspond to noise in the training data. should still be preferred over
because it has a wider margin, and thus, is less susceptible to overfitting.
However, the SVM formulation presented in the previous section only
constructs decision boundaries that are mistake-free.




B2 B1
B1 B2

Figure 4.35.
Decision boundary of SVM for the non-separable case.

This section examines how the formulation of SVM can be modified to learn a
separating hyperplane that is tolerable to small number of training errors using
a method known as the soft-margin approach. More importantly, the method
presented in this section allows SVM to learn linear hyperplanes even in
situations where the classes are not linearly separable. To do this, the learning
algorithm in SVM must consider the trade-off between the width of the margin
and the number of training errors committed by the linear hyperplane.

To introduce the concept of training errors in the SVM formulation, let us relax
the inequality constraints to accommodate for some violations on a small
number of training instances. This can be done by introducing a slack
variable for every training instance as follows:ξ≥0 xi

The variable allows for some slack in the inequalities of the SVM such that
every instance does not need to strictly satisfy . Further, is
non-zero only if the margin hyperplanes are not able to place on the same
side as the rest of the instances belonging to . To illustrate this, Figure
4.36 shows a circle P that falls on the opposite side of the separating
hyperplane as the rest of the circles, and thus satisfies . The
distance between P and the margin hyperplane is equal to .
Hence, provides a measure of the error of SVM in representing using soft
inequality constraints.

Figure 4.36.
Slack variables used in soft-margin SVM.

yi(wTxi+b)≥1−ξi (4.80)

xi yi(wTxi+b)≥1 ξi


wTx+b=−1 ξ/ǁ w ǁ

ξi xi

In the presence of slack variables, it is important to learn a separating
hyperplane that jointly maximizes the margin (ensuring good generalization
performance) and minimizes the values of slack variables (ensuring low
training error). This can be achieved by modifying the optimization problem of
SVM as follows:

where C is a hyper-parameter that makes a trade-off between maximizing the
margin and minimizing the training error. A large value of C pays more
emphasis on minimizing the training error than maximizing the margin. Notice
the similarity of the previous equation with the generic formula of
generalization error rate introduced in Section 3.4 of the previous chapter.
Indeed, SVM provides a natural way to balance between model complexity
and training error in order to maximize generalization performance.

To solve Equation 4.81 we apply the Lagrange multiplier method and
convert the primal problem to its corresponding dual problem, similar to the
approach described in the previous section. The Lagrangian primal problem of
Equation 4.81 can be written as follows:

where and are the Lagrange multipliers corresponding to the
inequality constraints of Equation 4.81 . Setting the derivative of with
respect to , b, and equal to 0, we obtain the following equations:

minw, b, ξiǁ w ǁ22+c∑i=1nξisubject toyi(wTxi+b)≥1−ξi,ξi≥0. (4.81)

LP=12ǁ w ǁ2+C∑i=1nξi−∑i=1nλi(yi(wTxi+b)−1+ξi)−∑i=1nμi(ξi), (4.82)

λi≥0 μi≥0


∂LP∂w=0⇒w=∑i=1nλiyixi. (4.83)

∂L∂b=0⇒∑i=1nλiyi=0. (4.84)

We can also obtain the complementary slackness conditions by using the
following KKT conditions:

Equation 4.86 suggests that is zero for all training instances except those
that reside on the margin hyperplanes , or have . These
instances with are known as support vectors. On the other hand, given
in Equation 4.87 is zero for any training instance that is misclassified, i.e.,

. Further, and are related with each other by Equation 4.85 . This
results in the following three configurations of :

1. If and , then does not reside on the margin hyperplanes
and is correctly classified on the same side as other instances
belonging to .

2. If and , then is misclassified and has a non-zero slack
variable .

3. If and , then resides on one of the margin

Substituting Equations 4.83 to 4.87 into Equation 4.82 , we obtain the
following dual optimization problem:

Notice that the previous problem looks almost identical to the dual problem of
SVM for the linearly separable case (Equation 4.77 ), except that is

∂L∂ξi=0⇒λi+μi=C. (4.85)

λi(yi(wTxi+b)−1+ξi)=0, (4.86)

μiξi=0. (4.87)

wTxi+b=±1 ξi>0

λi>0 μi

ξi>0 λi μi
(λi, μi)

λi=0 μi=C xi

λi=C μi=0 xi

0<λi<C 0<μi<C xi

maxλi∑i=1nλi−12∑i=1n∑j=1nλiλjyiyjxiTxjsubject to∑i=1nλiyi=0,0≤λi≤C. (4.88)


required to not only be greater than 0 but also smaller than a constant value
C. Clearly, when C reaches infinity, the previous optimization problem
becomes equivalent to Equation 4.77 , where the learned hyperplane
perfectly separates the classes (with no training errors). However, by capping
the values of to C, the learned hyperplane is able to tolerate a few training
errors that have .

Figure 4.37.
Hinge loss as a function of .

As before, Equation 4.88 can be solved by using any of the standard
solvers for QPP, and the optimal value of can be obtained by using
Equation 4.83 . To solve for b, we can use Equation 4.86 on the support
vectors that reside on the margin hyperplanes as follows:



b=1nS∑i∈S1−yiwTxiyi (4.89)

where S represents the set of support vectors residing on the margin
hyperplanes and is the number of elements in S.

SVM as a Regularizer of Hinge Loss
SVM belongs to a broad class of regularization techniques that use a loss
function to represent the training errors and a norm of the model parameters
to represent the model complexity. To realize this, notice that the slack
variable , used for measuring the training errors in SVM, is equivalent to the
hinge loss function, which can be defined as follows:

where . In the case of SVM, corresponds to . Figure
4.37 shows a plot of the hinge loss function as we vary . We can see
that the hinge loss is equal to 0 as long as y and have the same sign and

. However, the hinge loss grows linearly with whenever y and are
of the opposite sign or . This is similar to the notion of the slack variable
, which is used to measure the distance of a point from its margin
hyperplane. Hence, the optimization problem of SVM can be represented in
the following equivalent form:

Note that using the hinge loss ensures that the optimization problem is convex
and can be solved using standard optimization techniques. However, if we use
a different loss function, such as the squared loss function that was introduced
in Section 4.7 on ANN, it will result in a different optimization problem that
may or may not remain convex. Nevertheless, different loss functions can be
explored to capture varying notions of training error, depending on the
characteristics of the problem.

(S={i|0<λi<C}) nS


Loss (y, y^) =max(0, 1−yy^),

y∈{+1, −1} y^ wTx+b

|y^|≥1 |y^| y^


minw, bǁ w ǁ22+C∑i=1nLoss (yi, wTxi+b) (4.90)

Another interesting property of SVM that relates it to a broader class of
regularization techniques is the concept of a margin. Although minimizing
has the geometric interpretation of maximizing the margin of a separating

hyperplane, it is essentially the squared norm of the model parameters,
. In general, the norm of , , is equal to the Minkowski distance of

order q from to the origin, i.e.,

Minimizing the norm of to achieve lower model complexity is a generic
regularization concept that has several interpretations. For example,
minimizing the norm amounts to finding a solution on a hypersphere of
smallest radius that shows suitable training performance. To visualize this in
two-dimensions, Figure 4.38(a) shows the plot of a circle with constant
radius r, where every point has the same norm. On the other hand, using
the norm ensures that the solution lies on the surface of a hypercube with
smallest size, with vertices along the axes. This is illustrated in Figure
4.38(b) as a square with vertices on the axes at a distance of r from the
origin. The norm is commonly used as a regularizer to obtain sparse model
parameters with only a small number of non-zero parameter values, such as
the use of Lasso in regression problems (see Bibliographic Notes).

ǁ w

L2 ǁ w
ǁ22 Lq ǁ w ǁq

ǁ w ǁq=(∑ipwiq)1/q





Figure 4.38.
Plots showing the behavior of two-dimensional solutions with constant and

In general, depending on the characteristics of the problem, different
combinations of norms and training loss functions can be used for learning
the model parameters, each requiring a different optimization solver. This
forms the backbone of a wide range of modeling techniques that attempt to
improve the generalization performance by jointly minimizing training error
and model complexity. However, in this section, we focus only on the squared
norm and the hinge loss function, resulting in the classical formulation of


4.9.4 Nonlinear SVM




The SVM formulations described in the previous sections construct a linear
decision boundary to separate the training examples into their respective
classes. This section presents a methodology for applying SVM to data sets
that have nonlinear decision boundaries. The basic idea is to transform the
data from its original attribute space in into a new space so that a
linear hyperplane can be used to separate the instances in the transformed
space, using the SVM approach. The learned hyperplane can then be
projected back to the original attribute space, resulting in a nonlinear decision

Figure 4.39.
Classifying data with a nonlinear decision boundary.

Attribute Transformation
To illustrate how attribute transformation can lead to a linear decision
boundary, Figure 4.39(a) shows an example of a two-dimensional data set
consisting of squares (classified as ) and circles (classified as ). The


y=1 y=−1

data set is generated in such a way that all the circles are clustered near the
center of the diagram and all the squares are distributed farther away from the
center. Instances of the data set can be classified using the following

The decision boundary for the data can therefore be written as follows:

which can be further simplified into the following quadratic equation:

A nonlinear transformation is needed to map the data from its original
attribute space into a new space such that a linear hyperplane can separate
the classes. This can be achieved by using the following simple

Figure 4.39(b) shows the points in the transformed space, where we can
see that all the circles are located in the lower left-hand side of the diagram. A
linear hyperplane with parameters and b can therefore be constructed in the
transformed space, to separate the instances into their respective classes.

One may think that because the nonlinear transformation possibly increases
the dimensionality of the input space, this approach can suffer from the curse
of dimensionality that is often associated with high-dimensional data.

y={1if (x1−0.5)2+(x2−0.5)2>0.2,−1otherwise. (4.91)




φ:(x1, x2)→(x12−x1, x22−x2). (4.92)

However, as we will see in the following section, nonlinear SVM is able to
avoid this problem by using kernel functions.

Learning a Nonlinear SVM Model
Using a suitable function, , we can transform any data instance to .
(The details on how to choose will become clear later.) The linear
hyperplane in the transformed space can be expressed as . To
learn the optimal separating hyperplane, we can substitute for in the
formulation of SVM to obtain the following optimization problem:

Using Lagrange multipliers , this can be converted into a dual optimization
problem: max

where denotes the inner product between vectors and . Also, the
equation of the hyperplane in the transformed space can be represented

as follows:

Further, b is given by

φ(⋅) φ(x)


minw, b, ξiǁ w ǁ22+C∑i=1nξisubject toyi(wTφ(xi)+b)≥1−ξi,ξi≥0. (4.93)


maxλi∑i=1nλi−12∑i=1n∑j=1nλiλjyiyj⟨φ(xi), φ(xj)
⟩subject to∑i=1nλjyi=0,0≤λi≤C,


⟨ a, b ⟩


∑i=1nλiyi⟨φ(xi), φ(x) ⟩+b=0. (4.95)

b=1nS(∑i∈S1yi−∑i∈S∑j=1nλjyiyj⟨φ(xi), φ(xj) ⟩yi), (4.96)

where is the set of support vectors residing on the margin
hyperplanes and is the number of elements in S.

Note that in order to solve the dual optimization problem in Equation 4.94 ,
or to use the learned model parameters to make predictions using Equations
4.95 and 4.96 , we need only inner products of . Hence, even though

may be nonlinear and high-dimensional, it suffices to use a function of
the inner products of in the transformed space. This can be achieved by
using a kernel trick, which can be described as follows.

The inner product between two vectors is often regarded as a measure of
similarity between the vectors. For example, the cosine similarity described in
Section 2.4.5 on page 79 can be defined as the dot product between two
vectors that are normalized to unit length. Analogously, the inner product

can also be regarded as a measure of similarity between two
instances, and , in the transformed space. The kernel trick is a method
for computing this similarity as a function of the original attributes. Specifically,
the kernel function K(u, v) between two instances u and v can be defined as

where is a function that follows certain conditions as stated by the Mercer’s
Theorem. Although the details of this theorem are outside the scope of the
book, we provide a list of some of the commonly used kernel functions:




φ(xi), φ(xj)
xi xj

K(u, v)=⟨φ(u), φ(v) ⟩=f(u, v) (4.97)


Polynomial kernelK(u, v)=(uTv+1)p (4.98)

Radial Basis Function kernelK(u, v)=e−ǁu−v ǁ2/(2σ2) (4.99)

Sigmoid kernelK(u, v)=tanh(kuTv−δ) (4.100)

By using a kernel function, we can directly work with inner products in the
transformed space without dealing with the exact forms of the nonlinear
transformation function . Specifically, this allows us to use high-dimensional
transformations (sometimes even involving infinitely many dimensions), while
performing calculations only in the original attribute space. Computing the
inner products using kernel functions is also considerably cheaper than using
the transformed attribute set . Hence, the use of kernel functions provides
a significant advantage in representing nonlinear decision boundaries, without
suffering from the curse of dimensionality. This has been one of the major
reasons behind the widespread usage of SVM in highly complex and
nonlinear problems.

Figure 4.40.
Decision boundary produced by a nonlinear SVM with polynomial kernel.

Figure 4.40 shows the nonlinear decision boundary obtained by SVM using
the polynomial kernel function given in Equation 4.98 . We can see that the



learned decision boundary is quite close to the true decision boundary shown
in Figure 4.39(a) . Although the choice of kernel function depends on the
characteristics of the input data, a commonly used kernel function is the radial
basis function (RBF) kernel, which involves a single hyper-parameter ,
known as the standard deviation of the RBF kernel.

4.9.5 Characteristics of SVM

1. The SVM learning problem can be formulated as a convex optimization
problem, in which efficient algorithms are available to find the global
minimum of the objective function. Other classification methods, such
as rule-based classifiers and artificial neural networks, employ a greedy
strategy to search the hypothesis space. Such methods tend to find
only locally optimum solutions.

2. SVM provides an effective way of regularizing the model parameters by
maximizing the margin of the decision boundary. Furthermore, it is able
to create a balance between model complexity and training errors by
using a hyper-parameter C. This trade-off is generic to a broader class
of model learning techniques that capture the model complexity and the
training loss using different formulations.

3. Linear SVM can handle irrelevant attributes by learning zero weights
corresponding to such attributes. It can also handle redundant
attributes by learning similar weights for the duplicate attributes.
Furthermore, the ability of SVM to regularize its learning makes it more
robust to the presence of a large number of irrelevant and redundant
attributes than other classifiers, even in high-dimensional settings. For
this reason, nonlinear SVMs are less impacted by irrelevant and
redundant attributes than other highly expressive classifiers that can
learn nonlinear decision boundaries such as decision trees.


To compare the effect of irrelevant attributes on the performance of
nonlinear SVMs and decision trees, consider the two-dimensional data
set shown in Figure 4.41(a) containing and instances,
where the two classes can be easily separated using a nonlinear
decision boundary. We incrementally add irrelevant attributes to this
data set and compare the performance of two classifiers: decision tree
and nonlinear SVM (using radial basis function kernel), using 70% of
the data for training and the rest for testing. Figure 4.41(b) shows
the test error rates of the two classifiers as we increase the number of
irrelevant attributes. We can see that the test error rate of decision
trees swiftly reaches 0.5 (same as random guessing) in the presence of
even a small number of irrelevant attributes. This can be attributed to
the problem of multiple comparisons while choosing splitting attributes
at internal nodes as discussed in Example 3.7 of the previous
chapter. On the other hand, nonlinear SVM shows a more robust and
steady performance even after adding a moderately large number of
irrelevant attributes. Its test error rate gradually declines and eventually
reaches close to 0.5 after adding 125 irrelevant attributes, at which
point it becomes difficult to discern the discriminative information in the
original two attributes from the noise in the remaining attributes for
learning nonlinear decision boundaries.

500+ 500o

Figure 4.41.
Comparing the effect of adding irrelevant attributes on the performance
of nonlinear SVMs and decision trees.

4. SVM can be applied to categorical data by introducing dummy
variables for each categorical attribute value present in the data. For
example, if has three values ,
we can introduce a binary variable for each of the attribute values.

5. The SVM formulation presented in this chapter is for binary class
problems. However, multiclass extensions of SVM have also been

6. Although the training time of an SVM model can be large, the learned
parameters can be succinctly represented with the help of a small
number of support vectors, making the classification of test instances
quite fast.


4.10 Ensemble Methods
This section presents techniques for improving classification accuracy by
aggregating the predictions of multiple classifiers. These techniques are
known as ensemble or classifier combination methods. An ensemble
method constructs a set of base classifiers from training data and performs
classification by taking a vote on the predictions made by each base classifier.
This section explains why ensemble methods tend to perform better than any
single classifier and presents techniques for constructing the classifier

4.10.1 Rationale for Ensemble Method

The following example illustrates how an ensemble method can improve a
classifier’s performance.

Example 4.8.
Consider an ensemble of 25 binary classifiers, each of which has an error
rate of . The ensemble classifier predicts the class label of a test
example by taking a majority vote on the predictions made by the base
classifiers. If the base classifiers are identical, then all the base classifiers
will commit the same mistakes. Thus, the error rate of the ensemble
remains 0.35. On the other hand, if the base classifiers are independent—
i.e., their errors are uncorrelated—then the ensemble makes a wrong
prediction only if more than half of the base classifiers predict incorrectly.
In this case, the error rate of the ensemble classifier is


which is considerably lower than the error rate of the base classifiers.

Figure 4.42 shows the error rate of an ensemble of 25 binary classifiers
for different base classifier error rates . The diagonal line

represents the case in which the base classifiers are identical, while the solid
line represents the case in which the base classifiers are independent.
Observe that the ensemble classifier performs worse than the base classifiers
when is larger than 0.5.

The preceding example illustrates two necessary conditions for an ensemble
classifier to perform better than a single classifier: (1) the base classifiers
should be independent of each other, and (2) the base classifiers should do
better than a classifier that performs random guessing. In practice, it is difficult
to ensure total independence among the base classifiers. Nevertheless,
improvements in classification accuracies have been observed in ensemble
methods in which the base classifiers are somewhat correlated.

4.10.2 Methods for Constructing an
Ensemble Classifier

A logical view of the ensemble method is presented in Figure 4.43 . The
basic idea is to construct multiple classifiers from the original data and then
aggregate their predictions when classifying unknown examples. The
ensemble of classifiers can be constructed in many ways:

eensemble=∑i=1325(25i)∈i(1−∈)25−i=0.06, (4.101)

(eensemble) (∈)

1. By manipulating the training set. In this approach, multiple training
sets are created by resampling the original data according to some
sampling distribution and constructing a classifier from each training
set. The sampling distribution determines how likely it is that an
example will be selected for training, and it may vary from one trial to
another. Bagging and boosting are two examples of ensemble
methods that manipulate their training sets. These methods are
described in further detail in Sections 4.10.4 and 4.10.5 .

Figure 4.42.
Comparison between errors of base classifiers and errors of the
ensemble classifier.

Figure 4.43.
A logical view of the ensemble learning method.

2. By manipulating the input features. In this approach, a subset of
input features is chosen to form each training set. The subset can be
either chosen randomly or based on the recommendation of domain
experts. Some studies have shown that this approach works very well
with data sets that contain highly redundant features. Random forest,
which is described in Section 4.10.6 , is an ensemble method that
manipulates its input features and uses decision trees as its base

3. By manipulating the class labels. This method can be used when the
number of classes is sufficiently large. The training data is transformed
into a binary class problem by randomly partitioning the class labels
into two disjoint subsets, and . Training examples whose classA0 A1

label belongs to the subset are assigned to class 0, while those that
belong to the subset are assigned to class 1. The relabeled
examples are then used to train a base classifier. By repeating this
process multiple times, an ensemble of base classifiers is obtained.
When a test example is presented, each base classifier is used to
predict its class label. If the test example is predicted as class 0, then
all the classes that belong to will receive a vote. Conversely, if it is
predicted to be class 1, then all the classes that belong to will
receive a vote. The votes are tallied and the class that receives the
highest vote is assigned to the test example. An example of this
approach is the error-correcting output coding method described on
page 331.

4. By manipulating the learning algorithm. Many learning algorithms
can be manipulated in such a way that applying the algorithm several
times on the same training data will result in the construction of
different classifiers. For example, an artificial neural network can
change its network topology or the initial weights of the links between
neurons. Similarly, an ensemble of decision trees can be constructed
by injecting randomness into the tree-growing procedure. For example,
instead of choosing the best splitting attribute at each node, we can
randomly choose one of the top k attributes for splitting.

The first three approaches are generic methods that are applicable to any
classifier, whereas the fourth approach depends on the type of classifier used.
The base classifiers for most of these approaches can be generated
sequentially (one after another) or in parallel (all at once). Once an ensemble
of classifiers has been learned, a test example is classified by combining
the predictions made by the base classifiers :





C*(x)=f(C1(x), C2(x), …, Ck(x)).

where f is the function that combines the ensemble responses. One simple
approach for obtaining is to take a majority vote of the individual
predictions. An alternate approach is to take a weighted majority vote, where
the weight of a base classifier denotes its accuracy or relevance.

Ensemble methods show the most improvement when used with unstable
classifiers, i.e., base classifiers that are sensitive to minor perturbations in
the training set, because of high model complexity. Although unstable
classifiers may have a low bias in finding the optimal decision boundary, their
predictions have a high variance for minor changes in the training set or
model selection. This trade-off between bias and variance is discussed in
detail in the next section. By aggregating the responses of multiple unstable
classifiers, ensemble learning attempts to minimize their variance without
worsening their bias.

4.10.3 Bias-Variance Decomposition

Bias-variance decomposition is a formal method for analyzing the
generalization error of a predictive model. Although the analysis is slightly
different for classification than regression, we first discuss the basic intuition of
this decomposition by using an analogue of a regression problem.

Consider the illustrative task of reaching a target y by firing projectiles from a
starting position , as shown in Figure 4.44 . The target corresponds to the
desired output at a test instance, while the starting position corresponds to its
observed attributes. In this analogy, the projectile represents the model used
for predicting the target using the observed attributes. Let denote the point
where the projectile hits the ground, which is analogous of the prediction of
the model.



Figure 4.44.
Bias-variance decomposition.

Ideally, we would like our predictions to be as close to the true target as
possible. However, note that different trajectories of projectiles are possible
based on differences in the training data or in the approach used for model
selection. Hence, we can observe a variance in the predictions over
different runs of projectile. Further, the target in our example is not fixed but
has some freedom to move around, resulting in a noise component in the true
target. This can be understood as the non-deterministic nature of the output
variable, where the same set of attributes can have different output values.
Let represent the average prediction of the projectile over multiple runs,
and denote the average target value. The difference between and

is known as the bias of the model.

In the context of classification, it can be shown that the generalization error of
a classification model m can be decomposed into terms involving the bias,
variance, and noise components of the model in the following way:

where and are constants that depend on the characteristics of training
and test sets. Note that while the noise term is intrinsic to the target class, the


yavg y^avg



c1 c2

bias and variance terms depend on the choice of the classification model. The
bias of a model represents how close the average prediction of the model is to
the average target. Models that are able to learn complex decision
boundaries, e.g., models produced by k-nearest neighbor and multi-layer
ANN, generally show low bias. The variance of a model captures the stability
of its predictions in response to minor perturbations in the training set or the
model selection approach.

We can say that a model shows better generalization performance if it has a
lower bias and lower variance. However, if the complexity of a model is high
but the training size is small, we generally expect to see a lower bias but
higher variance, resulting in the phenomena of overfitting. This phenomena is
pictorially represented in Figure 4.45(a) . On the other hand, an overly
simplistic model that suffers from underfitting may show a lower variance but
would suffer from a high bias, as shown in Figure 4.45(b) . Hence, the
trade-off between bias and variance provides a useful way for interpreting the
effects of underfitting and overfitting on the generalization performance of a

Figure 4.45.
Plots showing the behavior of two-dimensional solutions with constant and

The bias-variance trade-off can be used to explain why ensemble learning
improves the generalization performance of unstable classifiers. If a base
classifier show low bias but high variance, it can become susceptible to
overfitting, as even a small change in the training set will result in different
predictions. However, by combining the responses of multiple base classifiers,
we can expect to reduce the overall variance. Hence, ensemble learning
methods show better performance primarily by lowering the variance in the
predictions, although they can even help in reducing the bias. One of the
simplest approaches for combining predictions and reducing their variance is
to compute their average. This forms the basis of the bagging method,
described in the following subsection.


4.10.4 Bagging

Bagging, which is also known as bootstrap aggregating, is a technique that
repeatedly samples (with replacement) from a data set according to a uniform
probability distribution. Each bootstrap sample has the same size as the
original data. Because the sampling is done with replacement, some
instances may appear several times in the same training set, while others may
be omitted from the training set. On average, a bootstrap sample contains
approximately 63% of the original training data because each sample has a
probability of being selected in each . If N is sufficiently large,
this probability converges to . The basic procedure for bagging is
summarized in Algorithm 4.5 . After training the k classifiers, a test
instance is assigned to the class that receives the highest number of votes.

To illustrate how bagging works, consider the data set shown in Table 4.4 .
Let x denote a one-dimensional attribute and y denote the class label.
Suppose we use only one-level binary decision trees, with a test condition

, where k is a split point chosen to minimize the entropy of the leaf nodes.
Such a tree is also known as a decision stump.

Table 4.4. Example of data set used to construct an ensemble of bagging

x 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

y 1 1 1 1 1 1

Without bagging, the best decision stump we can produce splits the instances
at either or . Either way, the accuracy of the tree is at most 70%.
Suppose we apply the bagging procedure on the data set using 10 bootstrap
samples. The examples chosen for training in each bagging round are shown


1−(1−1/N)N Di


−1 −1 −1 −1

x≤0.35 x≤0.75

in Figure 4.46 . On the right-hand side of each table, we also describe the
decision stump being used in each round.

We classify the entire data set given in Table 4.4 by taking a majority vote
among the predictions made by each base classifier. The results of the
predictions are shown in Figure 4.47 . Since the class labels are either
or , taking the majority vote is equivalent to summing up the predicted
values of y and examining the sign of the resulting sum (refer to the second to
last row in Figure 4.47 ). Notice that the ensemble classifier perfectly
classifies all 10 examples in the original data.

Algorithm 4.5 Bagging algorithm.


Figure 4.46.
Example of bagging.

The preceding example illustrates another advantage of using ensemble
methods in terms of enhancing the representation of the target function. Even
though each base classifier is a decision stump, combining the classifiers can
lead to a decision boundary that mimics a decision tree of depth 2.

Bagging improves generalization error by reducing the variance of the base
classifiers. The performance of bagging depends on the stability of the base
classifier. If a base classifier is unstable, bagging helps to reduce the errors
associated with random fluctuations in the training data. If a base classifier is
stable, i.e., robust to minor perturbations in the training set, then the error of
the ensemble is primarily caused by bias in the base classifier. In this
situation, bagging may not be able to improve the performance of the base
classifiers significantly. It may even degrade the classifier’s performance
because the effective size of each training set is about 37% smaller than the
original data.

Figure 4.47.
Example of combining classifiers constructed using the bagging approach.

4.10.5 Boosting

Boosting is an iterative procedure used to adaptively change the distribution of
training examples for learning base classifiers so that they increasingly focus
on examples that are hard to classify. Unlike bagging, boosting assigns a
weight to each training example and may adaptively change the weight at the
end of each boosting round. The weights assigned to the training examples
can be used in the following ways:

1. They can be used to inform the sampling distribution used to draw a set
of bootstrap samples from the original data.

2. They can be used to learn a model that is biased toward examples with
higher weight.

This section describes an algorithm that uses weights of examples to
determine the sampling distribution of its training set. Initially, the examples
are assigned equal weights, 1/N, so that they are equally likely to be chosen
for training. A sample is drawn according to the sampling distribution of the
training examples to obtain a new training set. Next, a classifier is built from
the training set and used to classify all the examples in the original data. The
weights of the training examples are updated at the end of each boosting
round. Examples that are classified incorrectly will have their weights
increased, while those that are classified correctly will have their weights
decreased. This forces the classifier to focus on examples that are difficult to
classify in subsequent iterations.

The following table shows the examples chosen during each boosting round,
when applied to the data shown in Table 4.4 .

Boosting (Round 1): 7 3 2 8 7 9 4 10 6 3

Boosting (Round 2): 5 4 9 4 2 5 1 7 4 2

Boosting (Round 3): 4 4 8 10 4 5 4 6 3 4

Initially, all the examples are assigned the same weights. However, some
examples may be chosen more than once, e.g., examples 3 and 7, because
the sampling is done with replacement. A classifier built from the data is then
used to classify all the examples. Suppose example 4 is difficult to classify.
The weight for this example will be increased in future iterations as it gets
misclassified repeatedly. Meanwhile, examples that were not chosen in the
previous round, e.g., examples 1 and 5, also have a better chance of being
selected in the next round since their predictions in the previous round were
likely to be wrong. As the boosting rounds proceed, examples that are the
hardest to classify tend to become even more prevalent. The final ensemble is
obtained by aggregating the base classifiers obtained from each boosting

Over the years, several implementations of the boosting algorithm have been
developed. These algorithms differ in terms of (1) how the weights of the
training examples are updated at the end of each boosting round, and (2) how
the predictions made by each classifier are combined. An implementation
called AdaBoost is explored in the next section.

Let denote a set of N training examples. In the AdaBoost
algorithm, the importance of a base classifier depends on its

{(xj, yj)|j=1, 2, …, N}

Figure 4.48.
Plot of as a function of training error .

error rate, which is defined as

where if the predicate p is true, and 0 otherwise. The importance of a
classifier is given by the following parameter,

Note that has a large positive value if the error rate is close to 0 and a large
negative value if the error rate is close to 1, as shown in Figure 4.48 .

The parameter is also used to update the weight of the training examples.
To illustrate, let denote the weight assigned to example ( during the

α ∈

∈i=1N[∑j=1Nwj I(Ci(xj)≠yj) ], (4.102)


αi=12ln (1−∈i∈i).


wi(j) xi, yi)


j boosting round. The weight update mechanism for AdaBoost is given by the

where is the normalization factor used to ensure that . The
weight update formula given in Equation 4.103 increases the weights of
incorrectly classified examples and decreases the weights of those classified

Instead of using a majority voting scheme, the prediction made by each
classifier is weighted according to . This approach allows AdaBoost to
penalize models that have poor accuracy, e.g., those generated at the earlier
boosting rounds. In addition, if any intermediate rounds produce an error rate
higher than 50%, the weights are reverted back to their original uniform
values, , and the resampling procedure is repeated. The AdaBoost
algorithm is summarized in Algorithm 4.6 .

Algorithm 4.6 AdaBoost algorithm.

∈ ∑


wi(j+1)=wi(j)Zj×{e−αjif Cj(xi)=yi,eαjif Cj(xi)≠yi (4.103)

Zj ∑iwi(j+1)=1

Cj αj


∈ ∈

Let us examine how the boosting approach works on the data set shown in
Table 4.4 . Initially, all the examples have identical weights. After three
boosting rounds, the examples chosen for training are shown in Figure
4.49(a) . The weights for each example are updated at the end of each
boosting round using Equation 4.103 , as shown in Figure 4.50(b) .

Without boosting, the accuracy of the decision stump is, at best, 70%. With
AdaBoost, the results of the predictions are given in Figure 4.50(b) . The
final prediction of the ensemble classifier is obtained by taking a weighted
average of the predictions made by each base classifier, which is shown in the
last row of Figure 4.50(b) . Notice that AdaBoost perfectly classifies all the
examples in the training data.

Figure 4.49.
Example of boosting.

An important analytical result of boosting shows that the training error of the
ensemble is bounded by the following expression:

where is the error rate of each base classifier i. If the error rate of the base
classifier is less than 50%, we can write , where measures how
much better the classifier is than random guessing. The bound on the training
error of the ensemble becomes

eensemble≤∏i[∈i(1−∈i) ], (4.104)

∈i=0.5 −γi γi

Hence, the training error of the ensemble decreases exponentially, which
leads to the fast convergence of the algorithm. By focusing on examples that
are difficult to classify by base classifiers, it is able to reduce the bias of the
final predictions along with the variance. AdaBoost has been shown to provide
significant improvements in performance over base classifiers on a range of
data sets. Nevertheless, because of its tendency to focus on training
examples that are wrongly classified, the boosting technique can be
susceptible to overfitting, resulting in poor generalization performance in some

Figure 4.50.
Example of combining classifiers constructed using the AdaBoost approach.

eensemble≤∏i1−4γi2≤exp(−2∑iγi2). (4.105)

4.10.6 Random Forests

Random forests attempt to improve the generalization performance by
constructing an ensemble of decorrelated decision trees. Random forests
build on the idea of bagging to use a different bootstrap sample of the training
data for learning decision trees. However, a key distinguishing feature of
random forests from bagging is that at every internal node of a tree, the best
splitting criterion is chosen among a small set of randomly selected attributes.
In this way, random forests construct ensembles of decision trees by not only
manipulating training instances (by using bootstrap samples similar to
bagging), but also the input attributes (by using different subsets of attributes
at every internal node).

Given a training set D consisting of n instances and d attributes, the basic
procedure of training a random forest classifier can be summarized using the
following steps:

1. Construct a bootstrap sample of the training set by randomly
sampling n instances (with replacement) from D.

2. Use to learn a decision tree as follows. At every internal node of
, randomly sample a set of p attributes and choose an attribute from

this subset that shows the maximum reduction in an impurity measure
for splitting. Repeat this procedure till every leaf is pure, i.e., containing
instances from the same class.

Once an ensemble of decision trees have been constructed, their average
prediction (majority vote) on a test instance is used as the final prediction of
the random forest. Note that the decision trees involved in a random forest are
unpruned trees, as they are allowed to grow to their largest possible size till
every leaf is pure. Hence, the base classifiers of random forest represent


Di Ti

unstable classifiers that have low bias but high variance, because of their
large size.

Another property of the base classifiers learned in random forests is the lack
of correlation among their model parameters and test predictions. This can be
attributed to the use of an independently sampled data set for learning
every decision tree , similar to the bagging approach. However, random
forests have the additional advantage of choosing a splitting criterion at every
internal node using a different (and randomly selected) subset of attributes.
This property significantly helps in breaking the correlation structure, if any,
among the decision trees .

To realize this, consider a training set involving a large number of attributes,
where only a small subset of attributes are strong predictors of the target
class, whereas other attributes are weak indicators. Given such a training set,
even if we consider different bootstrap samples for learning , we would
mostly be choosing the same attributes for splitting at internal nodes, because
the weak attributes would be largely overlooked when compared with the
strong predictors. This can result in a considerable correlation among the
trees. However, if we restrict the choice of attributes at every internal node to
a random subset of attributes, we can ensure the selection of both strong and
weak predictors, thus promoting diversity among the trees. This principle is
utilized by random forests for creating decorrelated decision trees.

By aggregating the predictions of an ensemble of strong and decorrelated
decision trees, random forests are able to reduce the variance of the trees
without negatively impacting their low bias. This makes random forests quite
robust to overfitting. Additionally, because of their ability to consider only a
small subset of attributes at every internal node, random forests are
computationally fast and robust even in high-dimensional settings.



Di Ti

The number of attributes to be selected at every node, p, is a hyper-parameter
of the random forest classifier. A small value of p can reduce the correlation
among the classifiers but may also reduce their strength. A large value can
improve their strength but may result in correlated trees similar to bagging.
Although common suggestions for p in the literature include and , a
suitable value of p for a given training set can always be selected by tuning it
over a validation set, as described in the previous chapter. However, there is
an alternative way for selecting hyper-parameters in random forests, which
does not require using a separate validation set. It involves computing a
reliable estimate of the generalization error rate directly during training, known
as the out-of-bag (oob) error estimate. The oob estimate can be computed
for any generic ensemble learning method that builds independent base
classifiers using bootstrap samples of the training set, e.g., bagging and
random forests. The approach for computing oob estimate can be described
as follows.

Consider an ensemble learning method that uses an independent base
classifier built on a bootstrap sample of the training set . Since every
training instance will be used for training approximately 63% of base
classifiers, we can call as an out-of-bag sample for the remaining 27% of
base classifiers that did not use it for training. If we use these remaining 27%
classifiers to make predictions on , we can obtain the oob error on by
taking their majority vote and comparing it with its class label. Note that the
oob error estimates the error of 27% classifiers on an instance that was not
used for training those classifiers. Hence, the oob error can be considered as
a reliable estimate of generalization error. By taking the average of oob errors
of all training instances, we can compute the overall oob error estimate. This
can be used as an alternative to the validation error rate for selecting hyper-
parameters. Hence, random forests do not need to use a separate partition of
the training set for validation, as it can simultaneously train the base
classifiers and compute generalization error estimates on the same data set.

d log2d+1

Ti Di

Random forests have been empirically found to provide significant
improvements in generalization performance that are often comparable, if not
superior, to the improvements provided by the AdaBoost algorithm. Random
forests are also more robust to overfitting and run much faster than the
AdaBoost algorithm.

4.10.7 Empirical Comparison among
Ensemble Methods

Table 4.5 shows the empirical results obtained when comparing the
performance of a decision tree classifier against bagging, boosting, and
random forest. The base classifiers used in each ensemble method consist of
50 decision trees. The classification accuracies reported in this table are
obtained from tenfold cross-validation. Notice that the ensemble classifiers
generally outperform a single decision tree classifier on many of the data sets.

Table 4.5. Comparing the accuracy of a decision tree classifier against
three ensemble methods.

Data Set Number of (Attributes,
Classes, Instances)

Tree (%)

Bagging(%) Boosting(%) RF(%)

Anneal (39, 6, 898) 92.09 94.43 95.43 95.43

Australia (15, 2, 690) 85.51 87.10 85.22 85.80

Auto (26, 7, 205) 81.95 85.37 85.37 84.39

Breast (11, 2, 699) 95.14 96.42 97.28 96.14

Cleve (14, 2, 303) 76.24 81.52 82.18 82.18

Credit (16, 2, 690) 85.8 86.23 86.09 85.8

Diabetes (9, 2, 768) 72.40 76.30 73.18 75.13

German (21, 2, 1000) 70.90 73.40 73.00 74.5

Glass (10, 7, 214) 67.29 76.17 77.57 78.04

Heart (14, 2, 270) 80.00 81.48 80.74 83.33

Hepatitis (20, 2, 155) 81.94 81.29 83.87 83.23

Horse (23, 2, 368) 85.33 85.87 81.25 85.33

Ionosphere (35, 2, 351) 89.17 92.02 93.73 93.45

Iris (5, 3, 150) 94.67 94.67 94.00 93.33

Labor (17, 2, 57) 78.95 84.21 89.47 84.21

Led7 (8, 10, 3200) 73.34 73.66 73.34 73.06

Lymphography (19, 4, 148) 77.03 79.05 85.14 82.43

Pima (9, 2, 768) 74.35 76.69 73.44 77.60

Sonar (61, 2, 208) 78.85 78.85 84.62 85.58

Tic-tac-toe (10, 2, 958) 83.72 93.84 98.54 95.82

Vehicle (19, 4, 846) 71.04 74.11 78.25 74.94

Waveform (22, 3, 5000) 76.44 83.30 83.90 84.04

Wine (14, 3, 178) 94.38 96.07 97.75 97.75

Zoo (17, 7, 101) 93.07 93.07 95.05 97.03

4.11 Class Imbalance Problem
In many data sets there are a disproportionate number of instances that
belong to different classes, a property known as skew or class
imbalance.For example, consider a health-care application where diagnostic
reports are used to decide whether a person has a rare disease. Because of
the infrequent nature of the disease, we can expect to observe a smaller
number of subjects who are positively diagnosed. Similarly, in credit card
fraud detection, fraudulent transactions are greatly outnumbered by legitimate

The degree of imbalance between the classes varies across different
applications and even across different data sets from the same application.
For example, the risk for a rare disease may vary across different populations
of subjects depending on their dietary and lifestyle choices. However, despite
their infrequent occurrences, a correct classification of the rare class often has
greater value than a correct classification of the majority class. For example, it
may be more dangerous to ignore a patient suffering from a disease than to
misdiagnose a healthy person.

More generally, class imbalance poses two challenges for classification. First,
it can be difficult to find sufficiently many labeled samples of a rare class. Note
that many of the classification methods discussed so far work well only when
the training set has a balanced representation of both classes. Although some
classifiers are more effective at handling imbalance in the training data than
others, e.g., rule-based classifiers and k-NN, they are all impacted if the
minority class is not well-represented in the training set. In general, a classifier
trained over an imbalanced data set shows a bias toward improving its
performance over the majority class, which is often not the desired behavior.

As a result, many existing classification models, when trained on an
imbalanced data set, may not effectively detect instances of the rare class.

Second, accuracy, which is the traditional measure for evaluating
classification performance, is not well-suited for evaluating models in the
presence of class imbalance in the test data. For example, if 1% of the credit
card transactions are fraudulent, then a trivial model that predicts every
transaction as legitimate will have an accuracy of 99% even though it fails to
detect any of the fraudulent activities. Thus, there is a need to use alternative
evaluation metrics that are sensitive to the skew and can capture different
criteria of performance than accuracy.

In this section, we first present some of the generic methods for building
classifiers when there is class imbalance in the training set. We then discuss
methods for evaluating classification performance and adapting classification
decisions in the presence of a skewed test set. In the remainder of this
section, we will consider binary classification problems for simplicity, where
the minority class is referred as the positive class while the majority class
is referred as the negative class.

4.11.1 Building Classifiers with Class

There are two primary considerations for building classifiers in the presence of
class imbalance in the training set. First, we need to ensure that the learning
algorithm is trained over a data set that has adequate representation of both
the majority as well as the minority classes. Some common approaches for
ensuring this includes the methodologies of oversampling and undersampling


the training set. Second, having learned a classification model, we need a way
to adapt its classification decisions (and thus create an appropriately tuned
classifier) to best match the requirements of the imbalanced test set. This is
typically done by converting the outputs of the classification model to real-
valued scores, and then selecting a suitable threshold on the classification
score to match the needs of a test set. Both these considerations are
discussed in detail in the following.

Oversampling and Undersampling
The first step in learning with imbalanced data is to transform the training set
to a balanced training set, where both classes have nearly equal
representation. The balanced training set can then be used with any of the
existing classification techniques (without making any modifications in the
learning algorithm) to learn a model that gives equal emphasis to both
classes. In the following, we present some of the common techniques for
transforming an imbalanced training set to a balanced one.

A basic approach for creating balanced training sets is to generate a sample
of training instances where the rare class has adequate representation. There
are two types of sampling methods that can be used to enhance the
representation of the minority class: (a) undersampling, where the frequency
of the majority class is reduced to match the frequency of the minority class,
and (b) oversampling, where artificial examples of the minority class are
created to make them equal in proportion to the number of negative instances.

To illustrate undersampling, consider a training set that contains 100 positive
examples and 1000 negative examples. To overcome the skew among the
classes, we can select a random sample of 100 examples from the negative
class and use them with the 100 positive examples to create a balanced
training set. A classifier built over the resultant balanced set will then be

unbiased toward both classes. However, one limitation of undersampling is
that some of the useful negative examples (e.g., those closer to the actual
decision boundary) may not be chosen for training, therefore, resulting in an
inferior classification model. Another limitation is that the smaller sample of
100 negative instances may have a higher variance than the larger set of

Oversampling attempts to create a balanced training set by artificially
generating new positive examples. A simple approach for oversampling is to
duplicate every positive instance times, where and are the
numbers of positive and negative training instances, respectively. Figure
4.51 illustrates the effect of oversampling on the learning of a decision
boundary using a classifier such as a decision tree. Without oversampling,
only the positive examples at the bottom right-hand side of Figure 4.51(a)
are classified correctly. The positive example in the middle of the diagram is
misclassified because there are not enough examples to justify the creation of
a new decision boundary to separate the positive and negative instances.
Oversampling provides the additional examples needed to ensure that the
decision boundary surrounding the positive example is not pruned, as
illustrated in Figure 4.51(b) . Note that duplicating a positive instance is
analogous to doubling its weight during the training stage. Hence, the effect of
oversampling can be alternatively achieved by assigning higher weights to
positive instances than negative instances. This method of weighting
instances can be used with a number of classifiers such as logistic regression,
ANN, and SVM.

n−/n+ n+ n−

Figure 4.51.
Illustrating the effect of oversampling of the rare class.

One limitation of the duplication method for oversampling is that the replicated
positive examples have an artificially lower variance when compared with their
true distribution in the overall data. This can bias the classifier to the specific
distribution of training instances, which may not be representative of the
overall distribution of test instances, leading to poor generalizability. To
overcome this limitation, an alternative approach for oversampling is to
generate synthetic positive instances in the neighborhood of existing positive
instances. In this approach, called the Synthetic Minority Oversampling
Technique (SMOTE), we first determine the k-nearest positive neighbors of
every positive instance , and then generate a synthetic positive instance at
some intermediate point along the line segment joining to one of its
randomly chosen k-nearest neighbor, . This process is repeated until the
desired number of positive instances is reached. However, one limitation of
this approach is that it can only generate new positive instances in the convex
hull of the existing positive class. Hence, it does not help improve the
representation of the positive class outside the boundary of existing positive


instances. Despite their complementary strengths and weaknesses,
undersampling and oversampling provide useful directions for generating
balanced training sets in the presence of class imbalance.

Assigning Scores to Test Instances
If a classifier returns an ordinal score s( )for every test instance such that a
higher score denotes a greater likelihood of belonging to the positive class,
then for every possible value of score threshold, , we can create a new
binary classifier where a test instance is classified positive only if .
Thus, every choice of can potentially lead to a different classifier, and we
are interested in finding the classifier that is best suited for our needs.

Ideally, we would like the classification score to vary monotonically with the
actual posterior probability of the positive class, i.e., if and are the
scores of any two instances, and , then

. However, this is difficult to guarantee in
practice as the properties of the classification score depends on several
factors such as the complexity of the classification algorithm and the
representative power of the training set. In general, we can only expect the
classification score of a reasonable algorithm to be weakly related to the
actual posterior probability of the positive class, even though the relationship
may not be strictly monotonic. Most classifiers can be easily modified to
produce such a real valued score. For example, the signed distance of an
instance from the positive margin hyperplane of SVM can be used as a
classification score. As another example, test instances belonging to a leaf in
a decision tree can be assigned a score based on the fraction of training
instances labeled as positive in the leaf. Also, probabilistic classifiers such as
naïve Bayes, Bayesian networks, and logistic regression naturally output
estimates of posterior probabilities, . Next, we discuss some



s(x1) s(x2)
x1 x2



evaluation measures for assessing the goodness of a classifier in the
presence of class imbalance.

Table 4.6. A confusion matrix for a binary classification problem in which
the classes are not equally important.

Predicted Class


4.11.2 Evaluating Performance with
Class Imbalance

The most basic approach for representing a classifier’s performance on a test
set is to use a confusion matrix, as shown in Table 4.6 . This table is
essentially the same as Table 3.4 , which was introduced in the context of
evaluating classification performance in Section 3.2 . A confusion matrix
summarizes the number of instances predicted correctly or incorrectly by a
classifier using the following four counts:

True positive (TP) or , which corresponds to the number of positive
examples correctly predicted by the classifier.
False positive (FP) or (also known as Type I error), which corresponds
to the number of negative examples wrongly predicted as positive by the

+ −

+ f++ (TP) f+− (FN)

− f−+ (FP) f−− (TN)



False negative (FN) or (also known as Type II error), which
corresponds to the number of positive examples wrongly predicted as
negative by the classifier.
True negative (TN) or , which corresponds to the number of negative
examples correctly predicted by the classifier.

The confusion matrix provides a concise representation of classification
performance on a given test data set. However, it is often difficult to interpret
and compare the performance of classifiers using the four-dimensional
representations (corresponding to the four counts) provided by their confusion
matrices. Hence, the counts in the confusion matrix are often summarized
using a number of evaluation measures. Accuracy is an example of one
such measure that combines these four counts into a single value, which is
used extensively when classes are balanced. However, the accuracy measure
is not suitable for handling data sets with imbalanced class distributions as it
tends to favor classifiers that correctly classify the majority class. In the
following, we describe other possible measures that capture different criteria
of performance when working with imbalanced classes.

A basic evaluation measure is the true positive rate (TPR), which is defined
as the fraction of positive test instances correctly predicted by the classifier:

In the medical community, TPR is also known as sensitivity, while in the
information retrieval literature, it is also called recall (r). A classifier with a high
TPR has a high chance of correctly identifying the positive instances of the

Analogously to TPR, the true negative rate (TNR) (also known as
specificity) is defined as the fraction of negative test instances correctly




predicted by the classifier, i.e.,

A high TNR value signifies that the classifier correctly classifies any randomly
chosen negative instance in the test set. A commonly used evaluation
measure that is closely related to TNR is the false positive rate (FPR), which
is defined as .

Similarly, we can define false negative rate (FNR) as .

Note that the evaluation measures defined above do not take into account the
skew among the classes, which can be formally defined as , where
P and N denote the number of actual positives and actual negatives,
respectively. As a result, changing the relative numbers of P and N will have
no effect on TPR, TNR, FPR, or FNR, since they depend only on the fraction
of correct classifications for every class, independently of the other class.
Furthermore, knowing the values of TPR and TNR (and consequently FNR
and FPR) does not by itself help us uniquely determine all four entries of the
confusion matrix. However, together with information about the skew factor, ,
and the total number of instances, N, we can compute the entire confusion
matrix using TPR and TNR, as shown in Table 4.7 .

Table 4.7. Entries of the confusion matrix in terms of the TPR, TNR,
skew, , and total number of instances, N.

Predicted Predicted









+ −




An evaluation measure that is sensitive to the skew is precision, which can
be defined as the fraction of correct predictions of the positive class over the
total number of positive predictions, i.e.,

Precision is also referred as the positive predicted value (PPV). A classifier
that has a high precision is likely to have most of its positive predictions
correct. Precision is a useful measure for highly skewed test sets where the
positive predictions, even though small in numbers, are required to be mostly
correct. A measure that is closely related to precision is the false discovery
rate (FDR), which can be defined as .

Although both FDR and FPR focus on FP, they are designed to capture
different evaluation objectives and thus can take quite contrasting values,
especially in the presence of class imbalance. To illustrate this, consider a
classifier with the following confusion matrix.

Predicted Class


100 0

+ TPR×α×N (1−TPR)×α×N α×N

− (1−TNR)×(1−α)×N TNR×(1−α)×N (1−α)×N

Precision, p=TPTP+FP.



+ −


100 900

Since half of the positive predictions made by the classifier are incorrect, it
has a FDR value of . However, its FPR is equal to

, which is quite low. This example shows that in the
presence of high skew (i.e., very small value of ), even a small FPR can
result in high FDR. See Section 10.6 for further discussion of this issue.

Note that the evaluation measures defined above provide an incomplete
representation of performance, because they either only capture the effect of
false positives (e.g., FPR and precision) or the effect of false negatives (e.g.,
TPR or recall), but not both. Hence, if we optimize only one of these
evaluation measures, we may end up with a classifier that shows low FN but
high FP, or vice-versa. For example, a classifier that declares every instance
to be positive will have a perfect recall, but high FPR and very poor precision.
On the other hand, a classifier that is very conservative in classifying an
instance as positive (to reduce FP) may end up having high precision but very
poor recall. We thus need evaluation measures that account for both types of
misclassifications, FP and FN. Some examples of such evaluation measures
are summarized by the following definitions.

While some of these evaluation measures are invariant to the skew (e.g., the
positive likelihood ratio), others (e.g., precision and the measure) are
sensitive to skew. Further, different evaluation measures capture the effects of
different types of misclassification errors in various ways. For example, the
measure represents a harmonic mean between recall and precision, i.e.,



Positive Likelihood Ratio=TPRFPR.F1 measure=2rpr+p=2×TP2×TP+FP+FN.G




Because the harmonic mean of two numbers tends to be closer to the smaller
of the two numbers, a high value of -measure ensures that both precision
and recall are reasonably high. Similarly, the G measure represents the
geometric mean between recall and precision. A comparison among
harmonic, geometric, and arithmetic means is given in the next example.

Example 4.9.
Consider two positive numbers and . Their arithmetic mean is

and their geometric mean is . Their harmonic mean
is , which is closer to the smaller value between a and
b than the arithmetic and geometric means.

A generic extension of the measure is the measure, which can be
defined as follows.

Both precision and recall can be viewed as special cases of by setting
and , respectively. Low values of make closer to precision, and high
values make it closer to recall.

A more general measure that captures as well as accuracy is the weighted
accuracy measure, which is defined by the following equation:

The relationship between weighted accuracy and other performance
measures is summarized in the following table:



a=1 b=5 μa=
(a+b)/2=3 μg=ab=2.236

F1 Fβ

Fβ=(β2+1)rpr+β2p=(β2+1)×TP(β2+1)TP+β2FP+FN. (4.106)

Fβ β=0
β=∞ β Fβ

Weighted accuracy=w1TP+w4TNw1TP+w2FP+w3FN+w4TN. (4.107)

w1 w2 w3 w4

Recall 1 1 0 0

Precision 1 0 1 0

1 0

Accuracy 1 1 1 1

4.11.3 Finding an Optimal Score

Given a suitably chosen evaluation measure E and a distribution of
classification scores, , on a validation set, we can obtain the optimal score
threshold on the validation set using the following approach:

1. Sort the scores in increasing order of their values.
2. For every unique value of score, s, consider the classification model

that assigns an instance as positive only if . Let E(s) denote
the performance of this model on the validation set.

3. Find that maximizes the evaluation measure, E(s).

Note that can be treated as a hyper-parameter of the classification
algorithm that is learned during model selection. Using , we can assign a
positive label to a future test instance only if . If the evaluation
measure E is skew invariant (e.g., Positive Likelihood Ratio), then we can
select without knowing the skew of the test set, and the resultant classifier
formed using can be expected to show optimal performance on the test set

Fβ β2+1 β2



s*=argmaxs E(s).




(with respect to the evaluation measure E). On the other hand, if E is sensitive
to the skew (e.g., precision or -measure), then we need to ensure that the
skew of the validation set used for selecting is similar to that of the test set,
so that the classifier formed using shows optimal test performance with
respect to E. Alternatively, given an estimate of the skew of the test data, ,
we can use it along with the TPR and TNR on the validation set to estimate all
entries of the confusion matrix (see Table 4.7 ), and thus the estimate of
any evaluation measure E on the test set. The score threshold selected
using this estimate of E can then be expected to produce optimal test
performance with respect to E. Furthermore, the methodology of selecting
on the validation set can help in comparing the test performance of different
classification algorithms, by using the optimal values of for each algorithm.

4.11.4 Aggregate Evaluation of

Although the above approach helps in finding a score threshold that
provides optimal performance with respect to a desired evaluation measure
and a certain amount of skew, , sometimes we are interested in evaluating
the performance of a classifier on a number of possible score thresholds,
each corresponding to a different choice of evaluation measure and skew
value. Assessing the performance of a classifier over a range of score
thresholds is called aggregate evaluation of performance. In this style of
analysis, the emphasis is not on evaluating the performance of a single
classifier corresponding to the optimal score threshold, but to assess the
general quality of ranking produced by the classification scores on the test set.
In general, this helps in obtaining robust estimates of classification
performance that are not sensitive to specific choices of score thresholds.








ROC Curve
One of the widely-used tools for aggregate evaluation is the receiver
operating characteristic (ROC) curve. An ROC curve is a graphical
approach for displaying the trade-off between TPR and FPR of a classifier,
over varying score thresholds. In an ROC curve, the TPR is plotted along the
y-axis and the FPR is shown on the x-axis. Each point along the curve
corresponds to a classification model generated by placing a threshold on the
test scores produced by the classifier. The following procedure describes the
generic approach for computing an ROC curve:

1. Sort the test instances in increasing order of their scores.
2. Select the lowest ranked test instance (i.e., the instance with lowest

score). Assign the selected instance and those ranked above it to the
positive class. This approach is equivalent to classifying all the test
instances as positive class. Because all the positive examples are
classified correctly and the negative examples are misclassified,

3. Select the next test instance from the sorted list. Classify the selected

instance and those ranked above it as positive, while those ranked
below it as negative. Update the counts of TP and FP by examining the
actual class label of the selected instance. If this instance belongs to
the positive class, the TP count is decremented and the FP count
remains the same as before. If the instance belongs to the negative
class, the FP count is decremented and TP count remains the same as

4. Repeat Step 3 and update the TP and FP counts accordingly until the
highest ranked test instance is selected. At this final threshold,

, as all instances are labeled as negative.
5. Plot the TPR against FPR of the classifier.



Example 4.10. [Generating ROC Curve]
Figure 4.52 shows an example of how to compute the TPR and FPR
values for every choice of score threshold. There are five positive
examples and five negative examples in the test set. The class labels of
the test instances are shown in the first row of the table, while the second
row corresponds to the sorted score values for each instance. The next six
rows contain the counts of TP , FP , TN, and FN, along with their
corresponding TPR and FPR. The table is then filled from left to right.
Initially, all the instances are predicted to be positive. Thus, and

. Next, we assign the test instance with the lowest score as
the negative class. Because the selected instance is actually a positive
example, the TP count decreases from 5 to 4 and the FP count is the
same as before. The FPR and TPR are updated accordingly. This process
is repeated until we reach the end of the list, where and .
The ROC curve for this example is shown in Figure 4.53 .

Figure 4.52.
Computing the TPR and FPR at every score threshold.



Figure 4.53.
ROC curve for the data shown in Figure 4.52 .

Note that in an ROC curve, the TPR monotonically increases with FPR,
because the inclusion of a test instance in the set of predicted positives can
either increase the TPR or the FPR. The ROC curve thus has a staircase
pattern. Furthermore, there are several critical points along an ROC curve that
have well-known interpretations:

: Model predicts every instance to be a negative class.

: Model predicts every instance to be a positive class.

: The perfect model with zero misclassifications.

A good classification model should be located as close as possible to the
upper left corner of the diagram, while a model that makes random guesses
should reside along the main diagonal, connecting the points
and . Random guessing means that an instance is classified
as a positive class with a fixed probability p, irrespective of its attribute set.

(TPR=0, FPR=0)

(TPR=1, FPR=1)

(TPR=1, FPR=0)

(TPR=0, FPR=0)
(TPR=1, FPR=1)

For example, consider a data set that contains positive instances and
negative instances. The random classifier is expected to correctly classify
of the positive instances and to misclassify of the negative instances.
Therefore, the TPR of the classifier is , while its FPR is .
Hence, this random classifier will reside at the point (p, p) in the ROC curve
along the main diagonal.

Figure 4.54.
ROC curves for two different classifiers.

Since every point on the ROC curve represents the performance of a classifier
generated using a particular score threshold, they can be viewed as different
operating points of the classifier. One may choose one of these operating
points depending on the requirements of the application. Hence, an ROC
curve facilitates the comparison of classifiers over a range of operating points.
For example, Figure 4.54 compares the ROC curves of two classifiers,

n+ n−

(pn+)/n+=p (pn−)/p=p


and , generated by varying the score thresholds. We can see that is
better than when FPR is less than 0.36, as shows better TPR than
for this range of operating points. On the other hand, is superior when
FPR is greater than 0.36, since the TPR of is higher than that of for
this range. Clearly, neither of the two classifiers dominates (is strictly better
than) the other, i.e., shows higher values of TPR and lower values of FPR
over all operating points.

To summarize the aggregate behavior across all operating points, one of the
commonly used measures is the area under the ROC curve (AUC). If the
classifier is perfect, then its area under the ROC curve will be equal 1. If the
algorithm simply performs random guessing, then its area under the ROC
curve will be equal to 0.5.

Although the AUC provides a useful summary of aggregate performance,
there are certain caveats in using the AUC for comparing classifiers. First,
even if the AUC of algorithm A is higher than the AUC of another algorithm B,
this does not mean that algorithm A is always better than B, i.e., the ROC
curve of A dominates that of B across all operating points. For example, even
though shows a slightly lower AUC than in Figure 4.54 , we can see
that both and are useful over different ranges of operating points and
none of them are strictly better than the other across all possible operating
points. Hence, we cannot use the AUC to determine which algorithm is better,
unless we know that the ROC curve of one of the algorithms dominates the

Second, although the AUC summarizes the aggregate performance over all
operating points, we are often interested in only a small range of operating
points in most applications. For example, even though shows slightly
lower AUC than , it shows higher TPR values than for small FPR
values (smaller than 0.36). In the presence of class imbalance, the behavior of

M2 M1
M2 M1 M2

M2 M1

M1 M2
M1 M2

M2 M2

an algorithm over small FPR values (also termed as early retrieval) is often
more meaningful for comparison than the performance over all FPR values.
This is because, in many applications, it is important to assess the TPR
achieved by a classifier in the first few instances with highest scores, without
incurring a large FPR. Hence, in Figure 4.54 , due to the high TPR values
of during early retrieval , we may prefer over for
imbalanced test sets, despite the lower AUC of . Hence, care must be
taken while comparing the AUC values of different classifiers, usually by
visualizing their ROC curves rather than just reporting their AUC.

A key characteristic of ROC curves is that they are agnostic to the skew in the
test set, because both the evaluation measures used in constructing ROC
curves (TPR and FPR) are invariant to class imbalance. Hence, ROC curves
are not suitable for measuring the impact of skew on classification
performance. In particular, we will obtain the same ROC curve for two test
data sets that have very different skew.

M1 (FPR<0.36) M1 M2

Figure 4.55.
PR curves for two different classifiers.

Precision-Recall Curve
An alternate tool for aggregate evaluation is the precision recall curve (PR
curve). The PR curve plots the precision and recall values of a classifier on
the y and x axes respectively, by varying the threshold on the test scores.
Figure 4.55 shows an example of PR curves for two hypothetical
classifiers, and . The approach for generating a PR curve is similar to
the approach described above for generating an ROC curve. However, there
are some key distinguishing features in the PR curve:

1. PR curves are sensitive to the skew factor , and different PR
curves are generated for different values of .

M1 M2


2. When the score threshold is lowest (every instance is labeled as
positive), the precision is equal to while recall is 1. As we increase
the score threshold, the number of predicted positives can stay the
same or decrease. Hence, the recall monotonically declines as the
score threshold increases. In general, the precision may increase or
decrease for the same value of recall, upon addition of an instance into
the set of predicted positives. For example, if the k ranked instance
belongs to the negative class, then including it will result in a drop in
the precision without affecting the recall. The precision may improve at
the next step, which adds the ranked instance, if this instance
belongs to the positive class. Hence, the PR curve is not a smooth,
monotonically increasing curve like the ROC curve, and generally has a
zigzag pattern. This pattern is more prominent in the left part of the
curve, where even a small change in the number of false positives can
cause a large change in precision.

3. As, as we increase the imbalance among the classes (reduce the value
of ), the rightmost points of all PR curves will move downwards. At
and near the leftmost point on the PR curve (corresponding to larger
values of score threshold), the recall is close to zero, while the
precision is equal to the fraction of positives in the top ranked instances
of the algorithm. Hence, different classifiers can have different values
of precision at the leftmost points of the PR curve. Also, if the
classification score of an algorithm monotonically varies with the
posterior probability of the positive class, we can expect the PR curve
to gradually decrease from a high value of precision on the leftmost
point to a constant value of at the rightmost point, albeit with some
ups and downs. This can be observed in the PR curve of algorithm
in Figure 4.55 , which starts from a higher value of precision on the
left that gradually decreases as we move towards the right. On the
other hand, the PR curve of algorithm starts from a lower value of
precision on the left and shows more drastic ups and downs as we







move right, suggesting that the classification score of shows a
weaker monotonic relationship with the posterior probability of the
positive class.

4. A random classifier that assigns an instance to be positive with a fixed
probability p has a precision of and a recall of p. Hence, a classifier
that performs random guessing has a horizontal PR curve with , as
shown using a dashed line in Figure 4.55 . Note that the random
baseline in PR curves depends on the skew in the test set, in contrast
to the fixed main diagonal of ROC curves that represents random

5. Note that the precision of an algorithm is impacted more strongly by
false positives in the top ranked test instances than the FPR of the
algorithm. For this reason, the PR curve generally helps to magnify the
differences between classifiers in the left portion of the PR curve.
Hence, in the presence of class imbalance in the test data, analyzing
the PR curves generally provides a deeper insight into the performance
of classifiers than the ROC curves, especially in the early retrieval
range of operating points.

6. The classifier corresponding to represents the
perfect classifier. Similar to AUC, we can also compute the area under
the PR curve of an algorithm, known as AUC-PR. The AUC-PR of a
random classifier is equal to , while that of a perfect algorithm is equal
to 1. Note that AUC-PR varies with changing skew in the test set, in
contrast to the area under the ROC curve, which is insensitive to the
skew. The AUC-PR helps in accentuating the differences between
classification algorithms in the early retrieval range of operating points.
Hence, it is more suited for evaluating classification performance in the
presence of class imbalance than the area under the ROC curve.
However, similar to ROC curves, a higher value of AUC-PR does not
guarantee the superiority of a classification algorithm over another. This
is because the PR curves of two algorithms can easily cross each



(precision=1, recall=1)


other, such that they both show better performances in different ranges
of operating points. Hence, it is important to visualize the PR curves
before comparing their AUC-PR values, in order to ensure a
meaningful evaluation.

4.12 Multiclass Problem
Some of the classification techniques described in this chapter are originally
designed for binary classification problems. Yet there are many real-world
problems, such as character recognition, face identification, and text
classification, where the input data is divided into more than two categories.
This section presents several approaches for extending the binary classifiers
to handle multiclass problems. To illustrate these approaches, let

be the set of classes of the input data.

The first approach decomposes the multiclass problem into K binary
problems. For each class , a binary problem is created where all
instances that belong to are considered positive examples, while the
remaining instances are considered negative examples. A binary classifier is
then constructed to separate instances of class from the rest of the classes.
This is known as the one-against-rest (1-r) approach.

The second approach, which is known as the one-against-one (1-1) approach,
constructs binary classifiers, where each classifier is used to
distinguish between a pair of classes, . Instances that do not belong to
either or are ignored when constructing the binary classifier for . In
both 1-r and 1-1 approaches, a test instance is classified by combining the
predictions made by the binary classifiers. A voting scheme is typically
employed to combine the predictions, where the class that receives the
highest number of votes is assigned to the test instance. In the 1-r approach,
if an instance is classified as negative, then all classes except for the positive
class receive a vote. This approach, however, may lead to ties among the
different classes. Another possibility is to transform the outputs of the binary

{y1, y2, … ,yK}



K(K −1)/2
(yi, yj)

yi yj (yi, yj)

classifiers into probability estimates and then assign the test instance to the
class that has the highest probability.

Example 4.11.
Consider a multiclass problem where . Suppose a test
instance is classified as according to the 1-r approach. In other
words, it is classified as positive when is used as the positive class and
negative when , and are used as the positive class. Using a
simple majority vote, notice that receives the highest number of votes,
which is four, while the remaining classes receive only three votes. The
test instance is therefore classified as .

Example 4.12.
Suppose the test instance is classified using the 1-1 approach as follows:

Binary pair of classes


The first two rows in this table correspond to the pair of classes
chosen to build the classifier and the last row represents the predicted
class for the test instance. After combining the predictions, and each
receive two votes, while and each receives only one vote. The test
instance is therefore classified as either or , depending on the tie-
breaking procedure.

Error-Correcting Output Coding

Y={y1, y2, y3, y4}
(+, −, −, −)

y2, y3 y4



+:y1−:y2 +:y1−:y3 +:y1−:y4 +:y2−:y3 +:y2−:y4 +:y3−:y4

+ + − + − +

(yi, yj)

y1 y4
y2 y3

y1 y4

A potential problem with the previous two approaches is that they may be
sensitive to binary classification errors. For the 1-r approach given in Example

4.12, if at least of one of the binary classifiers makes a mistake in its
prediction, then the classifier may end up declaring a tie between classes or
making a wrong prediction. For example, suppose the test instance is
classified as due to misclassification by the third classifier. In this
case, it will be difficult to tell whether the instance should be classified as or
, unless the probability associated with each class prediction is taken into


The error-correcting output coding (ECOC) method provides a more robust
way for handling multiclass problems. The method is inspired by an
information-theoretic approach for sending messages across noisy channels.

The idea behind this approach is to add redundancy into the transmitted
message by means of a codeword, so that the receiver may detect errors in
the received message and perhaps recover the original message if the
number of errors is small.

For multiclass learning, each class is represented by a unique bit string of
length n known as its codeword. We then train n binary classifiers to predict
each bit of the codeword string. The predicted class of a test instance is given
by the codeword whose Hamming distance is closest to the codeword
produced by the binary classifiers. Recall that the Hamming distance between
a pair of bit strings is given by the number of bits that differ.

Example 4.13.
Consider a multiclass problem where . Suppose we
encode the classes using the following seven bit codewords:

(+, −, +, −)



Y={y1, y2, y3, y4}

Class Codeword

1 1 1 1 1 1 1

0 0 0 0 1 1 1

0 0 1 1 0 0 1

0 1 0 1 0 1 0

Each bit of the codeword is used to train a binary classifier. If a test
instance is classified as (0,1,1,1,1,1,1) by the binary classifiers, then the
Hamming distance between the codeword and is 1, while the Hamming
distance to the remaining classes is 3. The test instance is therefore
classified as .

An interesting property of an error-correcting code is that if the minimum
Hamming distance between any pair of codewords is d, then any
errors in the output code can be corrected using its nearest codeword. In
Example 4.13 , because the minimum Hamming distance between any pair
of codewords is 4, the classifier may tolerate errors made by one of the seven
binary classifiers. If there is more than one classifier that makes a mistake,
then the classifier may not be able to compensate for the error.

An important issue is how to design the appropriate set of codewords for
different classes. From coding theory, a vast number of algorithms have been
developed for generating n-bit codewords with bounded Hamming distance.
However, the discussion of these algorithms is beyond the scope of this book.
It is worthwhile mentioning that there is a significant difference between the
design of error-correcting codes for communication tasks compared to those
used for multiclass learning. For communication, the codewords should
maximize the Hamming distance between the rows so that error correction







⌊ (d−1)/2) ⌋

can be performed. Multiclass learning, however, requires that both the row-
wise and column-wise distances of the codewords must be well separated. A
larger column-wise distance ensures that the binary classifiers are mutually
independent, which is an important requirement for ensemble learning

4.13 Bibliographic Notes
Mitchell [278] provides excellent coverage on many classification techniques
from a machine learning perspective. Extensive coverage on classification can
also be found in Aggarwal [195], Duda et al. [229], Webb [307], Fukunaga
[237], Bishop [204], Hastie et al. [249], Cherkassky and Mulier [215], Witten
and Frank [310], Hand et al. [247], Han and Kamber [244], and Dunham [230].

Direct methods for rule-based classifiers typically employ the sequential
covering scheme for inducing classification rules. Holte’s 1R [255] is the
simplest form of a rule-based classifier because its rule set contains only a
single rule. Despite its simplicity, Holte found that for some data sets that
exhibit a strong one-to-one relationship between the attributes and the class
label, 1R performs just as well as other classifiers. Other examples of rule-
based classifiers include IREP [234], RIPPER [218], CN2 [216, 217], AQ
[276], RISE [224], and ITRULE [296]. Table 4.8 shows a comparison of the
characteristics of four of these classifiers.

Table 4.8. Comparison of various rule-based classifiers.





General-to-specific General-to-


(seeded by a positive


Evaluation metric FOIL’s Info gain Laplace Entropy and

Number of true

condition forrule-

All examples
belong to the same






Rules cover only
positive class

Rule pruning Reduced error

None None None


Positive and

Positive only Positive only Positive and negative

condition for
adding rules

orbased on MDL




All positive examples
are covered

Rule setp runing Replace or modify


None None

Search strategy Greedy Beam


Beam search

For rule-based classifiers, the rule antecedent can be generalized to include
any propositional or first-order logical expression (e.g., Horn clauses).
Readers who are interested in first-order logic rule-based classifiers may refer
to references such as [278] or the vast literature on inductive logic
programming [279]. Quinlan [287] proposed the C4.5rules algorithm for
extracting classification rules from decision trees. An indirect method for
extracting rules from artificial neural networks was given by Andrews et al. in

Cover and Hart [220] presented an overview of the nearest neighbor
classification method from a Bayesian perspective. Aha provided both
theoretical and empirical evaluations for instance-based methods in [196].
PEBLS, which was developed by Cost and Salzberg [219], is a nearest
neighbor classifier that can handle data sets containing nominal attributes.


Each training example in PEBLS is also assigned a weight factor that
depends on the number of times the example helps make a correct prediction.
Han et al. [243] developed a weight-adjusted nearest neighbor algorithm, in
which the feature weights are learned using a greedy, hill-climbing
optimization algorithm. A more recent survey of k-nearest neighbor
classification is given by Steinbach and Tan [298].

Naïve Bayes classifiers have been investigated by many authors, including
Langley et al. [267], Ramoni and Sebastiani [288], Lewis [270], and Domingos
and Pazzani [227]. Although the independence assumption used in naïve
Bayes classifiers may seem rather unrealistic, the method has worked
surprisingly well for applications such as text classification. Bayesian networks
provide a more flexible approach by allowing some of the attributes to be
interdependent. An excellent tutorial on Bayesian networks is given by
Heckerman in [252] and Jensen in [258]. Bayesian networks belong to a
broader class of models known as probabilistic graphical models. A formal
introduction to the relationships between graphs and probabilities was
presented in Pearl [283]. Other great resources on probabilistic graphical
models include books by Bishop [205], and Jordan [259]. Detailed discussions
of concepts such as d-separation and Markov blankets are provided in Geiger
et al. [238] and Russell and Norvig [291].

Generalized linear models (GLM) are a rich class of regression models that
have been extensively studied in the statistical literature. They were
formulated by Nelder and Wedderburn in 1972 [280] to unify a number of
regression models such as linear regression, logistic regression, and Poisson
regression, which share some similarities in their formulations. An extensive
discussion of GLMs is provided in the book by McCullagh and Nelder [274].

Artificial neural networks (ANN) have witnessed a long and winding history of
developments, involving multiple phases of stagnation and resurgence. The

idea of a mathematical model of a neural network was first introduced in 1943
by McCulloch and Pitts [275]. This led to a series of computational machines
to simulate a neural network based on the theory of neural plasticity [289].
The perceptron, which is the simplest prototype of modern ANNs, was
developed by Rosenblatt in 1958 [290]. The perceptron uses a single layer of
processing units that can perform basic mathematical operations such as
addition and multiplication. However, the perceptron can only learn linear
decision boundaries and is guaranteed to converge only when the classes are
linearly separable. Despite the interest in learning multi-layer networks to
overcome the limitations of perceptron, progress in this area remain halted
until the invention of the backpropagation algorithm by Werbos in 1974 [309],
which allowed for the quick training of multi-layer ANNs using the gradient
descent method. This led to an upsurge of interest in the artificial intelligence
(AI) community to develop multi-layer ANN models, a trend that continued for
more than a decade. Historically, ANNs mark a paradigm shift in AI from
approaches based on expert systems (where knowledge is encoded using if-
then rules) to machine learning approaches (where the knowledge is encoded
in the parameters of a computational model). However, there were still a
number of algorithmic and computational challenges in learning large ANN
models, which remained unresolved for a long time. This hindered the
development of ANN models to the scale necessary for solving real-world
problems. Slowly, ANNs started getting outpaced by other classification
models such as support vector machines, which provided better performance
as well as theoretical guarantees of convergence and optimality. It is only
recently that the challenges in learning deep neural networks have been
circumvented, owing to better computational resources and a number of
algorithmic improvements in ANNs since 2006. This re-emergence of ANN
has been dubbed as “deep learning,” which has often outperformed existing
classification models and gained wide-spread popularity.

Deep learning is a rapidly evolving area of research with a number of
potentially impactful contributions being made every year. Some of the
landmark advancements in deep learning include the use of large-scale
restricted Boltzmann machines for learning generative models of data [201,
253], the use of autoencoders and its variants (denoising autoencoders) for
learning robust feature representations [199, 305, 306], and sophistical
architectures to promote sharing of parameters across nodes such as
convolutional neural networks for images [265, 268] and recurrent neural
networks for sequences [241, 242, 277]. Other major improvements include
the approach of unsupervised pretraining for initializing ANN models [232], the
dropout technique for regularization [254, 297], batch normalization for fast
learning of ANN parameters [256], and maxout networks for effective usage of
the dropout technique [240]. Even though the discussions in this chapter on
learning ANN models were centered around the gradient descent method,
most of the modern ANN models involving a large number of hidden layers
are trained using the stochastic gradient descent method since it is highly
scalable [207]. An extensive survey of deep learning approaches has been
presented in review articles by Bengio [200], LeCun et al. [269], and
Schmidhuber [293]. An excellent summary of deep learning approaches can
also be obtained from recent books by Goodfellow et al. [239] and Nielsen

Vapnik [303, 304] has written two authoritative books on Support Vector
Machines (SVM). Other useful resources on SVM and kernel methods include
the books by Cristianini and Shawe-Taylor [221] and Schölkopf and Smola
[294]. There are several survey articles on SVM, including those written by
Burges [212], Bennet et al. [202], Hearst [251], and Mangasarian [272]. SVM
can also be viewed as an norm regularizer of the hinge loss function, as
described in detail by Hastie et al. [249]. The norm regularizer of the
square loss function can be obtained using the least absolute shrinkage and
selection operator (Lasso), which was introduced by Tibshirani in 1996 [301].


The Lasso has several interesting properties such as the ability to
simultaneously perform feature selection as well as regularization, so that only
a subset of features are selected in the final model. An excellent review of
Lasso can be obtained from a book by Hastie et al. [250].

A survey of ensemble methods in machine learning was given by Diet-terich
[222]. The bagging method was proposed by Breiman [209]. Freund and
Schapire [236] developed the AdaBoost algorithm. Arcing, which stands for
adaptive resampling and combining, is a variant of the boosting algorithm
proposed by Breiman [210]. It uses the non-uniform weights assigned to
training examples to resample the data for building an ensemble of training
sets. Unlike AdaBoost, the votes of the base classifiers are not weighted when
determining the class label of test examples. The random forest method was
introduced by Breiman in [211]. The concept of bias-variance decomposition is
explained in detail by Hastie et al. [249]. While the bias-variance
decomposition was initially proposed for regression problems with squared
loss function, a unified framework for classification problems involving 0–1
losses was introduced by Domingos [226].

Related work on mining rare and imbalanced data sets can be found in the
survey papers written by Chawla et al. [214] and Weiss [308]. Sampling-based
methods for mining imbalanced data sets have been investigated by many
authors, such as Kubat and Matwin [266], Japkowitz [257], and Drummond
and Holte [228]. Joshi et al. [261] discussed the limitations of boosting
algorithms for rare class modeling. Other algorithms developed for mining rare
classes include SMOTE [213], PNrule [260], and CREDOS [262].

Various alternative metrics that are well-suited for class imbalanced problems
are available. The precision, recall, and -measure are widely-used metrics
in information retrieval [302]. ROC analysis was originally used in signal
detection theory for performing aggregate evaluation over a range of score


thresholds. A method for comparing classifier performance using the convex
hull of ROC curves was suggested by Provost and Fawcett in [286]. Bradley
[208] investigated the use of area under the ROC curve (AUC) as a
performance metric for machine learning algorithms. Despite the vast body of
literature on optimizing the AUC measure in machine learning models, it is
well-known that AUC suffers from certain limitations. For example, the AUC
can be used to compare the quality of two classifiers only if the ROC curve of
one classifier strictly dominates the other. However, if the ROC curves of two
classifiers intersect at any point, then it is difficult to assess the relative quality
of classifiers using the AUC measure. An in-depth discussion of the pitfalls in
using AUC as a performance measure can be obtained in works by Hand
[245, 246], and Powers [284]. The AUC has also been considered to be an
incoherent measure of performance, i.e., it uses different scales while
comparing the performance of different classifiers, although a coherent
interpretation of AUC has been provided by Ferri et al. [235]. Berrar and Flach
[203] describe some of the common caveats in using the ROC curve for
clinical microarray research. An alternate approach for measuring the
aggregate performance of a classifier is the precision-recall (PR) curve, which
is especially useful in the presence of class imbalance [292].

An excellent tutorial on cost-sensitive learning can be found in a review article
by Ling and Sheng [271]. The properties of a cost matrix had been studied by
Elkan in [231]. Margineantu and Dietterich [273] examined various methods
for incorporating cost information into the C4.5 learning algorithm, including
wrapper methods, class distribution-based methods, and loss-based methods.
Other cost-sensitive learning methods that are algorithm-independent include
AdaCost [233], MetaCost [225], and costing [312].

Extensive literature is also available on the subject of multiclass learning. This
includes the works of Hastie and Tibshirani [248], Allwein et al. [197], Kong
and Dietterich [264], and Tax and Duin [300]. The error-correcting output

coding (ECOC) method was proposed by Dietterich and Bakiri [223]. They
had also investigated techniques for designing codes that are suitable for
solving multiclass problems.

Apart from exploring algorithms for traditional classification settings where
every instance has a single set of features with a unique categorical label,
there has been a lot of recent interest in non-traditional classification
paradigms, involving complex forms of inputs and outputs. For example, the
paradigm of multi-label learning allows for an instance to be assigned multiple
class labels rather than just one. This is useful in applications such as object
recognition in images, where a photo image may include more than one
classification object, such as, grass, sky, trees, and mountains. A survey on
multi-label learning can be found in [313]. As another example, the paradigm
of multi-instance learning considers the problem where the instances are
available in the form of groups called bags, and training labels are available at
the level of bags rather than individual instances. Multi-instance learning is
useful in applications where an object can exist as multiple instances in
different states (e.g., the different isomers of a chemical compound), and even
if a single instance shows a specific characteristic, the entire bag of instances
associated with the object needs to be assigned the relevant class. A survey
on multi-instance learning is provided in [314].

In a number of real-world applications, it is often the case that the training
labels are scarce in quantity, because of the high costs associated with
obtaining gold-standard supervision. However, we almost always have
abundant access to unlabeled test instances, which do not have supervised
labels but contain valuable information about the structure or distribution of
instances. Traditional learning algorithms, which only make use of the labeled
instances in the training set for learning the decision boundary, are unable to
exploit the information contained in unlabeled instances. In contrast, learning
algorithms that make use of the structure in the unlabeled data for learning the

classification model are known as semi-supervised learning algorithms [315,
316]. The use of unlabeled data is also explored in the paradigm of multi-view
learning [299, 311], where every object is observed in multiple views of the
data, involving diverse sets of features. A common strategy used by multi-view
learning algorithms is co-training [206], where a different model is learned for
every view of the data, but the model predictions from every view are
constrained to be identical to each other on the unlabeled test instances.

Another learning paradigm that is commonly explored in the paucity of training
data is the framework of active learning, which attempts to seek the smallest
set of label annotations to learn a reasonable classification model. Active
learning expects the annotator to be involved in the process of model learning,
so that the labels are requested incrementally over the most relevant set of
instances, given a limited budget of label annotations. For example, it may be
useful to obtain labels over instances closer to the decision boundary that can
play a bigger role in fine-tuning the boundary. A review on active learning
approaches can be found in [285, 295].

In some applications, it is important to simultaneously solve multiple learning
tasks together, where some of the tasks may be similar to one another. For
example, if we are interested in translating a passage written in English into
different languages, the tasks involving lexically similar languages (such as
Spanish and Portuguese) would require similar learning of models. The
paradigm of multi-task learning helps in simultaneously learning across all
tasks while sharing the learning among related tasks. This is especially useful
when some of the tasks do not contain sufficiently many training samples, in
which case borrowing the learning from other related tasks helps in the
learning of robust models. A special case of multi-task learning is transfer
learning, where the learning from a source task (with sufficient number of
training samples) has to be transferred to a destination task (with paucity of

training data). An extensive survey of transfer learning approaches is provided
by Pan et al. [282].

Most classifiers assume every data instance must belong to a class, which is
not always true for some applications. For example, in malware detection, due
to the ease in which new malwares are created, a classifier trained on existing
classes may fail to detect new ones even if the features for the new malwares
are considerably different than those for existing malwares. Another example
is in critical applications such as medical diagnosis, where prediction errors
are costly and can have severe consequences. In this situation, it would be
better for the classifier to refrain from making any prediction on a data
instance if it is unsure of its class. This approach, known as classifier with
reject option, does not need to classify every data instance unless it
determines the prediction is reliable (e.g., if the class probability exceeds a
user-specified threshold). Instances that are unclassified can be presented to
domain experts for further determination of their true class labels.

Classifiers can also be distinguished in terms of how the classification model
is trained. A batch classifier assumes all the labeled instances are available
during training. This strategy is applicable when the training set size is not too
large and for stationary data, where the relationship between the attributes
and classes does not vary over time. An online classifier, on the other hand,
trains an initial model using a subset of the labeled data [263]. The model is
then updated incrementally as more labeled instances become available. This
strategy is effective when the training set is too large or when there is concept
drift due to changes in the distribution of the data over time.

[195] C. C. Aggarwal. Data classification: algorithms and applications. CRC

Press, 2014.

[196] D. W. Aha. A study of instance-based algorithms for supervised learning
tasks: mathematical, empirical, and psychological evaluations. PhD thesis,
University of California, Irvine, 1990.

[197] E. L. Allwein, R. E. Schapire, and Y. Singer. Reducing Multiclass to
Binary: A Unifying Approach to Margin Classifiers. Journal of Machine
Learning Research, 1: 113–141, 2000.

[198] R. Andrews, J. Diederich, and A. Tickle. A Survey and Critique of
Techniques For Extracting Rules From Trained Artificial Neural Networks.
Knowledge Based Systems, 8(6):373–389, 1995.

[199] P. Baldi. Autoencoders, unsupervised learning, and deep architectures.
ICML unsupervised and transfer learning, 27(37-50):1, 2012.

[200] Y. Bengio. Learning deep architectures for AI. Foundations and trends R
in Machine Learning, 2(1):1–127, 2009.

[201] Y. Bengio, A. Courville, and P. Vincent. Representation learning: A
review and new perspectives. IEEE transactions on pattern analysis and

machine intelligence, 35(8): 1798–1828, 2013.

[202] K. Bennett and C. Campbell. Support Vector Machines: Hype or
Hallelujah. SIGKDD Explorations, 2(2):1–13, 2000.

[203] D. Berrar and P. Flach. Caveats and pitfalls of ROC analysis in clinical
microarray research (and how to avoid them). Briefings in bioinformatics,
page bbr008, 2011.

[204] C. M. Bishop. Neural Networks for Pattern Recognition. Oxford
University Press, Oxford, U.K., 1995.

[205] C. M. Bishop. Pattern Recognition and Machine Learning. Springer,

[206] A. Blum and T. Mitchell. Combining labeled and unlabeled data with co-
training. In Proceedings of the eleventh annual conference on
Computational learning theory, pages 92–100. ACM, 1998.

[207] L. Bottou. Large-scale machine learning with stochastic gradient
descent. In Proceedings of COMPSTAT’2010, pages 177–186. Springer,

[208] A. P. Bradley. The use of the area under the ROC curve in the
Evaluation of Machine Learning Algorithms. Pattern Recognition,
30(7):1145–1149, 1997.

[209] L. Breiman. Bagging Predictors. Machine Learning, 24(2):123–140,

[210] L. Breiman. Bias, Variance, and Arcing Classifiers. Technical Report
486, University of California, Berkeley, CA, 1996.

[211] L. Breiman. Random Forests. Machine Learning, 45(1):5–32, 2001.

[212] C. J. C. Burges. A Tutorial on Support Vector Machines for Pattern
Recognition. Data Mining and Knowledge Discovery, 2(2):121–167, 1998.

[213] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. SMOTE:
Synthetic Minority Over-sampling Technique. Journal of Artificial
Intelligence Research, 16: 321–357, 2002.

[214] N. V. Chawla, N. Japkowicz, and A. Kolcz. Editorial: Special Issue on
Learning from Imbalanced Data Sets. SIGKDD Explorations, 6(1):1–6,

[215] V. Cherkassky and F. Mulier. Learning from Data: Concepts, Theory, and
Methods. Wiley Interscience, 1998.

[216] P. Clark and R. Boswell. Rule Induction with CN2: Some Recent
Improvements. In Machine Learning: Proc. of the 5th European Conf.
(EWSL-91), pages 151–163, 1991.

[217] P. Clark and T. Niblett. The CN2 Induction Algorithm. Machine Learning,
3(4): 261–283, 1989.

[218] W. W. Cohen. Fast Effective Rule Induction. In Proc. of the 12th Intl.
Conf. on Machine Learning, pages 115–123, Tahoe City, CA, July 1995.

[219] S. Cost and S. Salzberg. A Weighted Nearest Neighbor Algorithm for
Learning with Symbolic Features. Machine Learning, 10:57–78, 1993.

[220] T. M. Cover and P. E. Hart. Nearest Neighbor Pattern Classification.
Knowledge Based Systems, 8(6):373–389, 1995.

[221] N. Cristianini and J. Shawe-Taylor. An Introduction to Support Vector
Machines and Other Kernel-based Learning Methods. Cambridge
University Press, 2000.

[222] T. G. Dietterich. Ensemble Methods in Machine Learning. In First Intl.
Workshop on Multiple Classifier Systems, Cagliari, Italy, 2000.

[223] T. G. Dietterich and G. Bakiri. Solving Multiclass Learning Problems via
Error-Correcting Output Codes. Journal of Artificial Intelligence Research,
2:263–286, 1995.

[224] P. Domingos. The RISE system: Conquering without separating. In Proc.
of the 6th IEEE Intl. Conf. on Tools with Artificial Intelligence, pages 704–
707, New Orleans, LA, 1994.

[225] P. Domingos. MetaCost: A General Method for Making Classifiers Cost-
Sensitive. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data
Mining, pages 155–164, San Diego, CA, August 1999.

[226] P. Domingos. A unified bias-variance decomposition. In Proceedings of
17th International Conference on Machine Learning, pages 231–238, 2000.

[227] P. Domingos and M. Pazzani. On the Optimality of the Simple Bayesian
Classifier under Zero-One Loss. Machine Learning, 29(2-3):103–130,

[228] C. Drummond and R. C. Holte. C4.5, Class imbalance, and Cost
sensitivity: Why under-sampling beats over-sampling. In ICML’2004
Workshop on Learning from Imbalanced Data Sets II, Washington, DC,
August 2003.

[229] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John
Wiley & Sons, Inc., New York, 2nd edition, 2001.

[230] M. H. Dunham. Data Mining: Introductory and Advanced Topics.
Prentice Hall, 2006.

[231] C. Elkan. The Foundations of Cost-Sensitive Learning. In Proc. of the
17th Intl. Joint Conf. on Artificial Intelligence, pages 973–978, Seattle, WA,
August 2001.

[232] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, and S.
Bengio. Why does unsupervised pre-training help deep learning? Journal
of Machine Learning Research, 11(Feb):625–660, 2010.

[233] W. Fan, S. J. Stolfo, J. Zhang, and P. K. Chan. AdaCost:
misclassification cost-sensitive boosting. In Proc. of the 16th Intl. Conf. on
Machine Learning, pages 97–105, Bled, Slovenia, June 1999.

[234] J. Fürnkranz and G. Widmer. Incremental reduced error pruning. In
Proc. of the 11th Intl. Conf. on Machine Learning, pages 70–77, New
Brunswick, NJ, July 1994.

[235] C. Ferri, J. Hernández-Orallo, and P. A. Flach. A coherent interpretation
of AUC as a measure of aggregated classification performance. In
Proceedings of the 28th International Conference on Machine Learning
(ICML-11), pages 657–664, 2011.

[236] Y. Freund and R. E. Schapire. A decision-theoretic generalization of on-
line learning and an application to boosting. Journal of Computer and
System Sciences, 55(1): 119–139, 1997.

[237] K. Fukunaga. Introduction to Statistical Pattern Recognition. Academic
Press, New York, 1990.

[238] D. Geiger, T. S. Verma, and J. Pearl. d-separation: From theorems to
algorithms. arXiv preprint arXiv:1304.1505, 2013.

[239] I. Goodfellow, Y. Bengio, and A. Courville. Deep Learning. Book in
preparation for MIT Press, 2016.

[240] I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. C. Courville, and Y.
Bengio. Maxout networks. ICML (3), 28:1319–1327, 2013.

[241] A. Graves, M. Liwicki, S. Fernández, R. Bertolami, H. Bunke, and J.
Schmidhuber. A novel connectionist system for unconstrained handwriting
recognition. IEEE transactions on pattern analysis and machine
intelligence, 31(5):855–868, 2009.

[242] A. Graves and J. Schmidhuber. Offline handwriting recognition with
multidimensional recurrent neural networks. In Advances in neural
information processing systems, pages 545–552, 2009.

[243] E.-H. Han, G. Karypis, and V. Kumar. Text Categorization Using Weight
Adjusted k-Nearest Neighbor Classification. In Proc. of the 5th Pacific-Asia
Conf. on Knowledge Discovery and Data Mining, Lyon, France, 2001.

[244] J. Han and M. Kamber. Data Mining: Concepts and Techniques. Morgan
Kaufmann Publishers, San Francisco, 2001.

[245] D. J. Hand. Measuring classifier performance: a coherent alternative to
the area under the ROC curve. Machine learning, 77(1):103–123, 2009.

[246] D. J. Hand. Evaluating diagnostic tests: the area under the ROC curve
and the balance of errors. Statistics in medicine, 29(14):1502–1510, 2010.

[247] D. J. Hand, H. Mannila, and P. Smyth. Principles of Data Mining. MIT
Press, 2001.

[248] T. Hastie and R. Tibshirani. Classification by pairwise coupling. Annals
of Statistics, 26(2):451–471, 1998.

[249] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical
Learning: Data Mining, Inference, and Prediction. Springer, 2nd edition,

[250] T. Hastie, R. Tibshirani, and M. Wainwright. Statistical learning with
sparsity: the lasso and generalizations. CRC Press, 2015.

[251] M. Hearst. Trends & Controversies: Support Vector Machines. IEEE
Intelligent Systems, 13(4):18–28, 1998.

[252] D. Heckerman. Bayesian Networks for Data Mining. Data Mining and
Knowledge Discovery, 1(1):79–119, 1997.

[253] G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of
data with neural networks. Science, 313(5786):504–507, 2006.

[254] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R.
Salakhutdinov. Improving neural networks by preventing co-adaptation of
feature detectors. arXiv preprint arXiv:1207.0580, 2012.

[255] R. C. Holte. Very Simple Classification Rules Perform Well on Most
Commonly Used Data sets. Machine Learning, 11:63–91, 1993.

[256] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep
network training by reducing internal covariate shift. arXiv preprint
arXiv:1502.03167, 2015.

[257] N. Japkowicz. The Class Imbalance Problem: Significance and
Strategies. In Proc. of the 2000 Intl. Conf. on Artificial Intelligence: Special
Track on Inductive Learning, volume 1, pages 111–117, Las Vegas, NV,
June 2000.

[258] F. V. Jensen. An introduction to Bayesian networks, volume 210. UCL
press London, 1996.

[259] M. I. Jordan. Learning in graphical models, volume 89. Springer Science
& Business Media, 1998.

[260] M. V. Joshi, R. C. Agarwal, and V. Kumar. Mining Needles in a Haystack:
Classifying Rare Classes via Two-Phase Rule Induction. In Proc. of 2001
ACM-SIGMOD Intl. Conf. on Management of Data, pages 91–102, Santa
Barbara, CA, June 2001.

[261] M. V. Joshi, R. C. Agarwal, and V. Kumar. Predicting rare classes: can
boosting make any weak learner strong? In Proc. of the 8th Intl. Conf. on
Knowledge Discovery and Data Mining, pages 297–306, Edmonton,
Canada, July 2002.

[262] M. V. Joshi and V. Kumar. CREDOS: Classification Using Ripple Down
Structure (A Case for Rare Classes). In Proc. of the SIAM Intl. Conf. on
Data Mining, pages 321–332, Orlando, FL, April 2004.

[263] J. Kivinen, A. J. Smola, and R. C. Williamson. Online learning with
kernels. IEEE transactions on signal processing, 52(8):2165–2176, 2004.

[264] E. B. Kong and T. G. Dietterich. Error-Correcting Output Coding Corrects
Bias and Variance. In Proc. of the 12th Intl. Conf. on Machine Learning,
pages 313–321, Tahoe City, CA, July 1995.

[265] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification
with deep convolutional neural networks. In Advances in neural information
processing systems, pages 1097–1105, 2012.

[266] M. Kubat and S. Matwin. Addressing the Curse of Imbalanced Training
Sets: One Sided Selection. In Proc. of the 14th Intl. Conf. on Machine
Learning, pages 179–186, Nashville, TN, July 1997.

[267] P. Langley, W. Iba, and K. Thompson. An analysis of Bayesian
classifiers. In Proc. of the 10th National Conf. on Artificial Intelligence,
pages 223–228, 1992.

[268] Y. LeCun and Y. Bengio. Convolutional networks for images, speech,
and time series. The handbook of brain theory and neural networks,
3361(10):1995, 1995.

[269] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature,
521(7553):436–444, 2015.

[270] D. D. Lewis. Naive Bayes at Forty: The Independence Assumption in
Information Retrieval. In Proc. of the 10th European Conf. on Machine
Learning (ECML 1998), pages 4–15, 1998.

[271] C. X. Ling and V. S. Sheng. Cost-sensitive learning. In Encyclopedia of
Machine Learning, pages 231–235. Springer, 2011.

[272] O. Mangasarian. Data Mining via Support Vector Machines. Technical
Report Technical Report 01-05, Data Mining Institute, May 2001.

[273] D. D. Margineantu and T. G. Dietterich. Learning Decision Trees for
Loss Minimization in Multi-Class Problems. Technical Report 99-30-03,
Oregon State University, 1999.

[274] P. McCullagh and J. A. Nelder. Generalized linear models, volume 37.
CRC press, 1989.

[275] W. S. McCulloch and W. Pitts. A logical calculus of the ideas immanent
in nervous activity. The bulletin of mathematical biophysics, 5(4):115–133,

[276] R. S. Michalski, I. Mozetic, J. Hong, and N. Lavrac. The Multi-Purpose
Incremental Learning System AQ15 and Its Testing Application to Three

Medical Domains. In Proc. of 5th National Conf. on Artificial Intelligence,
Orlando, August 1986.

[277] T. Mikolov, M. Karafiát, L. Burget, J. Cernock`y, and S. Khudanpur.
Recurrent neural network based language model. In Interspeech, volume
2, page 3, 2010.

[278] T. Mitchell. Machine Learning. McGraw-Hill, Boston, MA, 1997.

[279] S. Muggleton. Foundations of Inductive Logic Programming. Prentice
Hall, Englewood Cliffs, NJ, 1995.

[280] J. A. Nelder and R. J. Baker. Generalized linear models. Encyclopedia of
statistical sciences, 1972.

[281] M. A. Nielsen. Neural networks and deep learning. Published online:
http: // neuralnetworksanddeeplearning. com/ .( visited: 10. 15. 2016) ,

[282] S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions
on knowledge and data engineering, 22(10):1345–1359, 2010.

[283] J. Pearl. Probabilistic reasoning in intelligent systems: networks of
plausible inference. Morgan Kaufmann, 2014.

[284] D. M. Powers. The problem of area under the curve. In 2012 IEEE
International Conference on Information Science and Technology, pages

567–573. IEEE, 2012.

[285] M. Prince. Does active learning work? A review of the research. Journal
of engineering education, 93(3):223–231, 2004.

[286] F. J. Provost and T. Fawcett. Analysis and Visualization of Classifier
Performance: Comparison under Imprecise Class and Cost Distributions.
In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining,
pages 43–48, Newport Beach, CA, August 1997.

[287] J. R. Quinlan. C4.5: Programs for Machine Learning. Morgan-Kaufmann
Publishers, San Mateo, CA, 1993.

[288] M. Ramoni and P. Sebastiani. Robust Bayes classifiers. Artificial
Intelligence, 125: 209–226, 2001.

[289] N. Rochester, J. Holland, L. Haibt, and W. Duda. Tests on a cell
assembly theory of the action of the brain, using a large digital computer.
IRE Transactions on information Theory, 2(3):80–93, 1956.

[290] F. Rosenblatt. The perceptron: a probabilistic model for information
storage and organization in the brain. Psychological review, 65(6):386,

[291] S. J. Russell, P. Norvig, J. F. Canny, J. M. Malik, and D. D. Edwards.
Artificial intelligence: a modern approach, volume 2. Prentice hall Upper
Saddle River, 2003.

[292] T. Saito and M. Rehmsmeier. The precision-recall plot is more
informative than the ROC plot when evaluating binary classifiers on
imbalanced datasets. PloS one, 10(3): e0118432, 2015.

[293] J. Schmidhuber. Deep learning in neural networks: An overview. Neural
Networks, 61:85–117, 2015.

[294] B. Schölkopf and A. J. Smola. Learning with Kernels: Support Vector
Machines, Regularization, Optimization, and Beyond. MIT Press, 2001.

[295] B. Settles. Active learning literature survey. University of Wisconsin,
Madison, 52 (55-66):11, 2010.

[296] P. Smyth and R. M. Goodman. An Information Theoretic Approach to
Rule Induction from Databases. IEEE Trans. on Knowledge and Data
Engineering, 4(4):301–316, 1992.

[297] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R.
Salakhutdinov. Dropout: a simple way to prevent neural networks from
overfitting. Journal of Machine Learning Research, 15(1):1929–1958,

[298] M. Steinbach and P.-N. Tan. kNN: k-Nearest Neighbors. In X. Wu and V.
Kumar, editors, The Top Ten Algorithms in Data Mining. Chapman and
Hall/CRC Reference, 1st edition, 2009.

[299] S. Sun. A survey of multi-view machine learning. Neural Computing and
Applications, 23(7-8):2031–2038, 2013.

[300] D. M. J. Tax and R. P. W. Duin. Using Two-Class Classifiers for
Multiclass Classification. In Proc. of the 16th Intl. Conf. on Pattern
Recognition (ICPR 2002), pages 124–127, Quebec, Canada, August 2002.

[301] R. Tibshirani. Regression shrinkage and selection via the lasso. Journal
of the Royal Statistical Society. Series B (Methodological), pages 267–288,

[302] C. J. van Rijsbergen. Information Retrieval. Butterworth-Heinemann,
Newton, MA, 1978.

[303] V. Vapnik. The Nature of Statistical Learning Theory. Springer Verlag,
New York, 1995.

[304] V. Vapnik. Statistical Learning Theory. John Wiley & Sons, New York,

[305] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and
composing robust features with denoising autoencoders. In Proceedings of
the 25th international conference on Machine learning, pages 1096–1103.
ACM, 2008.

[306] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol.
Stacked denoising autoencoders: Learning useful representations in a

deep network with a local denoising criterion. Journal of Machine Learning
Research, 11(Dec):3371–3408, 2010.

[307] A. R. Webb. Statistical Pattern Recognition. John Wiley & Sons, 2nd
edition, 2002.

[308] G. M. Weiss. Mining with Rarity: A Unifying Framework. SIGKDD
Explorations, 6 (1):7–19, 2004.

[309] P. Werbos. Beyond regression: new fools for prediction and analysis in
the behavioral sciences. PhD thesis, Harvard University, 1974.

[310] I. H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools
and Techniques with Java Implementations. Morgan Kaufmann, 1999.

[311] C. Xu, D. Tao, and C. Xu. A survey on multi-view learning. arXiv preprint
arXiv:1304.5634, 2013.

[312] B. Zadrozny, J. C. Langford, and N. Abe. Cost-Sensitive Learning by
Cost-Proportionate Example Weighting. In Proc. of the 2003 IEEE Intl.
Conf. on Data Mining, pages 435–442, Melbourne, FL, August 2003.

[313] M.-L. Zhang and Z.-H. Zhou. A review on multi-label learning algorithms.
IEEE transactions on knowledge and data engineering, 26(8):1819–1837,

[314] Z.-H. Zhou. Multi-instance learning: A survey. Department of Computer
Science & Technology, Nanjing University, Tech. Rep, 2004.

[315] X. Zhu. Semi-supervised learning. In Encyclopedia of machine learning,
pages 892–897. Springer, 2011.

[316] X. Zhu and A. B. Goldberg. Introduction to semi-supervised learning.
Synthesis lectures on artificial intelligence and machine learning, 3(1):1–
130, 2009.

4.14 Exercises
1. Consider a binary classification problem with the following set of attributes
and attribute values:

Suppose a rule-based classifier produces the following rule set:

a. Are the rules mutually exclusive?

b. Is the rule set exhaustive?

c. Is ordering needed for this set of rules?

d. Do you need a default class for the rule set?

2. The RIPPER algorithm (by Cohen [218]) is an extension of an earlier
algorithm called IREP (by Fürnkranz and Widmer [234]). Both algorithms
apply the reduced-error pruning method to determine whether a rule needs
to be pruned. The reduced error pruning method uses a validation set to
estimate the generalization error of a classifier. Consider the following pair of

Air Conditioner={Working, Broken}

Engine={Good, Bad}

Mileage={High, Medium, Low}

Rust={Yes, No}

Mileage=High→Mileage=HighMileage=Low→Value=HighAir Conditioner=Working


is obtained by adding a new conjunct, B, to the left-hand side of . For
this question, you will be asked to determine whether is preferred over
from the perspectives of rule-growing and rule-pruning. To determine whether
a rule should be pruned, IREP computes the following measure:

where P is the total number of positive examples in the validation set, N is the
total number of negative examples in the validation set, p is the number of
positive examples in the validation set covered by the rule, and n is the
number of negative examples in the validation set covered by the rule.
is actually similar to classification accuracy for the validation set. IREP favors
rules that have higher values of . On the other hand, RIPPER applies
the following measure to determine whether a rule should be pruned:

a. Suppose is covered by 350 positive examples and 150 negative
examples, while is covered by 300 positive examples and 50 negative
examples. Compute the FOIL’s information gain for the rule with
respect to .

b. Consider a validation set that contains 500 positive examples and 500
negative examples. For , suppose the number of positive examples
covered by the rule is 200, and the number of negative examples covered
bytheruleis50. For , suppose the number of positive examples covered
by the rule is 100 and the number of negative examples is 5. Compute

for both rules. Which rule does IREP prefer?

c. Compute for the previous problem. Which rule does RIPPER

R2 R1
R2 R1











3. C4.5rules is an implementation of an indirect method for generating rules
from a decision tree. RIPPER is an implementation of a direct method for
generating rules directly from data.

a. Discuss the strengths and weaknesses of both methods.

b. Consider a data set that has a large difference in the class size (i.e.,
some classes are much bigger than others). Which method (between
C4.5rules and RIPPER) is better in terms of finding high accuracy rules
for the small classes?

4. Consider a training set that contains 100 positive examples and 400
negative examples. For each of the following candidate rules,

determine which is the best and worst candidate rule according to:

a. Rule accuracy.

b. FOIL’s information gain.

c. The likelihood ratio statistic.

d. The Laplace measure.

e. The m-estimate measure (with and ).

5. Figure 4.3 illustrates the coverage of the classification rules R1, R2, and
R3. Determine which is the best and worst rule according to:

a. The likelihood ratio statistic.

b. The Laplace measure.

R1:A→+(covers 4 positive and 1 negative examples),R2:B→+
(covers 30 positive and 10 negative examples),R3:C→+
(covers 100 positive and 90 negative examples),

k=2 p+=0.2

c. The m-estimate measure (with and ).

d. The rule accuracy after R1 has been discovered, where none of the
examples covered by R1 are discarded.

e. The rule accuracy after R1 has been discovered, where only the positive
examples covered by R1 are discarded.

f. The rule accuracy after R1 has been discovered, where both positive and
negative examples covered by R1 are discarded.


a. Suppose the fraction of undergraduate students who smoke is 15% and
the fraction of graduate students who smoke is 23%. If one-fifth of the
college students are graduate students and the rest are undergraduates,
what is the probability that a student who smokes is a graduate student?

b. Given the information in part (a), is a randomly chosen college student
more likely to be a graduate or undergraduate student?

c. Repeat part (b) assuming that the student is a smoker.

d. Suppose 30% of the graduate students live in a dorm but only 10% of the
undergraduate students live in a dorm. If a student smokes and lives in
the dorm, is he or she more likely to be a graduate or undergraduate
student? You can assume independence between students who live in a
dorm and those who smoke.

7. Consider the data set shown in Table 4.9

Table 4.9. Data set for Exercise 7.

Instance A B C Class

1 0 0 0

k=2 p+=0.58


2 0 0 1

3 0 1 1

4 0 1 1

5 0 0 1

6 1 0 1

7 1 0 1

8 1 0 1

9 1 1 1

10 1 0 1

a. Estimate the conditional probabilities for
, and .

b. Use the estimate of conditional probabilities given in the previous
question to predict the class label for a test sample
using the naïve Bayes approach.

c. Estimate the conditional probabilities using the m-estimate approach, with
and .

d. Repeat part (b) using the conditional probabilities given in part (c).

e. Compare the two methods for estimating probabilities. Which method is
better and why?

8. Consider the data set shown in Table 4.10 .

Table 4.10. Data set for Exercise 8.





P(A|+), P(B|+), P(C|+), P(A|
−), P(B|−) P(C|−)

(A=0, B=1, C=0)

p=1/2 m=4

Instance A B C Class

1 0 0 1

2 1 0 1

3 0 1 0

4 1 0 0

5 1 0 1

6 0 0 1

7 1 1 0

8 0 0 0

9 0 1 0

10 1 1 1 +

a. Estimate the conditional probabilities for
, and using

the same approach as in the previous problem.

b. Use the conditional probabilities in part (a) to predict the class label for a
test sample using the naïve Bayes approach.

c. Compare , and . State the relationships
between A and B.

d. Repeat the analysis in part (c) using , and .

e. Compare against and
. Are the variables conditionally independent given the






P(A=1|+), P(B=1|+), P(C=1|+), P(A=1|−), P(B=1|−) P(C=1|−)

(A=1, B=1, C=1)

P(A=1), P(B=1) P(A=1, B=1)

P(A=1), P(B=0) P(A=1, B=0)

P(A=1, B=1|Class=+) P(A=1|Class=+)


a. Explain how naïve Bayes performs on the data set shown in Figure
4.56 .

b. If each class is further divided such that there are four classes (A1, A2,
B1, and B2), will naïve Bayes perform better?

c. How will a decision tree perform on this data set (for the two-class
problem)? What if there are four classes?

10. Figure 4.57 illustrates the Bayesian network for the data set shown in
Table 4.11 . (Assume that all the attributes are binary).

a. Draw the probability table for each node in the network.

b. Use the Bayesian network to compute

11. Given the Bayesian network shown in Figure 4.58 , compute the
following probabilities:

P(Engine=Bad, Air Conditioner=Broken)

Figure 4.56.
Data set for Exercise 9.

Figure 4.57.
Bayesian network.

a. .P(B=good,F=empty, G=empty, S=yes)

b. .

c. Given that the battery is bad, compute the probability that the car will

12. Consider the one-dimensional data set shown in Table 4.12 .

a. Classify the data point according to its 1-, 3-, 5-, and 9-nearest
neighbors (using majority vote).

b. Repeat the previous analysis using the distance-weighted voting
approach described in Section 4.3.1 .

Table 4.11. Data set for Exercise 10.

Mileage Engine Air

Number of Instances

Number of Instances

Hi Good Working 3 4

Hi Good Broken 1 2

Hi Bad Working 1 5

Hi Bad Broken 0 4

Lo Good Working 9 0

Lo Good Broken 5 1

Lo Bad Working 1 2

Lo Bad Broken 0 2

P(B=bad,F=empty, G=not empty, S=no)


Car Value=Hi Car Value=Lo

Figure 4.58.
Bayesian network for Exercise 11.

13. The nearest neighbor algorithm described in Section 4.3 can be
extended to handle nominal attributes. A variant of the algorithm called
PEBLS (Parallel Exemplar-Based Learning System) by Cost and Salzberg
[219] measures the distance between two values of a nominal attribute using
the modified value difference metric (MVDM). Given a pair of nominal attribute
values, and , the distance between them is defined as follows:

where is the number of examples from class i with attribute value and
is the number of examples with attribute value

Table 4.12. Data set for Exercise 12.

x 0.5 3.0 4.5 4.6 4.9 5.2 5.3 5.5 7.0 9.5

V1 V2

d(V1, V2)=∑i=1k| ni1n1−ni2n2, | (4.108)

nij Vj nj


Consider the training set for the loan classification problem shown in Figure
4.8 . Use the MVDM measure to compute the distance between every pair
of attribute values for the and attributes.

14. For each of the Boolean functions given below, state whether the problem
is linearly separable.



c. (A OR B) AND (A OR C)

d. (A XOR B) AND (A OR B)


a. Demonstrate how the perceptron model can be used to represent the
AND and OR functions between a pair of Boolean variables.

b. Comment on the disadvantage of using linear functions as activation
functions for multi-layer neural networks.

16. You are asked to evaluate the performance of two classification models,
and . The test set you have chosen contains 26 binary attributes,

labeled as A through Z. Table 4.13 shows the posterior probabilities
obtained by applying the models to the test set. (Only the posterior
probabilities for the positive class are shown). As this is a two-class problem,

and . Assume that we are mostly
interested in detecting instances from the positive class.

a. Plot the ROC curve for both and . (You should plot them on the
same graph.) Which model do you think is better? Explain your reasons.

− − + + + − − + − −

M1 M2

P(−)=1−P(+) P(−|A, …, Z)=1−P(+|A, …, Z)

M1 M2

b. For model , suppose you choose the cutoff threshold to be . In
other words, any test instances whose posterior probability is greater than
t will be classified as a positive example. Compute the precision, recall,
and F-measure for the model at this threshold value.

c. Repeat the analysis for part (b) using the same cutoff threshold on model
. Compare the F-measure results for both models. Which model is

better? Are the results consistent with what you expect from the ROC

d. Repeat part (b) for model using the threshold . Which threshold
do you prefer, or ? Are the results consistent with what you
expect from the ROC curve?

Table 4.13. Posterior probabilities for Exercise 16.

Instance True Class

1 0.73 0.61

2 0.69 0.03

3 0.44 0.68

4 0.55 0.31

5 0.67 0.45

6 0.47 0.09

7 0.08 0.38

8 0.15 0.05

9 0.45 0.01

10 0.35 0.04

M1 t=0.5


M1 t=0.1
t=0.5 t=0.1

P(+|A, …, Z, M1) P(+|A, …, Z, M2)






17. Following is a data set that contains two attributes, X and Y , and two
class labels, “ ” and “ ”. Each attribute can take three different values: 0, 1, or

X Y Number of Instances

0 0 0 100

1 0 0 0

2 0 0 100

0 1 10 100

1 1 10 0

2 1 10 100

0 2 0 100

1 2 0 0

2 2 0 100

The concept for the “ ” class is and the concept for the “ ” class is

a. Build a decision tree on the data set. Does the tree capture the “ ” and
“ ” concepts?

b. What are the accuracy, precision, recall, and -measure of the decision
tree? (Note that precision, recall, and -measure are defined with

+ −

+ −

+ Y=1 −



respect to the “ ” class.)

c. Build a new decision tree with the following cost function:

(Hint: only the leaves of the old decision tree need to be changed.) Does
the decision tree capture the “ ” concept?

d. What are the accuracy, precision, recall, and -measure of the new
decision tree?

18. Consider the task of building a classifier from random data, where the
attribute values are generated randomly irrespective of the class labels.
Assume the data set contains instances from two classes, “ ” and “ .” Half of
the data set is used for training while the remaining half is used for testing.

a. Suppose there are an equal number of positive and negative instances in
the data and the decision tree classifier predicts every test instance to be
positive. What is the expected error rate of the classifier on the test data?

b. Repeat the previous analysis assuming that the classifier predicts each
test instance to be positive class with probability 0.8 and negative class
with probability 0.2.

c. Suppose two-thirds of the data belong to the positive class and the
remaining one-third belong to the negative class. What is the expected
error of a classifier that predicts every test instance to be positive?

d. Repeat the previous analysis assuming that the classifier predicts each
test instance to be positive class with probability 2/3 and negative class
with probability 1/3.


C(i, j)={ 0,if i=j;1,if i=+, j=−;Number of
− instancesNumber of+ instancesif i=−, j=+;



+ −

19. Derive the dual Lagrangian for the linear SVM with non-separable data
where the objective function is

20. Consider the XOR problem where there are four training points:

Transform the data into the following feature space:

Find the maximum margin linear decision boundary in the transformed space.

21. Given the data sets shown in Figures 4.59 , explain how the decision
tree, naïve Bayes, and k-nearest neighbor classifiers would perform on these
data sets.

f(w)=ǁ w ǁ22+C(∑i=1Nξi)2.

(1, 1, −), (1, 0, +), (0, 1, +), (0, 0, −).

φ=(1, 2×1, 2×2, 2x1x2, x12, x22).

Figure 4.59.
Data set for Exercise 21.

5 Association Analysis: Basic
Concepts and Algorithms

Many business enterprises accumulate large quantities
of data from their day-to-day operations. For example,
huge amounts of customer purchase data are collected
daily at the checkout counters of grocery stores. Table
5.1 gives an example of such data, commonly known
as market basket transactions. Each row in this table
corresponds to a transaction, which contains a unique
identifier labeled TID and a set of items bought by a
given customer. Retailers are interested in analyzing
the data to learn about the purchasing behavior of their
customers. Such valuable information can be used to
support a variety of business-related applications such
as marketing promotions, inventory management, and
customer relationship management.

Table 5.1. An example of market basket

TID Items

1 {Bread, Milk}

2 {Bread, Diapers, Beer, Eggs}

3 {Milk, Diapers, Beer, Cola}

4 {Bread, Milk, Diapers, Beer}

5 {Bread, Milk, Diapers, Cola}

This chapter presents a methodology known as
association analysis, which is useful for discovering
interesting relationships hidden in large data sets. The
uncovered relationships can be represented in the form
of sets of items present in many transactions, which are
known as frequent itemsets, or association rules,
that represent relationships between two itemsets. For
example, the following rule can be extracted from the
data set shown in Table 5.1 :

The rule suggests a relationship between the sale of
diapers and beer because many customers who buy
diapers also buy beer. Retailers can use these types of
rules to help them identify new opportunities for cross-
selling their products to the customers.


Besides market basket data, association analysis is
also applicable to data from other application domains
such as bioinformatics, medical diagnosis, web mining,
and scientific data analysis. In the analysis of Earth
science data, for example, association patterns may
reveal interesting connections among the ocean, land,
and atmospheric processes. Such information may help
Earth scientists develop a better understanding of how
the different elements of the Earth system interact with
each other. Even though the techniques presented here
are generally applicable to a wider variety of data sets,
for illustrative purposes, our discussion will focus
mainly on market basket data.

There are two key issues that need to be addressed
when applying association analysis to market basket
data. First, discovering patterns from a large transaction
data set can be computationally expensive. Second,
some of the discovered patterns may be spurious
(happen simply by chance) and even for non-spurious
patterns, some are more interesting than others. The
remainder of this chapter is organized around these two
issues. The first part of the chapter is devoted to
explaining the basic concepts of association analysis
and the algorithms used to efficiently mine such
patterns. The second part of the chapter deals with the
issue of evaluating the discovered patterns in order to
help prevent the generation of spurious results and to
rank the patterns in terms of some interestingness

5.1 Preliminaries
This section reviews the basic terminology used in association analysis and
presents a formal description of the task.

Binary Representation

Market basket data can be represented in a binary format as shown in Table
5.2 , where each row corresponds to a transaction and each column
corresponds to an item. An item can be treated as a binary variable whose
value is one if the item is present in a transaction and zero otherwise.
Because the presence of an item in a transaction is often considered more
important than its absence, an item is an asymmetric binary variable. This
representation is a simplistic view of real market basket data because it
ignores important aspects of the data such as the quantity of items sold or the
price paid to purchase them. Methods for handling such non-binary data will
be explained in Chapter 6 .

Table 5.2. A binary 0/1 representation of market basket data.

TID Bread Milk Diapers Beer Eggs Cola

1 1 1 0 0 0 0

2 1 0 1 1 1 0

3 0 1 1 1 0 1

4 1 1 1 1 0 0

5 1 1 1 0 0 1

Itemset and Support Count

Let be the set of all items in a market basket data and
be the set of all transactions. Each transaction, contains a

subset of items chosen from I. In association analysis, a collection of zero or
more items is termed an itemset. If an itemset contains k items, it is called a k-
itemset. For instance, { , , } is an example of a 3-itemset. The
null (or empty) set is an itemset that does not contain any items.

A transaction is said to contain an itemset X if X is a subset of . For
example, the second transaction shown in Table 5.2 contains the itemset
{ , } but not { , }. An important property of an itemset is
its support count, which refers to the number of transactions that contain a
particular itemset. Mathematically, the support count, , for an itemset X
can be stated as follows:

where the symbol denotes the number of elements in a set. In the data set
shown in Table 5.2 , the support count for { , , } is equal to
two because there are only two transactions that contain all three items.

Often, the property of interest is the support, which is fraction of transactions
in which an itemset occurs:

An itemset X is called frequent if s(X) is greater than some user-defined
threshold, minsup.

I={i1, i2, … , id} T=
{t1, t2, …, tN} ti

tj tj


σ(X)=|{ti|X⊆ti, ti∈T}|,



Association Rule

An association rule is an implication expression of the form , where X
and Y are disjoint itemsets, i.e., . The strength of an association rule
can be measured in terms of its support and confidence. Support
determines how often a rule is applicable to a given data set, while confidence
determines how frequently items in Y appear in transactions that contain X.
The formal definitions of these metrics are

Example 5.1.
Consider the rule Because the support count for
{ , , } is 2 and the total number of transactions is 5, the
rule’s support is . The rule’s confidence is obtained by dividing the
support count for { , , } by the support count for { ,

}. Since there are 3 transactions that contain milk and diapers, the
confidence for this rule is .

Why Use Support and Confidence?

Support is an important measure because a rule that has very low support
might occur simply by chance. Also, from a business perspective a low
support rule is unlikely to be interesting because it might not be profitable to
promote items that customers seldom buy together (with the exception of the
situation described in Section 5.8 ). For these reasons, we are interested in
finding rules whose support is greater than some user-defined threshold. As


Support, s(X→Y)=σ(X∪Y)N; (5.1)

Confidence, c(X→Y)=σ(X∪Y)σ(X). (5.2)



will be shown in Section 5.2.1 , support also has a desirable property that
can be exploited for the efficient discovery of association rules.

Confidence, on the other hand, measures the reliability of the inference made
by a rule. For a given rule , the higher the confidence, the more likely it is
for Y to be present in transactions that contain X. Confidence also provides an
estimate of the conditional probability of Y given X.

Association analysis results should be interpreted with caution. The inference
made by an association rule does not necessarily imply causality. Instead, it
can sometimes suggest a strong co-occurrence relationship between items in
the antecedent and consequent of the rule. Causality, on the other hand,
requires knowledge about which attributes in the data capture cause and
effect, and typically involves relationships occurring over time (e.g.,
greenhouse gas emissions lead to global warming). See Section 5.7.1 for
additional discussion.

Formulation of the Association Rule Mining Problem

The association rule mining problem can be formally stated as follows:

Definition 5.1. (Association Rule
Given a set of transactions T , find all the rules having

and , where minsup and
minconf are the corresponding support and confidence


support ≥ minsup confidence ≥ minconf

A brute-force approach for mining association rules is to compute the support
and confidence for every possible rule. This approach is prohibitively
expensive because there are exponentially many rules that can be extracted
from a data set. More specifically, assuming that neither the left nor the right-
hand side of the rule is an empty set, the total number of possible rules, R,
extracted from a data set that contains d items is

The proof for this equation is left as an exercise to the readers (see Exercise 5
on page 440). Even for the small data set shown in Table 5.1 , this
approach requires us to compute the support and confidence for

rules. More than 80% of the rules are discarded after applying
and , thus wasting most of the computations. To

avoid performing needless computations, it would be useful to prune the rules
early without having to compute their support and confidence values.

An initial step toward improving the performance of association rule mining
algorithms is to decouple the support and confidence requirements. From
Equation 5.1 , notice that the support of a rule is the same as the
support of its corresponding itemset, . For example, the following rules
have identical support because they involve items from the same itemset,

{ , , }:

R=3d−2d+1+1. (5.3)

minsup=20% mincof=50%


{Beer, Diapers}→{Milk},{Beer, Milk}→{Diapers},{Diapers, Milk}→{Beer},{Beer}
→{Diapers, Milk},{Milk}→{Beer, Diapers},{Diapers}→{Beer, Milk}.

If the itemset is infrequent, then all six candidate rules can be pruned
immediately without our having to compute their confidence values.

Therefore, a common strategy adopted by many association rule mining
algorithms is to decompose the problem into two major subtasks:

1. Frequent Itemset Generation, whose objective is to find all the
itemsets that satisfy the minsup threshold.

2. Rule Generation, whose objective is to extract all the high confidence
rules from the frequent itemsets found in the previous step. These rules
are called strong rules.

The computational requirements for frequent itemset generation are generally
more expensive than those of rule generation. Efficient techniques for
generating frequent itemsets and association rules are discussed in Sections
5.2 and 5.3 , respectively.

5.2 Frequent Itemset Generation
A lattice structure can be used to enumerate the list of all possible itemsets.
Figure 5.1 shows an itemset lattice for . In general, a data
set that contains k items can potentially generate up to frequent
itemsets, excluding the null set. Because k can be very large in many practical
applications, the search space of itemsets that need to be explored is
exponentially large.

Figure 5.1.

I={a, b, c, d, e}

An itemset lattice.

A brute-force approach for finding frequent itemsets is to determine the
support count for every candidate itemset in the lattice structure. To do this,
we need to compare each candidate against every transaction, an operation
that is shown in Figure 5.2 . If the candidate is contained in a transaction,
its support count will be incremented. For example, the support for { ,

} is incremented three times because the itemset is contained in
transactions 1, 4, and 5. Such an approach can be very expensive because it
requires O(NMw) comparisons, where N is the number of transactions,
is the number of candidate itemsets, and w is the maximum transaction

width. Transaction width is the number of items present in a transaction.

Figure 5.2.
Counting the support of candidate itemsets.

There are three main approaches for reducing the computational complexity
of frequent itemset generation.

1. Reduce the number of candidate itemsets (M). The Apriori principle,
described in the next Section, is an effective way to eliminate some of


the candidate itemsets without counting their support values.
2. Reduce the number of comparisons. Instead of matching each

candidate itemset against every transaction, we can reduce the number
of comparisons by using more advanced data structures, either to store
the candidate itemsets or to compress the data set. We will discuss
these strategies in Sections 5.2.4 and 5.6 , respectively.

3. Reduce the number of transactions (N). As the size of candidate
itemsets increases, fewer transactions will be supported by the
itemsets. For instance, since the width of the first transaction in Table
5.1 is 2, it would be advantageous to remove this transaction before
searching for frequent itemsets of size 3 and larger. Algorithms that
employ such a strategy are discussed in the Bibliographic Notes.

5.2.1 The Apriori Principle

This Section describes how the support measure can be used to reduce the
number of candidate itemsets explored during frequent itemset generation.
The use of support for pruning candidate itemsets is guided by the following

Theorem 5.1 (Apriori Principle).
If an itemset is frequent, then all of its subsets must also be

To illustrate the idea behind the Apriori principle, consider the itemset lattice
shown in Figure 5.3 . Suppose {c, d, e} is a frequent itemset. Clearly, any
transaction that contains {c, d, e} must also contain its subsets, {c, d}, {c, e},
{d, e}, {c}, {d}, and {e}. As a result, if {c, d, e} is frequent, then all subsets of {c,
d, e} (i.e., the shaded itemsets in this figure) must also be frequent.

Figure 5.3.
An illustration of the Apriori principle. If {c, d, e} is frequent, then all subsets of
this itemset are frequent.

Conversely, if an itemset such as {a, b} is infrequent, then all of its supersets
must be infrequent too. As illustrated in Figure 5.4 , the entire subgraph

containing the supersets of {a, b} can be pruned immediately once {a, b} is
found to be infrequent. This strategy of trimming the exponential search space
based on the support measure is known as support-based pruning. Such a
pruning strategy is made possible by a key property of the support measure,
namely, that the support for an itemset never exceeds the support for its
subsets. This property is also known as the anti-monotone property of the
support measure.

Figure 5.4.
An illustration of support-based pruning. If {a, b} is infrequent, then all
supersets of {a, b} are infrequent.

Definition 5.2. (Anti-monotone Property.)
A measure f possesses the anti-monotone property if for every
itemset X that is a proper subset of itemset Y, i.e. , we have


More generally, a large number of measures—see Section 5.7.1 —can be
applied to itemsets to evaluate various properties of itemsets. As will be
shown in the next Section, any measure that has the anti-monotone property
can be incorporated directly into an itemset mining algorithm to effectively
prune the exponential search space of candidate itemsets.

5.2.2 Frequent Itemset Generation in
the Apriori Algorithm

Apriori is the first association rule mining algorithm that pioneered the use of
support-based pruning to systematically control the exponential growth of
candidate itemsets. Figure 5.5 provides a high-level illustration of the
frequent itemset generation part of the Apriori algorithm for the transactions
shown in Table 5.1 . We assume that the support threshold is 60%, which is
equivalent to a minimum support count equal to 3.


Figure 5.5.
Illustration of frequent itemset generation using the Apriori algorithm.

Initially, every item is considered as a candidate 1-itemset. After counting their
supports, the candidate itemsets { } and { } are discarded because
they appear in fewer than three transactions. In the next iteration, candidate 2-
itemsets are generated using only the frequent 1-itemsets because the Apriori
principle ensures that all supersets of the infrequent 1-itemsets must be
infrequent. Because there are only four frequent 1-itemsets, the number of
candidate 2-itemsets generated by the algorithm is . Two of these six
candidates, { , } and { , }, are subsequently found to be
infrequent after computing their support values. The remaining four
candidates are frequent, and thus will be used to generate candidate 3-
itemsets. Without support-based pruning, there are candidate 3-
itemsets that can be formed using the six items given in this example. With



the Apriori principle, we only need to keep candidate 3-itemsets whose
subsets are frequent. The only candidate that has this property is { ,

, }. However, even though the subsets of { , , }
are frequent, the itemset itself is not.

The effectiveness of the Apriori pruning strategy can be shown by counting
the number of candidate itemsets generated. A brute-force strategy of
enumerating all itemsets (up to size 3) as candidates will produce

candidates. With the Apriori principle, this number decreases to

candidates, which represents a 68% reduction in the number of candidate
itemsets even in this simple example.

The pseudocode for the frequent itemset generation part of the Apriori
algorithm is shown in Algorithm 5.1 . Let denote the set of candidate k-
itemsets and denote the set of frequent k-itemsets:

The algorithm initially makes a single pass over the data set to determine
the support of each item. Upon completion of this step, the set of all
frequent 1-itemsets, , will be known (steps 1 and 2).
Next, the algorithm will iteratively generate new candidate k-itemsets and
prune unnecessary candidates that are guaranteed to be infrequent given
the frequent -itemsets found in the previous iteration (steps 5 and 6).
Candidate generation and pruning is implemented using the functions
candidate-gen and candidate-prune, which are described in Section
5.2.3 .






To count the support of the candidates, the algorithm needs to make an
additional pass over the data set (steps 7–12). The subset function is used
to determine all the candidate itemsets in that are contained in each
transaction t. The implementation of this function is described in Section
5.2.4 .
After counting their supports, the algorithm eliminates all candidate
itemsets whose support counts are less than (step 13).
The algorithm terminates when there are no new frequent itemsets
generated, i.e., (step 14).

The frequent itemset generation part of the Apriori algorithm has two
important characteristics. First, it is a level-wise algorithm; i.e., it traverses the
itemset lattice one level at a time, from frequent 1-itemsets to the maximum
size of frequent itemsets. Second, it employs a generate-and-test strategy for
finding frequent itemsets. At each iteration (level), new candidate itemsets are
generated from the frequent itemsets found in the previous iteration. The
support for each candidate is then counted and tested against the minsup
threshold. The total number of iterations needed by the algorithm is ,
where is the maximum size of the frequent itemsets.

5.2.3 Candidate Generation and

The candidate-gen and candidate-prune functions shown in Steps 5 and 6 of
Algorithm 5.1 generate candidate itemsets and prunes unnecessary ones
by performing the following two operations, respectively:





1. Candidate Generation. This operation generates new candidate k-
itemsets based on the frequent -itemsets found in the previous

Algorithm 5.1 Frequent itemset generation of
the Apriori algorithm.

∈ ∧

∈ ∧

2. Candidate Pruning. This operation eliminates some of the candidate
k-itemsets using support-based pruning, i.e. by removing k-itemsets
whose subsets are known to be infrequent in previous iterations. Note


that this pruning is done without computing the actual support of these
k-itemsets (which could have required comparing them against each

Candidate Generation
In principle, there are many ways to generate candidate itemsets. An effective
candidate generation procedure must be complete and non-redundant. A
candidate generation procedure is said to be complete if it does not omit any
frequent itemsets. To ensure completeness, the set of candidate itemsets
must subsume the set of all frequent itemsets, i.e., . A candidate
generation procedure is non-redundant if it does not generate the same
candidate itemset more than once. For example, the candidate itemset {a, b,
c, d} can be generated in many ways—by merging {a, b, c} with {d}, {b, d} with
{a, c}, {c} with {a, b, d}, etc. Generation of duplicate candidates leads to
wasted computations and thus should be avoided for efficiency reasons. Also,
an effective candidate generation procedure should avoid generating too
many unnecessary candidates. A candidate itemset is unnecessary if at least
one of its subsets is infrequent, and thus, eliminated in the candidate pruning

Next, we will briefly describe several candidate generation procedures,
including the one used by the candidate-gen function.

Brute-Force Method

The brute-force method considers every k-itemset as a potential candidate
and then applies the candidate pruning step to remove any unnecessary
candidates whose subsets are infrequent (see Figure 5.6 ). The number of
candidate itemsets generated at level k is equal to , where d is the total
number of items. Although candidate generation is rather trivial, candidate



pruning becomes extremely expensive because a large number of itemsets
must be examined.

Figure 5.6.
A brute-force method for generating candidate 3-itemsets.


An alternative method for candidate generation is to extend each frequent
-itemset with frequent items that are not part of the -itemset. Figure

5.7 illustrates how a frequent 2-itemset such as { , } can be
augmented with a frequent item such as to produce a candidate 3-
itemset { , , }.


−1) (k−1)

Figure 5.7.
Generating and pruning candidate k-itemsets by merging a frequent –
itemset with a frequent item. Note that some of the candidates are
unnecessary because their subsets are infrequent.

The procedure is complete because every frequent k-itemset is composed of
a frequent -itemset and a frequent 1-itemset. Therefore, all frequent k-
itemsets are part of the candidate k-itemsets generated by this procedure.
Figure 5.7 shows that the candidate generation method only
produces four candidate 3-itemsets, instead of the

itemsets produced by the brute-force method. The method
generates lower number of candidates because every candidate is
guaranteed to contain at least one frequent -itemset. While this
procedure is a substantial improvement over the brute-force method, it can
still produce a large number of unnecessary candidates, as the remaining
subsets of a candidate itemset can still be infrequent.

Note that the approach discussed above does not prevent the same candidate




(63)=20 Fk−1×F1


itemset from being generated more than once. For instance, { , ,
} can be generated by merging { , } with { }, { ,
} with { }, or { , } with { }. One way to avoid

generating duplicate candidates is by ensuring that the items in each frequent
itemset are kept sorted in their lexicographic order. For example, itemsets
such as { , }, { , , }, and { , } follow
lexicographic order as the items within every itemset are arranged
alphabetically. Each frequent -itemset X is then extended with frequent
items that are lexicographically larger than the items in X. For example, the
itemset { , } can be augmented with { } because Milk is
lexicographically larger than Bread and Diapers. However, we should not
augment { , } with { } nor { , } with { }
because they violate the lexicographic ordering condition. Every candidate k-
itemset is thus generated exactly once, by merging the lexicographically
largest item with the remaining items in the itemset. If the
method is used in conjunction with lexicographic ordering, then only two
candidate 3-itemsets will be produced in the example illustrated in Figure
5.7 . { , , } and { , , } will not be generated
because { , } is not a frequent 2-itemset.


This candidate generation procedure, which is used in the candidate-gen
function of the Apriori algorithm, merges a pair of frequent -itemsets only
if their first items, arranged in lexicographic order, are identical. Let

and be a pair of frequent –
itemsets, arranged lexicographically. A and B are merged if they satisfy the
following conditions:


k−1 Fk−1×F1


k−2 A=

{a1, a2, …, ak−1} B={b1, b2, …, bk−1} (k−1)

ai=bi (for i=1, 2, …, k−2).

Note that in this case, because A and B are two distinct itemsets.
The candidate k-itemset generated by merging A and B consists of the first
common items followed by and in lexicographic order. This

candidate generation procedure is complete, because for every
lexicographically ordered frequent k-itemset, there exists two lexicographically
ordered frequent -itemsets that have identical items in the first

In Figure 5.8 , the frequent itemsets { , } and { , } are
merged to form a candidate 3-itemset { , , }. The algorithm
does not have to merge { , } with { , } because the first
item in both itemsets is different. Indeed, if { , , } is a viable
candidate, it would have been obtained by merging { , } with
{ , } instead. This example illustrates both the completeness of the
candidate generation procedure and the advantages of using lexicographic
ordering to prevent duplicate candidates. Also, if we order the frequent –
itemsets according to their lexicographic rank, itemsets with identical first
items would take consecutive ranks. As a result, the candidate
generation method would consider merging a frequent itemset only with ones
that occupy the next few ranks in the sorted list, thus saving some


−2 ak−1 bk−1

(k−1) k−2



Figure 5.8.
Generating and pruning candidate k-itemsets by merging pairs of frequent


Figure 5.8 shows that the candidate generation procedure
results in only one candidate 3-itemset. This is a considerable reduction from
the four candidate 3-itemsets generated by the method. This is
because the method ensures that every candidate k-itemset
contains at least two frequent -itemsets, thus greatly reducing the
number of candidates that are generated in this step.

Note that there can be multiple ways of merging two frequent -itemsets
in the procedure, one of which is merging if their first items
are identical. An alternate approach could be to merge two frequent –
itemsets A and B if the last items of A are identical to the first
itemsets of B. For example, { , } and { , } could be
merged using this approach to generate the candidate 3-itemset { ,

, }. As we will see later, this alternate procedure is





Fk−1×Fk−1 k−2

k−2 k−2


useful in generating sequential patterns, which will be discussed in Chapter
6 .

Candidate Pruning
To illustrate the candidate pruning operation for a candidate k-itemset,

, consider its k proper subsets, . If any of
them are infrequent, then X is immediately pruned by using the Apriori
principle. Note that we don’t need to explicitly ensure that all subsets of X of
size less than are frequent (see Exercise 7). This approach greatly
reduces the number of candidate itemsets considered during support
counting. For the brute-force candidate generation method, candidate pruning
requires checking only k subsets of size for each candidate k-itemset.
However, since the candidate generation strategy ensures that at
least one of the -size subsets of every candidate k-itemset is frequent,
we only need to check for the remaining subsets. Likewise, the
strategy requires examining only subsets of every candidate k-itemset,

since two of its -size subsets are already known to be frequent in the
candidate generation step.

5.2.4 Support Counting

Support counting is the process of determining the frequency of occurrence
for every candidate itemset that survives the candidate pruning step. Support
counting is implemented in steps 6 through 11 of Algorithm 5.1 . A brute-
force approach for doing this is to compare each transaction against every
candidate itemset (see Figure 5.2 ) and to update the support counts of
candidates contained in a transaction. This approach is computationally

{i1, i2, …, ik} X−{ij}(∀j=1, 2, …, k)



k−1 Fk−1×Fk

−1 k−2

expensive, especially when the numbers of transactions and candidate
itemsets are large.

An alternative approach is to enumerate the itemsets contained in each
transaction and use them to update the support counts of their respective
candidate itemsets. To illustrate, consider a transaction t that contains five
items, {1, 2, 3, 5, 6}. There are itemsets of size 3 contained in this
transaction. Some of the itemsets may correspond to the candidate 3-itemsets
under investigation, in which case, their support counts are incremented.
Other subsets of t that do not correspond to any candidates can be ignored.

Figure 5.9 shows a systematic way for enumerating the 3-itemsets
contained in t. Assuming that each itemset keeps its items in increasing
lexicographic order, an itemset can be enumerated by specifying the smallest
item first, followed by the larger items. For instance, given , all
the 3-itemsets contained in t must begin with item 1, 2, or 3. It is not possible
to construct a 3-itemset that begins with items 5 or 6 because there are only
two items in t whose labels are greater than or equal to 5. The number of
ways to specify the first item of a 3-itemset contained in t is illustrated by the
Level 1 prefix tree structure depicted in Figure 5.9 . For instance, 1
represents a 3-itemset that begins with item 1, followed by two more items
chosen from the set {2, 3, 5, 6}.


t={1, 2, 3, 5, 6}

2 3 5 6

Figure 5.9.
Enumerating subsets of three items from a transaction t.

After fixing the first item, the prefix tree structure at Level 2 represents the
number of ways to select the second item. For example, 1 2
corresponds to itemsets that begin with the prefix (1 2) and are followed by
the items 3, 5, or 6. Finally, the prefix tree structure at Level 3 represents the
complete set of 3-itemsets contained in t. For example, the 3-itemsets that
begin with prefix {1 2} are {1, 2, 3}, {1, 2, 5}, and {1, 2, 6}, while those that
begin with prefix {2 3} are {2, 3, 5} and {2, 3, 6}.

The prefix tree structure shown in Figure 5.9 demonstrates how itemsets
contained in a transaction can be systematically enumerated, i.e., by
specifying their items one by one, from the leftmost item to the rightmost item.
We still have to determine whether each enumerated 3-itemset corresponds

3 5 6

to an existing candidate itemset. If it matches one of the candidates, then the
support count of the corresponding candidate is incremented. In the next
Section, we illustrate how this matching operation can be performed efficiently
using a hash tree structure.

Support Counting Using a Hash Tree*
In the Apriori algorithm, candidate itemsets are partitioned into different
buckets and stored in a hash tree. During support counting, itemsets
contained in each transaction are also hashed into their appropriate buckets.
That way, instead of comparing each itemset in the transaction with every
candidate itemset, it is matched only against candidate itemsets that belong to
the same bucket, as shown in Figure 5.10 .

Figure 5.10.
Counting the support of itemsets using hash structure.

Figure 5.11 shows an example of a hash tree structure. Each internal node
of the tree uses the following hash function, , where modeh(p)=(p−1) mod 3,

refers to the modulo (remainder) operator, to determine which branch of the
current node should be followed next. For example, items 1, 4, and 7 are
hashed to the same branch (i.e., the leftmost branch) because they have the
same remainder after dividing the number by 3. All candidate itemsets are
stored at the leaf nodes of the hash tree. The hash tree shown in Figure
5.11 contains 15 candidate 3-itemsets, distributed across 9 leaf nodes.

Figure 5.11.
Hashing a transaction at the root node of a hash tree.

Consider the transaction, . To update the support counts of
the candidate itemsets, the hash tree must be traversed in such a way that all

t={1, 2, 3, 4, 5, 6}

the leaf nodes containing candidate 3-itemsets belonging to t must be visited
at least once. Recall that the 3-itemsets contained in t must begin with items
1, 2, or 3, as indicated by the Level 1 prefix tree structure shown in Figure
5.9 . Therefore, at the root node of the hash tree, the items 1, 2, and 3 of the
transaction are hashed separately. Item 1 is hashed to the left child of the root
node, item 2 is hashed to the middle child, and item 3 is hashed to the right
child. At the next level of the tree, the transaction is hashed on the second
item listed in the Level 2 tree structure shown in Figure 5.9 . For example,
after hashing on item 1 at the root node, items 2, 3, and 5 of the transaction
are hashed. Based on the hash function, items 2 and 5 are hashed to the
middle child, while item 3 is hashed to the right child, as shown in Figure
5.12 . This process continues until the leaf nodes of the hash tree are
reached. The candidate itemsets stored at the visited leaf nodes are
compared against the transaction. If a candidate is a subset of the transaction,
its support count is incremented. Note that not all the leaf nodes are visited
while traversing the hash tree, which helps in reducing the computational cost.
In this example, 5 out of the 9 leaf nodes are visited and 9 out of the 15
itemsets are compared against the transaction.

Figure 5.12.
Subset operation on the leftmost subtree of the root of a candidate hash tree.

5.2.5 Computational Complexity

The computational complexity of the Apriori algorithm, which includes both its
runtime and storage, can be affected by the following factors.

Support Threshold

Lowering the support threshold often results in more itemsets being declared
as frequent. This has an adverse effect on the computational complexity of the
algorithm because more candidate itemsets must be generated and counted

at every level, as shown in Figure 5.13 . The maximum size of frequent
itemsets also tends to increase with lower support thresholds. This increases
the total number of iterations to be performed by the Apriori algorithm, further
increasing the computational cost.

Figure 5.13.
Effect of support threshold on the number of candidate and frequent itemsets
obtained from a benchmark data set.

Number of Items (Dimensionality)

As the number of items increases, more space will be needed to store the
support counts of items. If the number of frequent items also grows with the
dimensionality of the data, the runtime and storage requirements will increase
because of the larger number of candidate itemsets generated by the

Number of Transactions

Because the Apriori algorithm makes repeated passes over the transaction
data set, its run time increases with a larger number of transactions.

Average Transaction Width

For dense data sets, the average transaction width can be very large. This
affects the complexity of the Apriori algorithm in two ways. First, the maximum
size of frequent itemsets tends to increase as the average transaction width
increases. As a result, more candidate itemsets must be examined during
candidate generation and support counting, as illustrated in Figure 5.14 .
Second, as the transaction width increases, more itemsets are contained in
the transaction. This will increase the number of hash tree traversals
performed during support counting.

A detailed analysis of the time complexity for the Apriori algorithm is presented

Figure 5.14.

Effect of average transaction width on the number of candidate and frequent
itemsets obtained from a synthetic data set.

Generation of frequent 1-itemsets

For each transaction, we need to update the support count for every item
present in the transaction. Assuming that w is the average transaction width,
this operation requires O(Nw) time, where N is the total number of

Candidate generation

To generate candidate k-itemsets, pairs of frequent -itemsets are merged
to determine whether they have at least items in common. Each merging
operation requires at most equality comparisons. Every merging step can
produce at most one viable candidate k-itemset, while in the worst-case, the
algorithm must try to merge every pair of frequent -itemsets found in the
previous iteration. Therefore, the overall cost of merging frequent itemsets is

where w is the maximum transaction width. A hash tree is also constructed
during candidate generation to store the candidate itemsets. Because the
maximum depth of the tree is k, the cost for populating the hash tree with
candidate itemsets is . During candidate pruning, we need to
verify that the subsets of every candidate k-itemset are frequent. Since
the cost for looking up a candidate in a hash tree is O(k), the candidate
pruning step requires time.

Support counting




∑k=2w(k−2)|Ck|<Cost of merging<∑k=2w(k−2)|Fk−1|2,



Each transaction of width produces itemsets of size k. This is also the
effective number of hash tree traversals performed for each transaction. The
cost for support counting is , where w is the maximum
transaction width and is the cost for updating the support count of a
candidate k-itemset in the hash tree.

|t| (|t|k)


5.3 Rule Generation
This Section describes how to extract association rules efficiently from a given
frequent itemset. Each frequent k-itemset, Y, can produce up to
association rules, ignoring rules that have empty antecedents or consequents

or ). An association rule can be extracted by partitioning the
itemset Y into two non-empty subsets, X and , such that satisfies
the confidence threshold. Note that all such rules must have already met the
support threshold because they are generated from a frequent itemset.

Example 5.2.
Let be a frequent itemset. There are six candidate association
rules that can be generated from

, and . As each of their support is identical to
the support for X, all the rules satisfy the support threshold.

Computing the confidence of an association rule does not require additional
scans of the transaction data set. Consider the rule , which is
generated from the frequent itemset . The confidence for this rule is

. Because {1, 2, 3} is frequent, the anti-monotone property
of support ensures that {1, 2} must be frequent, too. Since the support counts
for both itemsets were already found during frequent itemset generation, there
is no need to read the entire data set again.

5.3.1 Confidence-Based Pruning


∅→Y Y→∅

X={a, b, c}
X:{a, b}→{c}, {a, c}→{b}, {b, c}→{a}, {a}

→{b, c}, {b}→{a, c} {c}→{a, b}

{1, 2}→{3}
X={1, 2, 3}

σ{(1, 2, 3})/σ({1, 2})

Confidence does not show the anti-monotone property in the same way as the
support measure. For example, the confidence for can be larger,
smaller, or equal to the confidence for another rule , where and

(see Exercise 3 on page 439). Nevertheless, if we compare rules
generated from the same frequent itemset Y, the following theorem holds for
the confidence measure.

Theorem 5.2.
Let Y be an itemset and X is a subset of Y. If a rule
does not satisfy the confidence threshold, then any rule

, where is a subset of X, must not satisfy the confidence
threshold as well.

To prove this theorem, consider the following two rules: and
, where . The confidence of the rules are and ,

respectively. Since is a subset of X, . Therefore, the former rule
cannot have a higher confidence than the latter rule.

5.3.2 Rule Generation in Apriori

The Apriori algorithm uses a level-wise approach for generating association
rules, where each level corresponds to the number of items that belong to the

X˜→Y˜ X˜⊆X



−X˜ X˜

X˜→Y−X˜ X→Y
−X X˜⊂X σ(Y)/σ(X˜) σ(Y)/σ(X)

X˜ σ(X˜)/σ(X)

rule consequent. Initially, all the high confidence rules that have only one item
in the rule consequent are extracted. These rules are then used to generate
new candidate rules. For example, if and are high
confidence rules, then the candidate rule is generated by merging
the consequents of both rules. Figure 5.15 shows a lattice structure for the
association rules generated from the frequent itemset {a, b, c, d}. If any node
in the lattice has low confidence, then according to Theorem 5.2 , the entire
subgraph spanned by the node can be pruned immediately. Suppose the
confidence for is low. All the rules containing item a in its
consequent, including , and can
be discarded.

Figure 5.15.
Pruning of association rules using the confidence measure.

{acd}→{b} {abd}→{c}

{cd}→{ab}, {bd}→{ac}, {bc}→{ad} {d}→{abc}

A pseudocode for the rule generation step is shown in Algorithms 5.2 and
5.3 . Note the similarity between the procedure given in
Algorithm 5.3 and the frequent itemset generation procedure given in
Algorithm 5.1 . The only difference is that, in rule generation, we do not
have to make additional passes over the data set to compute the confidence
of the candidate rules. Instead, we determine the confidence of each rule by
using the support counts computed during frequent itemset generation.

Algorithm 5.2 Rule generation of the Apriori

Algorithm 5.3 Procedure ap-genrules .

(fk, Hm)

5.3.3 An Example: Congressional
Voting Records

This Section demonstrates the results of applying association analysis to the
voting records of members of the United States House of Representatives.
The data is obtained from the 1984 Congressional Voting Records Database,
which is available at the UCI machine learning data repository. Each
transaction contains information about the party affiliation for a representative
along with his or her voting record on 16 key issues. There are 435
transactions and 34 items in the data set. The set of items are listed in Table
5.3 .

Table 5.3. List of binary attributes from the 1984 United States
Congressional Voting Records. Source: The UCI machine learning

1. Republican
2. Democrat

water project cost sharing=yes
water project cost sharing=no



The Apriori algorithm is then applied to the data set with and
. Some of the high confidence rules extracted by the algorithm

are shown in Table 5.4 . The first two rules suggest that most of the
members who voted yes for aid to El Salvador and no for budget resolution
and MX missile are Republicans; while those who voted no for aid to El
Salvador and yes for budget resolution and MX missile are Democrats. These

physician fee freeze=yes
physician fee freeze=no
aid to El Salvador=yes
aid to El Salvador=no
religious groups in schools=yes
religious groups in schools=no
anti-satellite test ban=yes
anti-satellite test ban=no
aid to Nicaragua=yes
aid to Nicaragua=no
synfuel corporation cutback=yes
synfuel corporation cutback=no
education spending=yes
education spending=no
export administration act=yes
export administration act=no


high confidence rules show the key issues that divide members from both
political parties.

Table 5.4. Association rules extracted from the 1984 United States
Congressional Voting Records.

Association Rule Confidence





{budget resolution=no, MX-missile=no, aid to El Salvador=yes }→{Republican}

{budget resolution=yes, MX-missile=yes, aid to El Salvador=no }→{Democrat}

{crime=yes, right-to-sue=yes, physician fee freeze=yes }→{Republican}

{crime=no, right-to-sue=no, physician fee freeze=no }→{Democrat}

5.4 Compact Representation of
Frequent Itemsets
In practice, the number of frequent itemsets produced from a transaction data
set can be very large. It is useful to identify a small representative set of
frequent itemsets from which all other frequent itemsets can be derived. Two
such representations are presented in this Section in the form of maximal and
closed frequent itemsets.

5.4.1 Maximal Frequent Itemsets

Definition 5.3. (Maximal Frequent Itemset.)
A frequent itemset is maximal if none of its immediate supersets
are frequent.

To illustrate this concept, consider the itemset lattice shown in Figure 5.16 .
The itemsets in the lattice are divided into two groups: those that are frequent
and those that are infrequent. A frequent itemset border, which is represented
by a dashed line, is also illustrated in the diagram. Every itemset located
above the border is frequent, while those located below the border (the
shaded nodes) are infrequent. Among the itemsets residing near the border,

{a, d}, {a, c, e}, and {b, c, d, e} are maximal frequent itemsets because all of
their immediate supersets are infrequent. For example, the itemset {a, d} is
maximal frequent because all of its immediate supersets, {a, b, d}, {a, c, d},
and {a, d, e}, are infrequent. In contrast, {a, c} is non-maximal because one of
its immediate supersets, {a, c, e}, is frequent.

Figure 5.16.
Maximal frequent itemset.

Maximal frequent itemsets effectively provide a compact representation of
frequent itemsets. In other words, they form the smallest set of itemsets from
which all frequent itemsets can be derived. For example, every frequent
itemset in Figure 5.16 is a subset of one of the three maximal frequent

itemsets, {a, d}, {a, c, e}, and {b, c, d, e}. If an itemset is not a proper subset of
any of the maximal frequent itemsets, then it is either infrequent (e.g., {a, d,
e}) or maximal frequent itself (e.g., {b, c, d, e}). Hence, the maximal frequent
itemsets {a, c, e}, {a, d}, and {b, c, d, e} provide a compact representation of
the frequent itemsets shown in Figure 5.16 . Enumerating all the subsets of
maximal frequent itemsets generates the complete list of all frequent itemsets.

Maximal frequent itemsets provide a valuable representation for data sets that
can produce very long, frequent itemsets, as there are exponentially many
frequent itemsets in such data. Nevertheless, this approach is practical only if
an efficient algorithm exists to explicitly find the maximal frequent itemsets.
We briefly describe one such approach in Section 5.5 .

Despite providing a compact representation, maximal frequent itemsets do not
contain the support information of their subsets. For example, the support of
the maximal frequent itemsets {a, c, e}, {a, d}, and {b, c, d, e} do not provide
any information about the support of their subsets except that it meets the
support threshold. An additional pass over the data set is therefore needed to
determine the support counts of the non-maximal frequent itemsets. In some
cases, it is desirable to have a minimal representation of itemsets that
preserves the support information. We describe such a representation in the
next Section.

5.4.2 Closed Itemsets

Closed itemsets provide a minimal representation of all itemsets without losing
their support information. A formal definition of a closed itemset is presented

Definition 5.4. (Closed Itemset.)
An itemset X is closed if none of its immediate supersets has
exactly the same support count as X.

Put another way, X is not closed if at least one of its immediate supersets has
the same support count as X. Examples of closed itemsets are shown in
Figure 5.17 . To better illustrate the support count of each itemset, we have
associated each node (itemset) in the lattice with a list of its corresponding
transaction IDs. For example, since the node {b, c} is associated with
transaction IDs 1, 2, and 3, its support count is equal to three. From the
transactions given in this diagram, notice that the support for {b} is identical to
{b, c}.This is because every transaction that contains b also contains c.
Hence, {b} is not a closed itemset. Similarly, since c occurs in every
transaction that contains both a and d, the itemset {a, d} is not closed as it has
the same support as its superset {a, c, d}. On the other hand, {b, c} is a closed
itemset because it does not have the same support count as any of its

Figure 5.17.
An example of the closed frequent itemsets (with minimum support equal to

An interesting property of closed itemsets is that if we know their support
counts, we can derive the support count of every other itemset in the itemset
lattice without making additional passes over the data set. For example,
consider the 2-itemset {b, e} in Figure 5.17 . Since {b, e} is not closed, its
support must be equal to the support of one of its immediate supersets, {a, b,
e}, {b, c, e}, and {b, d, e}. Further, none of the supersets of {b, e} can have a
support greater than the support of {b, e}, due to the anti-monotone nature of
the support measure. Hence, the support of {b, e} can be computed by
examining the support counts of all of its immediate supersets of size three

and taking their maximum value. If an immediate superset is closed (e.g., {b,
c, e}), we would know its support count. Otherwise, we can recursively
compute its support by examining the supports of its immediate supersets of
size four. In general, the support count of any non-closed -itemset can be
determined as long as we know the support counts of all k-itemsets. Hence,
one can devise an iterative algorithm that computes the support counts of
itemsets at level using the support counts of itemsets at level k, starting
from the level , where is the size of the largest closed itemset.

Even though closed itemsets provide a compact representation of the support
counts of all itemsets, they can still be exponentially large in number.
Moreover, for most practical applications, we only need to determine the
support count of all frequent itemsets. In this regard, closed frequent item-sets
provide a compact representation of the support counts of all frequent
itemsets, which can be defined as follows.

Definition 5.5. (Closed Frequent Itemset.)
An itemset is a closed frequent itemset if it is closed and its
support is greater than or equal to minsup.

In the previous example, assuming that the support threshold is 40%, {b, c} is
a closed frequent itemset because its support is 60%. In Figure 5.17 , the
closed frequent itemsets are indicated by the shaded nodes.

Algorithms are available to explicitly extract closed frequent itemsets from a
given data set. Interested readers may refer to the Bibliographic Notes at the


kmax kmax

end of this chapter for further discussions of these algorithms. We can use
closed frequent itemsets to determine the support counts for all non-closed
frequent itemsets. For example, consider the frequent itemset {a, d} shown in
Figure 5.17 . Because this itemset is not closed, its support count must be
equal to the maximum support count of its immediate supersets, {a, b, d}, {a,
c, d}, and {a, d, e}. Also, since {a, d} is frequent, we only need to consider the
support of its frequent supersets. In general, the support count of every non-
closed frequent k-itemset can be obtained by considering the support of all its
frequent supersets of size . For example, since the only frequent superset
of {a, d} is {a, c, d}, its support is equal to the support of {a, c, d}, which is 2.
Using this methodology, an algorithm can be developed to compute the
support for every frequent itemset. The pseudocode for this algorithm is
shown in Algorithm 5.4 . The algorithm proceeds in a specific-to-general
fashion, i.e., from the largest to the smallest frequent itemsets. This is
because, in order to find the support for a non-closed frequent itemset, the
support for all of its supersets must be known. Note that the set of all frequent
itemsets can be easily computed by taking the union of all subsets of frequent
closed itemsets.

Algorithm 5.4 Support counting using closed
frequent itemsets.


⋅ ′⋅ ′∈ ⊂ ′

To illustrate the advantage of using closed frequent itemsets, consider the
data set shown in Table 5.5 , which contains ten transactions and fifteen
items. The items can be divided into three groups: (1) Group A, which
contains items through ; (2) Group B, which contains items through
; and (3) Group C, which contains items through . Assuming that the

support threshold is 20%, itemsets involving items from the same group are
frequent, but itemsets involving items from different groups are infrequent.
The total number of frequent itemsets is thus . However, there
are only four closed frequent itemsets in the data:

and . It is
often sufficient to present only the closed frequent itemsets to the analysts
instead of the entire set of frequent itemsets.

Table 5.5. A transaction data set for mining closed itemsets.


1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0

2 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0

3 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0

4 0 0 1 1 0 1 1 1 1 1 0 0 0 0 0

5 0 0 0 0 0 1 1 1 1 1 0 0 0 0 0

a1 a5 b1
b5 c1 c5


({a3, a4}, {a1, a2, a3, a4, a5}, {b1,b2,b3,b4,b5}, {c1, c2, c3, c4, c5})

a1 a2 a3 a4 a5 b1 b2 b3 b4 b5 c1 c2 c3 c4 c5

6 0 0 0 0 0 1 1 1 1 1 0 0 0 0 0

7 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1

8 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1

9 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1

10 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1

Finally, note that all maximal frequent itemsets are closed because none of
the maximal frequent itemsets can have the same support count as their
immediate supersets. The relationships among frequent, closed, closed
frequent, and maximal frequent itemsets are shown in Figure 5.18 .

Figure 5.18.
Relationships among frequent, closed, closed frequent, and maximal frequent

5.5 Alternative Methods for Generating
Frequent Itemsets*
Apriori is one of the earliest algorithms to have successfully addressed the
combinatorial explosion of frequent itemset generation. It achieves this by
applying the Apriori principle to prune the exponential search space. Despite
its significant performance improvement, the algorithm still incurs considerable
I/O overhead since it requires making several passes over the transaction
data set. In addition, as noted in Section 5.2.5 , the performance of the
Apriori algorithm may degrade significantly for dense data sets because of the
increasing width of transactions. Several alternative methods have been
developed to overcome these limitations and improve upon the efficiency of
the Apriori algorithm. The following is a high-level description of these

Traversal of Itemset Lattice

A search for frequent itemsets can be conceptually viewed as a traversal on
the itemset lattice shown in Figure 5.1 . The search strategy employed by
an algorithm dictates how the lattice structure is traversed during the frequent
itemset generation process. Some search strategies are better than others,
depending on the configuration of frequent itemsets in the lattice. An overview
of these strategies is presented next.

General-to-Specific versus Specific-to-General: The Apriori algorithm
uses a general-to-specific search strategy, where pairs of frequent –
itemsets are merged to obtain candidate k-itemsets. This general-to-


specific search strategy is effective, provided the maximum length of a
frequent itemset is not too long. The configuration of frequent itemsets that
works best with this strategy is shown in Figure 5.19(a) , where the
darker nodes represent infrequent itemsets. Alternatively, a specificto-
general search strategy looks for more specific frequent itemsets first,
before finding the more general frequent itemsets. This strategy is useful
to discover maximal frequent itemsets in dense transactions, where the
frequent itemset border is located near the bottom of the lattice, as shown
in Figure 5.19(b) . The Apriori principle can be applied to prune all
subsets of maximal frequent itemsets. Specifically, if a candidate k-itemset
is maximal frequent, we do not have to examine any of its subsets of size
. However, if the candidate k-itemset is infrequent, we need to check all

of its subsets in the next iteration. Another approach is to combine
both general-to-specific and specific-to-general search strategies. This
bidirectional approach requires more space to store the candidate
itemsets, but it can help to rapidly identify the frequent itemset border,
given the configuration shown in Figure 5.19(c) .

Figure 5.19.
General-to-specific, specific-to-general, and bidirectional search.



Equivalence Classes: Another way to envision the traversal is to first
partition the lattice into disjoint groups of nodes (or equivalence classes). A
frequent itemset generation algorithm searches for frequent itemsets within
a particular equivalence class first before moving to another equivalence
class. As an example, the level-wise strategy used in the Apriori algorithm
can be considered to be partitioning the lattice on the basis of itemset
sizes; i.e., the algorithm discovers all frequent 1-itemsets first before
proceeding to larger-sized itemsets. Equivalence classes can also be
defined according to the prefix or suffix labels of an itemset. In this case,
two itemsets belong to the same equivalence class if they share a
common prefix or suffix of length k. In the prefix-based approach, the
algorithm can search for frequent itemsets starting with the prefix a before
looking for those starting with prefixes b, c, and so on. Both prefix-based
and suffix-based equivalence classes can be demonstrated using the tree-
like structure shown in Figure 5.20 .

Figure 5.20.
Equivalence classes based on the prefix and suffix labels of itemsets.

Breadth-First versus Depth-First: The Apriori algorithm traverses the
lattice in a breadth-first manner, as shown in Figure 5.21(a) . It first
discovers all the frequent 1-itemsets, followed by the frequent 2-itemsets,
and so on, until no new frequent itemsets are generated. The itemset
lattice can also be traversed in a depth-first manner, as shown in Figures
5.21(b) and 5.22 . The algorithm can start from, say, node a in Figure
5.22 , and count its support to determine whether it is frequent. If so, the
algorithm progressively expands the next level of nodes, i.e., ab, abc, and
so on, until an infrequent node is reached, say, abcd. It then backtracks to
another branch, say, abce, and continues the search from there.

Figure 5.21.
Breadth-first and depth-first traversals.

Figure 5.22.
Generating candidate itemsets using the depth-first approach.

The depth-first approach is often used by algorithms designed to find
maximal frequent itemsets. This approach allows the frequent itemset
border to be detected more quickly than using a breadth-first approach.

Once a maximal frequent itemset is found, substantial pruning can be
performed on its subsets. For example, if the node bcde shown in Figure
5.22 is maximal frequent, then the algorithm does not have to visit the
subtrees rooted at bd, be, c, d, and e because they will not contain any
maximal frequent itemsets. However, if abc is maximal frequent, only the
nodes such as ac and bc are not maximal frequent (but the subtrees of ac
and bc may still contain maximal frequent itemsets). The depth-first
approach also allows a different kind of pruning based on the support of
itemsets. For example, suppose the support for {a, b, c} is identical to the
support for {a, b}. The subtrees rooted at abd and abe can be skipped

because they are guaranteed not to have any maximal frequent itemsets.
The proof of this is left as an exercise to the readers.

Representation of Transaction Data Set

There are many ways to represent a transaction data set. The choice of
representation can affect the I/O costs incurred when computing the support
of candidate itemsets. Figure 5.23 shows two different ways of
representing market basket transactions. The representation on the left is
called a horizontal data layout, which is adopted by many association rule
mining algorithms, including Apriori. Another possibility is to store the list of
transaction identifiers (TID-list) associated with each item. Such a
representation is known as the vertical data layout. The support for each
candidate itemset is obtained by intersecting the TID-lists of its subset items.
The length of the TID-lists shrinks as we progress to larger sized itemsets.
However, one problem with this approach is that the initial set of TID-lists
might be too large to fit into main memory, thus requiring more sophisticated
techniques to compress the TID-lists. We describe another effective approach
to represent the data in the next Section.

Figure 5.23.
Horizontal and vertical data format.

Horizontal Data Layout

5.6 FP-Growth Algorithm*
This Section presents an alternative algorithm called FP-growth that takes a
radically different approach to discovering frequent itemsets. The algorithm
does not subscribe to the generate-and-test paradigm of Apriori. Instead, it
encodes the data set using a compact data structure called an FP-tree and
extracts frequent itemsets directly from this structure. The details of this
approach are presented next.

5.6.1 FP-Tree Representation

An FP-tree is a compressed representation of the input data. It is constructed
by reading the data set one transaction at a time and mapping each
transaction onto a path in the FP-tree. As different transactions can have
several items in common, their paths might overlap. The more the paths
overlap with one another, the more compression we can achieve using the
FP-tree structure. If the size of the FP-tree is small enough to fit into main
memory, this will allow us to extract frequent itemsets directly from the
structure in memory instead of making repeated passes over the data stored
on disk.

Figure 5.24 shows a data set that contains ten transactions and five items.
The structures of the FP-tree after reading the first three transactions are also
depicted in the diagram. Each node in the tree contains the label of an item
along with a counter that shows the number of transactions mapped onto the
given path. Initially, the FP-tree contains only the root node represented by the
null symbol. The FP-tree is subsequently extended in the following way:

Figure 5.24.
Construction of an FP-tree.

1. The data set is scanned once to determine the support count of each
item. Infrequent items are discarded, while the frequent items are
sorted in decreasing support counts inside every transaction of the data

set. For the data set shown in Figure 5.24 , a is the most frequent
item, followed by b, c, d, and e.

2. The algorithm makes a second pass over the data to construct the FP-
tree. After reading the first transaction, {a, b}, the nodes labeled as a
and b are created. A path is then formed from to encode
the transaction. Every node along the path has a frequency count of 1.

3. After reading the second transaction, {b, c, d}, a new set of nodes is
created for items b, c, and d. A path is then formed to represent the
transaction by connecting the nodes . Every node
along this path also has a frequency count equal to one. Although the
first two transactions have an item in common, which is b, their paths
are disjoint because the transactions do not share a common prefix.

4. The third transaction, {a, c, d, e}, shares a common prefix item (which
is a) with the first transaction. As a result, the path for the third
transaction, , overlaps with the path for the first
transaction, . Because of their overlapping path, the
frequency count for node a is incremented to two, while the frequency
counts for the newly created nodes, c, d, and e, are equal to one.

5. This process continues until every transaction has been mapped onto
one of the paths given in the FP-tree. The resulting FP-tree after
reading all the transactions is shown at the bottom of Figure 5.24 .

The size of an FP-tree is typically smaller than the size of the uncompressed
data because many transactions in market basket data often share a few
items in common. In the best-case scenario, where all the transactions have
the same set of items, the FP-tree contains only a single branch of nodes. The
worst-case scenario happens when every transaction has a unique set of
items. As none of the transactions have any items in common, the size of the
FP-tree is effectively the same as the size of the original data. However, the
physical storage requirement for the FP-tree is higher because it requires
additional space to store pointers between nodes and counters for each item.




The size of an FP-tree also depends on how the items are ordered. The
notion of ordering items in decreasing order of support counts relies on the
possibility that the high support items occur more frequently across all paths
and hence must be used as most commonly occurring prefixes. For example,
if the ordering scheme in the preceding example is reversed, i.e., from lowest
to highest support item, the resulting FP-tree is shown in Figure 5.25 . The
tree appears to be denser because the branching factor at the root node has
increased from 2 to 5 and the number of nodes containing the high support
items such as a and b has increased from 3 to 12. Nevertheless, ordering by
decreasing support counts does not always lead to the smallest tree,
especially when the high support items do not occur frequently together with
the other items. For example, suppose we augment the data set given in
Figure 5.24 with 100 transactions that contain {e}, 80 transactions that
contain {d}, 60 transactions that contain {c}, and 40 transactions that contain
{b}. Item e is now most frequent, followed by d, c, b, and a. With the
augmented transactions, ordering by decreasing support counts will result in
an FP-tree similar to Figure 5.25 , while a scheme based on increasing
support counts produces a smaller FP-tree similar to Figure 5.24(iv) .

Figure 5.25.

An FP-tree representation for the data set shown in Figure 5.24 with a
different item ordering scheme.

An FP-tree also contains a list of pointers connecting nodes that have the
same items. These pointers, represented as dashed lines in Figures 5.24
and 5.25 , help to facilitate the rapid access of individual items in the tree.
We explain how to use the FP-tree and its corresponding pointers for frequent
itemset generation in the next Section.

5.6.2 Frequent Itemset Generation in
FP-Growth Algorithm

FP-growth is an algorithm that generates frequent itemsets from an FP-tree by
exploring the tree in a bottom-up fashion. Given the example tree shown in
Figure 5.24 , the algorithm looks for frequent itemsets ending in e first,
followed by d, c, b, and finally, a. This bottom-up strategy for finding frequent
itemsets ending with a particular item is equivalent to the suffix-based
approach described in Section 5.5 . Since every transaction is mapped onto
a path in the FP-tree, we can derive the frequent itemsets ending with a
particular item, say, e, by examining only the paths containing node e. These
paths can be accessed rapidly using the pointers associated with node e. The
extracted paths are shown in Figure 5.26 (a) . Similar paths for itemsets
ending in d, c, b, and a are shown in Figures 5.26 (b) , (c) , (d) , and
(e) , respectively.

Figure 5.26.
Decomposing the frequent itemset generation problem into multiple
subproblems, where each subproblem involves finding frequent itemsets
ending in e, d, c, b, and a.

FP-growth finds all the frequent itemsets ending with a particular suffix by
employing a divide-and-conquer strategy to split the problem into smaller
subproblems. For example, suppose we are interested in finding all frequent
itemsets ending in e. To do this, we must first check whether the itemset {e}
itself is frequent. If it is frequent, we consider the subproblem of finding
frequent itemsets ending in de,followedby ce, be,and ae. In turn, each of
these subproblems are further decomposed into smaller subproblems. By
merging the solutions obtained from the subproblems, all the frequent
itemsets ending in e can be found. Finally, the set of all frequent itemsets can
be generated by merging the solutions to the subproblems of finding frequent

itemsets ending in e, d, c, b, and a. This divide-and-conquer approach is the
key strategy employed by the FP-growth algorithm.

For a more concrete example on how to solve the subproblems, consider the
task of finding frequent itemsets ending with e.

1. The first step is to gather all the paths containing node e. These initial
paths are called prefix paths and are shown in Figure 5.27(a) .

Figure 5.27.
Example of applying the FP-growth algorithm to find frequent itemsets
ending in e.

2. From the prefix paths shown in Figure 5.27(a) , the support count for
e is obtained by adding the support counts associated with node e.
Assuming that the minimum support count is 2, {e} is declared a
frequent itemset because its support count is 3.

3. Because {e} is frequent, the algorithm has to solve the subproblems of
finding frequent itemsets ending in de, ce, be,and ae. Before solving
these subproblems, it must first convert the prefix paths into a
conditional FP-tree, which is structurally similar to an FP-tree, except
it is used to find frequent itemsets ending with a particular suffix. A
conditional FP-tree is obtained in the following way:

a. First, the support counts along the prefix paths must be updated
because some of the counts include transactions that do not
contain item e. For example, the rightmost path shown in Figure
5.27(a) , , includes a transaction {b, c}
that does not contain item e. The counts along the prefix path
must therefore be adjusted to 1 to reflect the actual number of
transactions containing {b, c, e}.

b. The prefix paths are truncated by removing the nodes for e.
These nodes can be removed because the support counts along
the prefix paths have been updated to reflect only transactions
that contain e and the subproblems of finding frequent itemsets
ending in de, ce, be, and ae no longer need information about
node e.

c. After updating the support counts along the prefix paths, some
of the items may no longer be frequent. For example, the node b
appears only once and has a support count equal to 1, which
means that there is only one transaction that contains both b
and e.Item b can be safely ignored from subsequent analysis
because all itemsets ending in be must be infrequent.

The conditional FP-tree for e is shown in Figure 5.27(b) . The tree
looks different than the original prefix paths because the frequency
counts have been updated and the nodes b and e have been

4. FP-growth uses the conditional FP-tree for e to solve the subproblems
of finding frequent itemsets ending in de, ce,and ae. To find the
frequent itemsets ending in de, the prefix paths for d are gathered from
the conditional FP-tree for e (Figure 5.27(c) ). By adding the
frequency counts associated with node d, we obtain the support count
for {d, e}. Since the support count is equal to 2, {d, e} is declared a
frequent itemset. Next, the algorithm constructs the conditional FP-tree
for de using the approach described in step 3. After updating the
support counts and removing the infrequent item c, the conditional FP-
tree for de is shown in Figure 5.27(d) . Since the conditional FP-tree
contains only one item, a, whose support is equal to minsup, the
algorithm extracts the frequent itemset {a, d, e} and moves on to the
next subproblem, which is to generate frequent itemsets ending in ce.
After processing the prefix paths for c, {c, e} is found to be frequent.
However, the conditional FP-tree for ce will have no frequent items and
thus will be eliminated. The algorithm proceeds to solve the next
subproblem and finds {a, e} to be the only frequent itemset remaining.

This example illustrates the divide-and-conquer approach used in the FP-
growth algorithm. At each recursive step, a conditional FP-tree is constructed
by updating the frequency counts along the prefix paths and removing all
infrequent items. Because the subproblems are disjoint, FP-growth will not
generate any duplicate itemsets. In addition, the counts associated with the
nodes allow the algorithm to perform support counting while generating the
common suffix itemsets.

FP-growth is an interesting algorithm because it illustrates how a compact
representation of the transaction data set helps to efficiently generate frequent
itemsets. In addition, for certain transaction data sets, FP-growth outperforms
the standard Apriori algorithm by several orders of magnitude. The run-time
performance of FP-growth depends on the compaction factor of the data set.
If the resulting conditional FP-trees are very bushy (in the worst case, a full
prefix tree), then the performance of the algorithm degrades significantly
because it has to generate a large number of subproblems and merge the
results returned by each subproblem.

5.7 Evaluation of Association Patterns
Although the Apriori principle significantly reduces the exponential search
space of candidate itemsets, association analysis algorithms still have the
potential to generate a large number of patterns. For example, although the
data set shown in Table 5.1 contains only six items, it can produce
hundreds of association rules at particular support and confidence thresholds.
As the size and dimensionality of real commercial databases can be very
large, we can easily end up with thousands or even millions of patterns, many
of which might not be interesting. Identifying the most interesting patterns from
the multitude of all possible ones is not a trivial task because “one person’s
trash might be another person’s treasure.” It is therefore important to establish
a set of well-accepted criteria for evaluating the quality of association patterns.

The first set of criteria can be established through a data-driven approach to
define objective interestingness measures. These measures can be used
to rank patterns—itemsets or rules—and thus provide a straightforward way of
dealing with the enormous number of patterns that are found in a data set.
Some of the measures can also provide statistical information, e.g., itemsets
that involve a set of unrelated items or cover very few transactions are
considered uninteresting because they may capture spurious relationships in
the data and should be eliminated. Examples of objective interestingness
measures include support, confidence, and correlation.

The second set of criteria can be established through subjective arguments. A
pattern is considered subjectively uninteresting unless it reveals unexpected
information about the data or provides useful knowledge that can lead to
profitable actions. For example, the rule may not be
interesting, despite having high support and confidence values, because the


relationship represented by the rule might seem rather obvious. On the other
hand, the rule is interesting because the relationship is
quite unexpected and may suggest a new cross-selling opportunity for
retailers. Incorporating subjective knowledge into pattern evaluation is a
difficult task because it requires a considerable amount of prior information
from domain experts. Readers interested in subjective interestingness
measures may refer to resources listed in the bibliography at the end of this

5.7.1 Objective Measures of

An objective measure is a data-driven approach for evaluating the quality of
association patterns. It is domain-independent and requires only that the user
specifies a threshold for filtering low-quality patterns. An objective measure is
usually computed based on the frequency counts tabulated in a contingency
table. Table 5.6 shows an example of a contingency table for a pair of
binary variables, A and B.We use the notation to indicate that A
(B)isabsent from a transaction. Each entry in this table denotes a
frequency count. For example, is the number of times A and B appear
together in the same transaction, while is the number of transactions that
contain B but not A. The row sum represents the support count for A,
while the column sum represents the support count for B. Finally, even
though our discussion focuses mainly on asymmetric binary variables, note
that contingency tables are also applicable to other attribute types such as
symmetric binary, nominal, and ordinal variables.

Table 5.6. A 2-way contingency table for variables A and B.


fij 2×2






Limitations of the Support-Confidence Framework
The classical association rule mining formulation relies on the support and
confidence measures to eliminate uninteresting patterns. The drawback of
support, which is described more fully in Section 5.8 , is that many
potentially interesting patterns involving low support items might be eliminated
by the support threshold. The drawback of confidence is more subtle and is
best demonstrated with the following example.

Example 5.3.
Suppose we are interested in analyzing the relationship between people
who drink tea and coffee. We may gather information about the beverage
preferences among a group of people and summarize their responses into
a contingency table such as the one shown in Table 5.7 .

Table 5.7. Beverage preferences among a group of 1000 people.


Tea 150 50 200

650 150 800

800 200 1000

f11 f10 f1+

A¯ f01 f00 f0+

f+1 f+0



The information given in this table can be used to evaluate the association
rule . At first glance, it may appear that people who drink
tea also tend to drink coffee because the rule’s support (15%) and
confidence (75%) values are reasonably high. This argument would have
been acceptable except that the fraction of people who drink coffee,
regardless of whether they drink tea, is 80%, while the fraction of tea
drinkers who drink coffee is only 75%. Thus knowing that a person is a tea
drinker actually decreases her probability of being a coffee drinker from
80% to 75%! The rule is therefore misleading despite its
high confidence value.

Now consider a similar problem where we are interested in analyzing the
relationship between people who drink tea and people who use honey in
their beverage. Table 5.8 summarizes the information gathered over the
same group of people about their preferences for drinking tea and using
honey. If we evaluate the association rule using this
information, we will find that the confidence value of this rule is merely
50%, which might be easily rejected using a reasonable threshold on the
confidence value, say 70%. One thus might consider that the preference of
a person for drinking tea has no influence on her preference for using
honey. However, the fraction of people who use honey, regardless of
whether they drink tea, is only 12%. Hence, knowing that a person drinks
tea significantly increases her probability of using honey from 12% to 50%.
Further, the fraction of people who do not drink tea but use honey is only
2.5%! This suggests that there is definitely some information in the
preference of a person of using honey given that she drinks tea. The rule

may therefore be falsely rejected if confidence is used as
the evaluation measure.

Table 5.8. Information about people who drink tea and people who
use honey in their beverage.






Tea 100 100 200

20 780 800

120 880 1000

Note that if we take the support of coffee drinkers into account, we would not
be surprised to find that many of the people who drink tea also drink coffee,
since the overall number of coffee drinkers is quite large by itself. What is
more surprising is that the fraction of tea drinkers who drink coffee is actually
less than the overall fraction of people who drink coffee, which points to an
inverse relationship between tea drinkers and coffee drinkers. Similarly, if we
account for the fact that the support of using honey is inherently small, it is
easy to understand that the fraction of tea drinkers who use honey will
naturally be small. Instead, what is important to measure is the change in the
fraction of honey users, given the information that they drink tea.

The limitations of the confidence measure are well-known and can be
understood from a statistical perspective as follows. The support of a variable
measures the probability of its occurrence, while the support s(A, B) of a pair
of a variables A and B measures the probability of the two variables occurring
together. Hence, the joint probability P (A, B) can be written as

If we assume A and B are statistically independent, i.e. there is no relationship
between the occurrences of A and B, then . Hence, under
the assumption of statistical independence between A and B, the support
sindep(A, B) of A and B can be written as



P(A, B)=s(A, B)=f11N.

P(A, B)=P(A)×P(B)

If the support between two variables, s(A, B) is equal to , then A
and B can be considered to be unrelated to each other. However, if s(A, B) is
widely different from , then A and B are most likely dependent.
Hence, any deviation of s(A, B) from can be seen as an indication
of a statistical relationship between A and B. Since the confidence measure
only considers the deviance of s(A, B) from s(A) and not from , it
fails to account for the support of the consequent, namely s(B). This results in
the detection of spurious patterns (e.g., ) and the rejection of
truly interesting patterns (e.g., ), as illustrated in the previous

Various objective measures have been used to capture the deviance of s(A,
B) from , that are not susceptible to the limitations of the
confidence measure. Below, we provide a brief description of some of these
measures and discuss some of their properties.

Interest Factor
The interest factor, which is also called as the “lift,” can be defined as follows:

Notice that . Hence, the interest factor measures the
ratio of the support of a pattern s(A, B) against its baseline support (A,
B) computed under the statistical independence assumption. Using
Equations 5.5 and 5.4 , we can interpret the measure as follows:

sindep(A, B)=s(A)×s(B)or equivalently,sindep(A, B)=f1+N×f+1N. (5.4)

sindep(A, B)

sindep(A, B)



sindep(A, B)

I(A, B)=s(A, B)s(A)×s(B)=Nf11f1+f+1. (5.5)

s(A)×s(B)=sindep(A, B)

I(A, B)={=1,if A and B are independent;>1,if A and B are positively related;
<1,if A and B are negatively related.


For the tea-coffee example shown in Table 5.7 , , thus
suggesting a slight negative relationship between tea drinkers and coffee
drinkers. Also, for the tea-honey example shown in Table 5.8 ,

, suggesting a strong positive relationship between
people who drink tea and people who use honey in their beverage. We can
thus see that the interest factor is able to detect meaningful patterns in the
tea-coffee and tea-honey examples. Indeed, the interest factor has a number
of statistical advantages over the confidence measure that make it a suitable
measure for analyzing statistical independence between variables.

Piatesky-Shapiro (PS) Measure
Instead of computing the ratio between s(A, B) and ,
the PS measure considers the difference between s(A, B) and as

The PS value is 0 when A and B are mutually independent of each other.
Otherwise, when there is a positive relationship between the two
variables, and when there is a negative relationship.

Correlation Analysis
Correlation analysis is one of the most popular techniques for analyzing
relationships between a pair of variables. For continuous variables, correlation
is defined using Pearson’s correlation coefficient (see Equation 2.10 on
page 83). For binary variables, correlation can be measured using the

, which is defined as



sindep(A, B)=s(A)×s(B)

PS=s(A, B)−s(A)×s(B)=f11N−f1+f+1N2 (5.7)



ϕ=f11f00−f01f10f1+f+1f0+f+0. (5.8)

If we rearrange the terms in 5.8, we can show that the can be
rewritten in terms of the support measures of A, B, and {A, B} as follows:

Note that the numerator in the above equation is identical to the PS measure.
Hence, the can be understood as a normalized version of the PS
measure, where that the value of the ranges from to . From
a statistical viewpoint, the correlation captures the normalized difference
between s(A, B) and (A, B). A correlation value of 0 means no
relationship, while a value of suggests a perfect positive relationship and a
value of suggests a perfect negative relationship. The correlation measure
has a statistical meaning and hence is widely used to evaluate the strength of
statistical independence among variables. For instance, the correlation
between tea and coffee drinkers in Table 5.7 is which is slightly
less than 0. On the other hand, the correlation between people who drink tea
and people who use honey in Table 5.8 is 0.5847, suggesting a positive

IS Measure
IS is an alternative measure for capturing the relationship between s(A, B) and

. The IS measure is defined as follows:

Although the definition of IS looks quite similar to the interest factor, they
share some interesting differences. Since IS is the geometric mean between
the interest factor and the support of a pattern, IS is large when both the
interest factor and support are large. Hence, if the interest factor of two
patterns are identical, the IS has a preference of selecting the pattern with
higher support. It is also possible to show that IS is mathematically equivalent


ϕ=s(A, B)−s(A)×s(B)s(A)×(1−s(A))×s(B)×(1−s(B)). (5.9)

ϕ-coefficient −1 +1





IS(A, B)=I(A, B)×s(A, B)=s(A, B)s(A)s(B)=f11f1+f+1. (5.10)

to the cosine measure for binary variables (see Equation 2.6 on page
81 ). The value of IS thus varies from 0 to 1, where an IS value of 0
corresponds to no co-occurrence of the two variables, while an IS value of 1
denotes perfect relationship, since they occur in exactly the same
transactions. For the tea-coffee example shown in Table 5.7 , the value of
IS is equal to 0.375, while the value of IS for the tea-honey example in Table
5.8 is 0.6455. The IS measure thus gives a higher value for the

rule than the rule, which is consistent with our
understanding of the two rules.

Alternative Objective Interestingness Measures
Note that all of the measures defined in the previous Section use different
techniques to capture the deviance between s(A, B) and .
Some measures use the ratio between s(A, B) and (A, B), e.g., the
interest factor and IS, while some other measures consider the difference
between the two, e.g., the PS and the . Some measures are
bounded in a particular range, e.g., the IS and the , while others
are unbounded and do not have a defined maximum or minimum value, e.g.,
the Interest Factor. Because of such differences, these measures behave
differently when applied to different types of patterns. Indeed, the measures
defined above are not exhaustive and there exist many alternative measures
for capturing different properties of relationships between pairs of binary
variables. Table 5.9 provides the definitions for some of these measures in
terms of the frequency counts of a contingency table.

Table 5.9. Examples of objective measures for the itemset {A, B}.

Measure (Symbol) Definition


→{Honey} {Tea}→{Coffee}




(ϕ) Nf11−f1+f+1f1+f+1f0+f+0

Odds ratio


Interest (I)

Cosine (IS)

Piatetsky-Shapiro (PS)

Collective strength (S)


All-confidence (h)

Consistency among Objective Measures
Given the wide variety of measures available, it is reasonable to question
whether the measures can produce similar ordering results when applied to a
set of association patterns. If the measures are consistent, then we can
choose any one of them as our evaluation metric. Otherwise, it is important to
understand what their differences are in order to determine which measure is
more suitable for analyzing certain types of patterns.

Suppose the measures defined in Table 5.9 are applied to rank the ten
contingency tables shown in Table 5.10 . These contingency tables are
chosen to illustrate the differences among the existing measures. The
ordering produced by these measures is shown in Table 5.11 (with 1 as the
most interesting and 10 as the least interesting table). Although some of the
measures appear to be consistent with each other, others produce quite
different ordering results. For example, the rankings given by the
agrees mostly with those provided by and collective strength, but are quite

(α) (f11f00)/(f10f01)

(κ) Nf11+Nf00−f1+f+1−f0+f+0N2−f1+f+1−f0+f+0





(ζ) f11/(f1++f+1−f11)

min[f11f1+, f11f+1]


different than the rankings produced by interest factor. Furthermore, a
contingency table such as is ranked lowest according to the ,
but highest according to interest factor.

Table 5.10. Example of contingency tables.


8123 83 424 1370

8330 2 622 1046

3954 3080 5 2961

2886 1363 1320 4431

1500 2000 500 6000

4000 2000 1000 3000

9481 298 127 94

4000 2000 2000 2000

7450 2483 4 63

61 2483 4 7452

Table 5.11. Rankings of contingency tables using the measures given in
Table 5.9 .


1 3 1 6 2 2 1 2 2

2 1 2 7 3 5 2 3 3

E10 ϕ-coefficient

f11 f10 f01 f00











ϕ α κ ζ



3 2 4 4 5 1 3 6 8

4 8 3 3 7 3 4 7 5

5 7 6 2 9 6 6 9 9

6 9 5 5 6 4 5 5 7

7 6 7 9 1 8 7 1 1

8 10 8 8 8 7 8 8 7

9 4 9 10 4 9 9 4 4

10 5 10 1 10 10 10 10 10

Properties of Objective Measures
The results shown in Table 5.11 suggest that the measures greatly differ
from each other and can provide conflicting information about the quality of a
pattern. In fact, no measure is universally best for all applications. In the
following, we describe some properties of the measures that play an important
role in determining if they are suited for a certain application.

Inversion Property

Consider the binary vectors shown in Figure 5.28 . The 0/1 value in each
column vector indicates whether a transaction (row) contains a particular item
(column). For example, the vector A indicates that the item appears in the first
and last transactions, whereas the vector B indicates that the item is
contained only in the fifth transaction. The vectors and are the inverted
versions of A and B, i.e., their 1 values have been changed to 0 values
(absence to presence) and vice versa. Applying this transformation to a binary









A¯ B¯

vector is called inversion. If a measure is invariant under the inversion
operation, then its value for the vector pair should be identical to its
value for {A, B}. The inversion property of a measure can be tested as follows.

Figure 5.28.
Effect of the inversion operation. The vectors and are inversions of
vectors A and B, respectively.

Definition 5.6. (Inversion Property.)
An objective measure M is invariant under the inversion
operation if its value remains the same when exchanging the
frequency counts with and with .

Measures that are invariant to the inversion property include the correlation
( ), odds ratio, , and collective strength. These measures are
especially useful in scenarios where the presence (1’s) of a variable is as

{A¯, B¯}

A¯ E¯

f11 f00 f10 f01

ϕ-coefficient κ

important as its absence (0’s). For example, if we compare two sets of
answers to a series of true/false questions where 0’s (true) and 1’s (false) are
equally meaningful, we should use a measure that gives equal importance to
occurrences of 0–0’s and 1–1’s. For the vectors shown in Figure 5.28 , the

is equal to -0.1667 regardless of whether we consider the pair {A,
B} or pair . Similarly, the odds ratio for both pairs of vectors is equal to
a constant value of 0. Note that even though the and the odds
ratio are invariant to inversion, they can still show different results, as will be
shown later.

Measures that do not remain invariant under the inversion operation include
the interest factor and the IS measure. For example, the IS value for the pair

in Figure 5.28 is 0.825, which reflects the fact that the 1’s in
and occur frequently together. However, the IS value of its inverted pair {A,
B} is equal to 0, since A and B do not have any co-occurrence of 1’s. For
asymmetric binary variables, e.g., the occurrence of words in documents, this
is indeed the desired behavior. A desired similarity measure between
asymmetric variables should not be invariant to inversion, since for these
variables, it is more meaningful to capture relationships based on the
presence of a variable rather than its absence. On the other hand, if we are
dealing with symmetric binary variables where the relationships between 0’s
and 1’s are equally meaningful, care should be taken to ensure that the
chosen measure is invariant to inversion.

Although the values of the interest factor and IS change with the inversion
operation, they can still be inconsistent with each other. To illustrate this,
consider Table 5.12 , which shows the contingency tables for two pairs of
variables, {p, q} and {r, s}. Note that r and s are inverted transformations of p
and q, respectively, where the roles of 0’s and 1’s have just been reversed.
The interest factor for {p, q} is 1.02 and for {r, s} is 4.08, which means that the
interest factor finds the inverted pair {r, s} more related than the original pair

{A¯, B¯}


{A¯, B¯} A¯

{p, q}. On the contrary, the IS value decreases upon inversion from 0.9346 for
{p, q} to 0.286 for {r, s}, suggesting quite an opposite trend to that of the
interest factor. Even though these measures conflict with each other for this
example, they may be the right choice of measure in different applications.

Table 5.12. Contingency tables for the pairs {p,q} and {r,s}.


q 880 50 930

50 20 70

930 70 1000


s 20 50 70

50 880 930

70 930 1000

Scaling Property

Table 5.13 shows two contingency tables for gender and the grades
achieved by students enrolled in a particular course. These tables can be
used to study the relationship between gender and performance in the course.
The second contingency table has data from the same population but has
twice as many males and three times as many females. The actual number of
males or females can depend upon the samples available for study, but the
relationship between gender and grade should not change just because of
differences in sample sizes. Similarly, if the number of students with high and
low grades are changed in a new study, a measure of association between

gender and grades should remain unchanged. Hence, we need a measure
that is invariant to scaling of rows or columns. The process of multiplying a
row or column of a contingency table by a constant value is called a row or
column scaling operation. A measure that is invariant to scaling does not
change its value after any row or column scaling operation.

Table 5.13. The grade-gender example. (a) Sample data of size 100.

Male Female

High 30 20 50

Low 40 10 50

70 30 100

(b) Sample data of size 230.

Male Female

High 60 60 120

Low 80 30 110

140 90 230

Definition 5.7. (Scaling Invariance
Let T be a contingency table with frequency counts

. Let be the transformed a contingency table[f11; f10; f01; f00] T′

with scaled frequency counts
, where are

positive constants used to scale the two rows and the two
columns of T . An objective measure M is invariant under the
row/column scaling operation if .

Note that the use of the term ‘scaling’ here should not be confused with the
scaling operation for continuous variables introduced in Chapter 2 on page
23, where all the values of a variable were being multiplied by a constant
factor, instead of scaling a row or column of a contingency table.

Scaling of rows and columns in contingency tables occurs in multiple ways in
different applications. For example, if we are measuring the effect of a
particular medical procedure on two sets of subjects, healthy and diseased,
the ratio of healthy and diseased subjects can widely vary across different
studies involving different groups of participants. Further, the fraction of
healthy and diseased subjects chosen for a controlled study can be quite
different from the true fraction observed in the complete population. These
differences can result in a row or column scaling in the contingency tables for
different populations of subjects. In general, the frequencies of items in a
contingency table closely depends on the sample of transactions used to
generate the table. Any change in the sampling procedure may affect a row or
column scaling transformation. A measure that is expected to be invariant to
differences in the sampling procedure must not change with row or column

Of all the measures introduced in Table 5.9 , only the odds ratio is
invariant to row and column scaling operations. For example, the value of
odds ratio for both the tables in Table 5.13 is equal to 0.375. All other

[k1k3f11; k2k3f10; k1k4f01; k2k4f00] k1, k2, k3, k4



measures such as the , IS, interest factor, and collective
strength (S) change their values when the rows and columns of the
contingency table are rescaled. Indeed, the odds ratio is a preferred choice of
measure in the medical domain, where it is important to find relationships that
do not change with differences in the population sample chosen for a study.

Null Addition Property

Suppose we are interested in analyzing the relationship between a pair of
words, such as and , in a set of documents. If a collection of
articles about ice fishing is added to the data set, should the association
between and be affected? This process of adding unrelated data
(in this case, documents) to a given data set is known as the null addition

Definition 5.8. (Null Addition Property.)
An objective measure M is invariant under the null addition
operation if it is not affected by increasing , while all other
frequencies in the contingency table stay the same.

For applications such as document analysis or market basket analysis, we
would like to use a measure that remains invariant under the null addition
operation. Otherwise, the relationship between words can be made to change
simply by adding enough documents that do not contain both words!
Examples of measures that satisfy this property include cosine (IS) and

ϕ-coefficient, κ


Jaccard measures, while those that violate this property include interest
factor, PS, odds ratio, and the .

To demonstrate the effect of null addition, consider the two contingency tables
and shown in Table 5.14 . Table has been obtained from by

adding 1000 extra transactions with both A and B absent. This operation only
affects the entry of Table , which has increased from 100 to 1100,
whereas all the other frequencies in the table , and remain the
same. Since IS is invariant to null addition, it gives a constant value of 0.875
to both the tables. However, the addition of 1000 extra transactions with
occurrences of 0–0’s changes the value of interest factor from 0.972 for
(denoting a slightly negative correlation) to 1.944 for (positive correlation).
Similarly, the value of odds ratio increases from 7 for to 77 for . Hence,
when the interest factor or odds ratio are used as the association measure,
the relationships between variables changes by the addition of null
transactions where both the variables are absent. In contrast, the IS measure
is invariant to null addition, since it considers two variables to be related only if
they frequently occur together. Indeed, the IS measure (cosine measure) is
widely used to measure similarity among documents, which is expected to
depend only on the joint occurrences (1’s) of words in documents, but not
their absences (0’s).

Table 5.14. An example demonstrating the effect of null addition.
(a) Table .


A 700 100 800

100 100 200

800 200 1000


T1 T2 T2 T1

f00 T2
(f11, f10 f01)

T1 T2


(b) Table .


A 700 100 800

10 1100 1200

800 1200 2000

Table 5.15 provides a summary of properties for the measures defined in
Table 5.9 . Even though this list of properties is not exhaustive, it can serve
as a useful guide for selecting the right choice of measure for an application.
Ideally, if we know the specific requirements of a certain application, we can
ensure that the selected measure shows properties that adhere to those
requirements. For example, if we are dealing with asymmetric variables, we
would prefer to use a measure that is not invariant to null addition or inversion.
On the other hand, if we require the measure to remain invariant to changes in
the sample size, we would like to use a measure that does not change with

Table 5.15. Properties of symmetric measures.

Symbol Measure Inversion Null Addition Scaling

Yes No No

odds ratio Yes No Yes

Cohen’s Yes No No

I Interest No No No

IS Cosine No Yes No

PS Piatetsky-Shapiro’s Yes No No


ϕ ϕ-coefficient



S Collective strength Yes No No

Jaccard No Yes No

h All-confidence No Yes No

s Support No No No

Asymmetric Interestingness Measures
Note that in the discussion so far, we have only considered measures that do
not change their value when the order of the variables are reversed. More
specifically, if M is a measure and A and B are two variables, then M(A, B) is
equal to M(B, A) if the order of the variables does not matter. Such measures
are called symmetric. On the other hand, measures that depend on the order
of variables are called asymmetric measures. For
example, the interest factor is a symmetric measure because its value is
identical for the rules and . In contrast, confidence is an
asymmetric measure since the confidence for and may not be the
same. Note that the use of the term ‘asymmetric’ to describe a particular type
of measure of relationship—one in which the order of the variables is
important—should not be confused with the use of ‘asymmetric’ to describe a
binary variable for which only 1’s are important. Asymmetric measures are
more suitable for analyzing association rules, since the items in a rule do have
a specific order. Even though we only considered symmetric measures to
discuss the different properties of association measures, the above discussion
is also relevant for the asymmetric measures. See Bibliographic Notes for
more information about different kinds of asymmetric measures and their


(M(A, B)≠M(B, A))


5.7.2 Measures beyond Pairs of Binary

The measures shown in Table 5.9 are defined for pairs of binary variables
(e.g., 2-itemsets or association rules). However, many of them, such as
support and all-confidence, are also applicable to larger-sized itemsets. Other
measures, such as interest factor, IS, PS, and Jaccard coefficient, can be
extended to more than two variables using the frequency tables tabulated in a
multidimensional contingency table. An example of a three-dimensional
contingency table for a, b, and c is shown in Table 5.16 . Each entry in
this table represents the number of transactions that contain a particular
combination of items a, b, and c. For example, is the number of
transactions that contain a and c, but not b. On the other hand, a marginal
frequency such as is the number of transactions that contain a and c,
irrespective of whether b is present in the transaction.

Table 5.16. Example of a three-dimensional contingency table.

c b


c b





f111 f101 f1+1

a¯ f011 f001 f0+1

f+11 f+01 f++1

f110 f100 f1+0

a¯ f010 f000 f0+0

Given a k-itemset , the condition for statistical independence can
be stated as follows:

With this definition, we can extend objective measures such as interest factor
and PS, which are based on deviations from statistical independence, to more
than two variables:

Another approach is to define the objective measure as the maximum,
minimum, or average value for the associations between pairs of items in a
pattern. For example, given a k-itemset , we may define the

for X as the average between every pair of items
in X. However, because the measure considers only pairwise

associations, it may not capture all the underlying relationships within a
pattern. Also, care should be taken in using such alternate measures for more
than two variables, since they may not always show the anti-monotone
property in the same way as the support measure, making them unsuitable for
mining patterns using the Apriori principle.

Analysis of multidimensional contingency tables is more complicated because
of the presence of partial associations in the data. For example, some
associations may appear or disappear when conditioned upon the value of
certain variables. This problem is known as Simpson’s paradox and is
described in Section 5.7.3 . More sophisticated statistical techniques are

f+10 f+00 f++0

{i1, i2, …, ik}

fi1i2…ik=fi1+…+×f+i2…+×…×f++…ikNk−1. (5.11)


X={i1, i2, …, ik} ϕ-
coefficient ϕ-coefficient
(ip, iq)

available to analyze such relationships, e.g., loglinear models, but these
techniques are beyond the scope of this book.

5.7.3 Simpson’s Paradox

It is important to exercise caution when interpreting the association between
variables because the observed relationship may be influenced by the
presence of other confounding factors, i.e., hidden variables that are not
included in the analysis. In some cases, the hidden variables may cause the
observed relationship between a pair of variables to disappear or reverse its
direction, a phenomenon that is known as Simpson’s paradox. We illustrate
the nature of this paradox with the following example.

Consider the relationship between the sale of high-definition televisions
(HDTV) and exercise machines, as shown in Table 5.17 . The rule

has a confidence of and
the rule has a confidence of

. Together, these rules suggest that customers who buy high-
definition televisions are more likely to buy exercise machines than those who
do not buy high-definition televisions.

Table 5.17. A two-way contingency table between the sale of high-
definition television and exercise machine.


Buy Exercise Machine

Yes No

Yes 99 81 180

No 54 66 120

{HDTV=Yes}→{Exercise machine=Yes} 99/180=55%
{HDTV=No}→{Exercise machine=Yes}


153 147 300

However, a deeper analysis reveals that the sales of these items depend on
whether the customer is a college student or a working adult. Table 5.18
summarizes the relationship between the sale of HDTVs and exercise
machines among college students and working adults. Notice that the support
counts given in the table for college students and working adults sum up to
the frequencies shown in Table 5.17 . Furthermore, there are more working
adults than college students who buy these items. For college students:

Table 5.18. Example of a three-way contingency table.



Buy Exercise Machine Total

Yes No

College Students Yes 1 9 10

No 4 30 34

Working Adult Yes 98 72 170

No 50 36 86

while for working adults:

c({HDTV=Yes}→{Exercise machine=Yes})=1/10=10%,c({HDTV=No}
→{Exercise machine=Yes})=4/34=11.8%.

c({HDTV=Yes}→{Exercise machine=Yes})=98/170=57.7%,c({HDTV=No}
→{Exercise machine=Yes})=50/86=58.1%.

The rules suggest that, for each group, customers who do not buy high-
definition televisions are more likely to buy exercise machines, which
contradicts the previous conclusion when data from the two customer groups
are pooled together. Even if alternative measures such as correlation, odds
ratio, or interest are applied, we still find that the sale of HDTV and exercise
machine is positively related in the combined data but is negatively related in
the stratified data (see Exercise 21 on page 449). The reversal in the direction
of association is known as Simpson’s paradox.

The paradox can be explained in the following way. First, notice that most
customers who buy HDTVs are working adults. This is reflected in the high
confidence of the rule .
Second, the high confidence of the rule

suggests that most customers who buy
exercise machines are also working adults. Since working adults form the
largest fraction of customers for both HDTVs and exercise machines, they
both look related and the rule turns
out to be stronger in the combined data than what it would have been if the
data is stratified. Hence, customer group acts as a hidden variable that affects
both the fraction of customers who buy HDTVs and those who buy exercise
machines. If we factor out the effect of the hidden variable by stratifying the
data, we see that the relationship between buying HDTVs and buying exercise
machines is not direct, but shows up as an indirect consequence of the effect
of the hidden variable.

The Simpson’s paradox can also be illustrated mathematically as follows.

{HDTV=Yes}→{Working Adult}(170/180=94.4%)
{Exercise machine=Yes}

→{Working Adult}(148/153=96.7%)

{HDTV=Yes}→{Exercise machine=Yes}


where a/b and p/q may represent the confidence of the rule in two
different strata, while c/d and r/s may represent the confidence of the rule

in the two strata. When the data is pooled together, the confidence
values of the rules in the combined data are and ,
respectively. Simpson’s paradox occurs when

thus leading to the wrong conclusion about the relationship between the
variables. The lesson here is that proper stratification is needed to avoid
generating spurious patterns resulting from Simpson’s paradox. For example,
market basket data from a major supermarket chain should be stratified
according to store locations, while medical records from various patients
should be stratified according to confounding factors such as age and gender.


(a+p)/(b+q) (c+r)/(d+s)


5.8 Effect of Skewed Support
The performances of many association analysis algorithms are influenced by
properties of their input data. For example, the computational complexity of
the Apriori algorithm depends on properties such as the number of items in
the data, the average transaction width, and the support threshold used. This
Section examines another important property that has significant influence on
the performance of association analysis algorithms as well as the quality of
extracted patterns. More specifically, we focus on data sets with skewed
support distributions, where most of the items have relatively low to moderate
frequencies, but a small number of them have very high frequencies.

Figure 5.29.
A transaction data set containing three items, p, q, and r, where p is a high
support item and q and r are low support items.

Figure 5.29 shows an illustrative example of a data set that has a skewed
support distribution of its items. While p has a high support of 83.3% in the
data, q and r are low-support items with a support of 16.7%. Despite their low
support, q and r always occur together in the limited number of transactions
that they appear and hence are strongly related. A pattern mining algorithm
therefore should report {q, r} as interesting.

However, note that choosing the right support threshold for mining item-sets
such as {q, r} can be quite tricky. If we set the threshold too high (e.g., 20%),

then we may miss many interesting patterns involving low-support items such
as {q, r}. Conversely, setting the support threshold too low can be detrimental
to the pattern mining process for the following reasons. First, the
computational and memory requirements of existing association analysis
algorithms increase considerably with low support thresholds. Second, the
number of extracted patterns also increases substantially with low support
thresholds, which makes their analysis and interpretation difficult. In particular,
we may extract many spurious patterns that relate a high-frequency item such
as p to a low-frequency item such as q. Such patterns, which are called
cross-support patterns, are likely to be spurious because the association
between p and q is largely influenced by the frequent occurrence of p instead
of the joint occurrence of p and q together. Because the support of {p, q} is
quite close to the support of {q, r}, we may easily select {p, q} if we set the
support threshold low enough to include {q, r}.

An example of a real data set that exhibits a skewed support distribution is
shown in Figure 5.30 . The data, taken from the PUMS (Public Use
Microdata Sample) census data, contains 49,046 records and 2113
asymmetric binary variables. We shall treat the asymmetric binary variables
as items and records as transactions. While more than 80% of the items have
support less than 1%, a handful of them have support greater than 90%. To
understand the effect of skewed support distribution on frequent itemset
mining, we divide the items into three groups, , and , according to
their support levels, as shown in Table 5.19 . We can see that more than
82% of items belong to and have a support less than 1%. In market basket
analysis, such low support items may correspond to expensive products (such
as jewelry) that are seldom bought by customers, but whose patterns are still
interesting to retailers. Patterns involving such low-support items, though
meaningful, can easily be rejected by a frequent pattern mining algorithm with
a high support threshold. On the other hand, setting a low support threshold
may result in the extraction of spurious patterns that relate a high-frequency

G1, G2 G3


item in to a low-frequency item in . For example, at a support threshold
equal to 0.05%, there are 18,847 frequent pairs involving items from and

. Out of these, 93% of them are cross-support patterns; i.e., the patterns
contain items from both and .

Figure 5.30.
Support distribution of items in the census data set.

Table 5.19. Grouping the items in the census data set based on their
support values.



Number of Items 1735 358 20

G3 G1

G1 G3

G1 G2 G3

<1% 1%−90% >90%

This example shows that a large number of weakly related cross-support
patterns can be generated when the support threshold is sufficiently low. Note
that finding interesting patterns in data sets with skewed support distributions
is not just a challenge for the support measure, but similar statements can be
made about many other objective measures discussed in the previous
Sections. Before presenting a methodology for finding interesting patterns and
pruning spurious ones, we formally define the concept of cross-support

Definition 5.9. (Cross-support Pattern.)
Let us define the support ratio, r(X), of an itemset


Given a user-specified threshold , an itemset X is a cross-
support pattern if .

Example 5.4.
Suppose the support for milk is 70%, while the support for sugar is 10%
and caviar is 0.04%. Given , the frequent itemset {milk, sugar,
caviar} is a cross-support pattern because its support ratio is

{i1, i2, …, ik}

r(X)=min[s(i1), s(i2), …, s(ik)}max[s(i1), s(i2), …, s(ik)} (5.12)



r=min[0.7, 0.1, 0.0004]max[0.7, 0.1, 0.0004]=0.0040.7=0.00058<0.01.

Existing measures such as support and confidence may not be sufficient to
eliminate cross-support patterns. For example, if we assume for the
data set presented in Figure 5.29 , the itemsets {p, q}, {p, r}, and {p, q, r}
are cross-support patterns because their support ratios, which are equal to
0.2, are less than the threshold . However, their supports are comparable to
that of {q, r}, making it difficult to eliminate cross-support patterns without
loosing interesting ones using a support-based pruning strategy. Confidence
pruning also does not help because the confidence of the rules extracted from
cross-support patterns can be very high. For example, the confidence for

is 80% even though {p, q} is a cross-support pattern. The fact that the
cross-support pattern can produce a high confidence rule should not come as
a surprise because one of its items (p) appears very frequently in the data.
Therefore, p is expected to appear in many of the transactions that contain q.
Meanwhile, the rule also has high confidence even though {q, r} is not
a cross-support pattern. This example demonstrates the difficulty of using the
confidence measure to distinguish between rules extracted from cross-support
patterns and interesting patterns involving strongly connected but low-support

Even though the rule has very high confidence, notice that the rule
has very low confidence because most of the transactions that contain p

do not contain q. In contrast, the rule , which is derived from {q, r}, has
very high confidence. This observation suggests that cross-support patterns
can be detected by examining the lowest confidence rule that can be
extracted from a given itemset. An approach for finding the rule with the
lowest confidence given an itemset can be described as follows.

1. Recall the following anti-monotone property of confidence:





{q}→{p} {p}


conf({i1i2}→{i3, i4, …, ik})≤conf({i1i2i3}→{i4, i5, …, ik}).

This property suggests that confidence never increases as we shift
more items from the left- to the right-hand side of an association rule.
Because of this property, the lowest confidence rule extracted from a
frequent itemset contains only one item on its left-hand side. We
denote the set of all rules with only one item on its left-hand side as .

2. Given a frequent itemset , the rule

has the lowest confidence in if .This
follows directly from the definition of confidence as the ratio between
the rule’s support and the support of the rule antecedent. Hence, the
confidence of a rule will be lowest when the support of the antecedent
is highest.

3. Summarizing the previous points, the lowest confidence attainable from
a frequent itemset is

This expression is also known as the h-confidence or all-confidence
measure. Because of the anti-monotone property of support, the
numerator of the h-confidence measure is bounded by the minimum
support of any item that appears in the frequent itemset. In other
words, the h-confidence of an itemset must not exceed
the following expression:

Note that the upper bound of h-confidence in the above equation is exactly
same as support ratio (r) given in Equation 5.12 . Because the support ratio
for a cross-support pattern is always less than , the h-confidence of the
pattern is also guaranteed to be less than . Therefore, cross-support
patterns can be eliminated by ensuring that the h-confidence values for the
patterns exceed . As a final note, the advantages of using h-confidence go

{i1, i2, …, ik}

{ij}→{i1, i2, …, ij−1, ij+1, …,ik}

R1 s(ij)=max[s(i1), s(i2), …, s(ik)]

{i1, i2, …, ik}
s({i1, i2, …, ik})max[s(i1), s(i2), …, s(ik)].

X={i1, i2, …, ik}

h-confidence(X)≤min[s(i1), s(i2), …, s(ik)]max[s(i1), s(i2), …, s(ik)].



beyond eliminating cross-support patterns. The measure is also anti-
monotone, i.e.,

and thus can be incorporated directly into the mining algorithm. Furthermore,
h-confidence ensures that the items contained in an itemset are strongly
associated with each other. For example, suppose the h-confidence of an
itemset X is 80%. If one of the items in X is present in a transaction, there is at
least an 80% chance that the rest of the items in X also belong to the same
transaction. Such strongly associated patterns involving low-support items are
called hyperclique patterns.

Definition 5.10. (Hyperclique Pattern.)
An itemset X is a hyperclique pattern if h-confidence ,
where is a user-specified threshold.

h-confidence({i1, i2, …, ik})≥h-confidence({i1, i2, …, ik+1}),


5.9 Bibliographic Notes
The association rule mining task was first introduced by Agrawal et al. [324,
325] to discover interesting relationships among items in market basket
transactions. Since its inception, extensive research has been conducted to
address the various issues in association rule mining, from its fundamental
concepts to its implementation and applications. Figure 5.31 shows a
taxonomy of the various research directions in this area, which is generally
known as association analysis. As much of the research focuses on finding
patterns that appear significantly often in the data, the area is also known as
frequent pattern mining. A detailed review on some of the research topics in
this area can be found in [362] and in [319].

Figure 5.31.
An overview of the various research directions in association analysis.

Conceptual Issues

Research on the conceptual issues of association analysis has focused on
developing a theoretical formulation of association analysis and extending the
formulation to new types of patterns and going beyond asymmetric binary

Following the pioneering work by Agrawal et al. [324, 325], there has been a
vast amount of research on developing a theoretical formulation for the
association analysis problem. In [357], Gunopoulos et al. showed the
connection between finding maximal frequent itemsets and the hypergraph
transversal problem. An upper bound on the complexity of the association
analysis task was also derived. Zaki et al. [454, 456] and Pasquier et al. [407]
have applied formal concept analysis to study the frequent itemset generation
problem. More importantly, such research has led to the development of a
class of patterns known as closed frequent itemsets [456]. Friedman et al.
[355] have studied the association analysis problem in the context of bump
hunting in multidimensional space. Specifically, they consider frequent
itemset generation as the task of finding high density regions in
multidimensional space. Formalizing association analysis in a statistical
learning framework is another active research direction [414, 435, 444] as it
can help address issues related to identifying statistically significant patterns
and dealing with uncertain data [320, 333, 343].

Over the years, the association rule mining formulation has been expanded to
encompass other rule-based patterns, such as, profile association rules [321],
cyclic association rules [403], fuzzy association rules [379], exception rules
[431], negative association rules [336, 418], weighted association rules [338,
413], dependence rules [422], peculiar rules[462], inter-transaction
association rules [353, 440], and partial classification rules [327, 397].
Additionally, the concept of frequent itemset has been extended to other types
of patterns including closed itemsets [407, 456], maximal itemsets [330],
hyperclique patterns [449], support envelopes [428], emerging patterns [347],

contrast sets [329], high-utility itemsets [340, 390], approximate or error-
tolerant item-sets [358, 389, 451], and discriminative patterns [352, 401, 430].
Association analysis techniques have also been successfully applied to
sequential [326, 426], spatial [371], and graph-based [374, 380, 406, 450,
455] data.

Substantial research has been conducted to extend the original association
rule formulation to nominal [425], ordinal [392], interval [395], and ratio [356,
359, 425, 443, 461] attributes. One of the key issues is how to define the
support measure for these attributes. A methodology was proposed by
Steinbach et al. [429] to extend the traditional notion of support to more
general patterns and attribute types.

Implementation Issues
Research activities in this area revolve around (1) integrating the mining
capability into existing database technology, (2) developing efficient and
scalable mining algorithms, (3) handling user-specified or domain-specific
constraints, and (4) post-processing the extracted patterns.

There are several advantages to integrating association analysis into existing
database technology. First, it can make use of the indexing and query
processing capabilities of the database system. Second, it can also exploit the
DBMS support for scalability, check-pointing, and parallelization [415]. The
SETM algorithm developed by Houtsma et al. [370] was one of the earliest
algorithms to support association rule discovery via SQL queries. Since then,
numerous methods have been developed to provide capabilities for mining
association rules in database systems. For example, the DMQL [363] and M-
SQL [373] query languages extend the basic SQL with new operators for
mining association rules. The Mine Rule operator [394] is an expressive SQL
operator that can handle both clustered attributes and item hierarchies. Tsur et

al. [439] developed a generate-and-test approach called query flocks for
mining association rules. A distributed OLAP-based infrastructure was
developed by Chen et al. [341] for mining multilevel association rules.

Despite its popularity, the Apriori algorithm is computationally expensive
because it requires making multiple passes over the transaction database. Its
runtime and storage complexities were investigated by Dunkel and Soparkar
[349]. The FP-growth algorithm was developed by Han et al. in [364]. Other
algorithms for mining frequent itemsets include the DHP (dynamic hashing
and pruning) algorithm proposed by Park et al. [405] and the Partition
algorithm developed by Savasere et al [417]. A sampling-based frequent
itemset generation algorithm was proposed by Toivonen [436]. The algorithm
requires only a single pass over the data, but it can produce more candidate
item-sets than necessary. The Dynamic Itemset Counting (DIC) algorithm
[337] makes only 1.5 passes over the data and generates less candidate
itemsets than the sampling-based algorithm. Other notable algorithms include
the tree-projection algorithm [317] and H-Mine [408]. Survey articles on
frequent itemset generation algorithms can be found in [322, 367]. A
repository of benchmark data sets and software implementation of association
rule mining algorithms is available at the Frequent Itemset Mining
Implementations (FIMI) repository (http://fimi.cs.helsinki.fi).

Parallel algorithms have been developed to scale up association rule mining
for handling big data [318, 360, 399, 420, 457]. A survey of such algorithms
can be found in [453]. Online and incremental association rule mining
algorithms have also been proposed by Hidber [365] and Cheung et al. [342].
More recently, new algorithms have been developed to speed up frequent
itemset mining by exploiting the processing power of GPUs [459] and the
MapReduce/Hadoop distributed computing framework [382, 384, 396]. For
example, an implementation of frequent itemset mining for the Hadoop
framework is available in the Apache Mahout software .


1 http://mahout.apache.org

Srikant et al. [427] have considered the problem of mining association rules in
the presence of Boolean constraints such as the following:

Given such a constraint, the algorithm looks for rules that contain both cookies
and milk, or rules that contain the descendent items of cookies but not
ancestor items of wheat bread. Singh et al. [424] and Ng et al. [400] had also
developed alternative techniques for constrained-based association rule
mining. Constraints can also be imposed on the support for different itemsets.
This problem was investigated by Wang et al. [442], Liu et al. in [387], and
Seno et al. [419]. In addition, constraints arising from privacy concerns of
mining sensitive data have led to the development of privacy-preserving
frequent pattern mining techniques [334, 350, 441, 458].

One potential problem with association analysis is the large number of
patterns that can be generated by current algorithms. To overcome this
problem, methods to rank, summarize, and filter patterns have been
developed. Toivonen et al. [437] proposed the idea of eliminating redundant
rules using structural rule covers and grouping the remaining rules using
clustering. Liu et al. [388] applied the statistical chi-square test to prune
spurious patterns and summarized the remaining patterns using a subset of
the patterns called direction setting rules. The use of objective measures to
filter patterns has been investigated by many authors, including Brin et al.
[336], Bayardo and Agrawal [331], Aggarwal and Yu [323], and DuMouchel
and Pregibon[348]. The properties for many of these measures were analyzed
by Piatetsky-Shapiro [410], Kamber and Singhal [376], Hilderman and
Hamilton [366], and Tan et al. [433]. The grade-gender example used to
highlight the importance of the row and column scaling invariance property

(Cookies∧Milk)∨(descendants(Cookies)∧¬ancestors(Wheat Bread))

was heavily influenced by the discussion given in [398] by Mosteller.
Meanwhile, the tea-coffee example illustrating the limitation of confidence was
motivated by an example given in [336] by Brin et al. Because of the limitation
of confidence, Brin et al. [336] had proposed the idea of using interest factor
as a measure of interestingness. The all-confidence measure was proposed
by Omiecinski [402]. Xiong et al. [449] introduced the cross-support property
and showed that the all-confidence measure can be used to eliminate cross-
support patterns. A key difficulty in using alternative objective measures
besides support is their lack of a monotonicity property, which makes it difficult
to incorporate the measures directly into the mining algorithms. Xiong et al.
[447] have proposed an efficient method for mining correlations by introducing
an upper bound function to the . Although the measure is non-
monotone, it has an upper bound expression that can be exploited for the
efficient mining of strongly correlated item pairs.

Fabris and Freitas [351] have proposed a method for discovering interesting
associations by detecting the occurrences of Simpson’s paradox [423].
Megiddo and Srikant [393] described an approach for validating the extracted
patterns using hypothesis testing methods. A resampling-based technique
was also developed to avoid generating spurious patterns because of the
multiple comparison problem. Bolton et al. [335] have applied the Benjamini-
Hochberg [332] and Bonferroni correction methods to adjust the p-values of
discovered patterns in market basket data. Alternative methods for handling
the multiple comparison problem were suggested by Webb [445], Zhang et al.
[460], and Llinares-Lopez et al. [391].

Application of subjective measures to association analysis has been
investigated by many authors. Silberschatz and Tuzhilin [421] presented two
principles in which a rule can be considered interesting from a subjective point
of view. The concept of unexpected condition rules was introduced by Liu et
al. in [385]. Cooley et al. [344] analyzed the idea of combining soft belief sets


using the Dempster-Shafer theory and applied this approach to identify
contradictory and novel association patterns in web data. Alternative
approaches include using Bayesian networks [375] and neighborhood-based
information [346] to identify subjectively interesting patterns.

Visualization also helps the user to quickly grasp the underlying structure of
the discovered patterns. Many commercial data mining tools display the
complete set of rules (which satisfy both support and confidence threshold
criteria) as a two-dimensional plot, with each axis corresponding to the
antecedent or consequent itemsets of the rule. Hofmann et al. [368] proposed
using Mosaic plots and Double Decker plots to visualize association rules.
This approach can visualize not only a particular rule, but also the overall
contingency table between itemsets in the antecedent and consequent parts
of the rule. Nevertheless, this technique assumes that the rule consequent
consists of only a single attribute.

Application Issues
Association analysis has been applied to a variety of application domains
such as web mining [409, 432], document analysis [369], telecommunication
alarm diagnosis [377], network intrusion detection [328, 345, 381], and
bioinformatics [416, 446]. Applications of association and correlation pattern
analysis to Earth Science studies have been investigated in [411, 412, 434].
Trajectory pattern mining [339, 372, 438] is another application of spatio-
temporal association analysis to identify frequently traversed paths of moving

Association patterns have also been applied to other learning problems such
as classification [383, 386], regression [404], and clustering [361, 448, 452]. A
comparison between classification and association rule mining was made by
Freitas in his position paper [354]. The use of association patterns for

clustering has been studied by many authors including Han et al.[361],
Kosters et al. [378], Yang et al. [452] and Xiong et al. [448].

[317] R. C. Agarwal, C. C. Aggarwal, and V. V. V. Prasad. A Tree Projection

Algorithm for Generation of Frequent Itemsets. Journal of Parallel and
Distributed Computing (Special Issue on High Performance Data Mining),
61(3):350–371, 2001.

[318] R. C. Agarwal and J. C. Shafer. Parallel Mining of Association Rules.
IEEE Transactions on Knowledge and Data Engineering, 8(6):962–969,
March 1998.

[319] C. Aggarwal and J. Han. Frequent Pattern Mining. Springer, 2014.

[320] C. C. Aggarwal, Y. Li, J. Wang, and J. Wang. Frequent pattern mining
with uncertain data. In Proceedings of the 15th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pages 29–38,
Paris, France, 2009.

[321] C. C. Aggarwal, Z. Sun, and P. S. Yu. Online Generation of Profile
Association Rules. In Proc. of the 4th Intl. Conf. on Knowledge Discovery
and Data Mining, pages 129— 133, New York, NY, August 1996.

[322] C. C. Aggarwal and P. S. Yu. Mining Large Itemsets for Association
Rules. Data Engineering Bulletin, 21(1):23–31, March 1998.

[323] C. C. Aggarwal and P. S. Yu. Mining Associations with the Collective
Strength Approach. IEEE Trans. on Knowledge and Data Engineering,
13(6):863–873, January/February 2001.

[324] R. Agrawal, T. Imielinski, and A. Swami. Database mining: A
performance perspective. IEEE Transactions on Knowledge and Data
Engineering, 5:914–925, 1993.

[325] R. Agrawal, T. Imielinski, and A. Swami. Mining association rules
between sets of items in large databases. In Proc. ACM SIGMOD Intl.
Conf. Management of Data, pages 207–216, Washington, DC, 1993.

[326] R. Agrawal and R. Srikant. Mining Sequential Patterns. In Proc. of Intl.
Conf. on Data Engineering, pages 3–14, Taipei, Taiwan, 1995.

[327] K. Ali, S. Manganaris, and R. Srikant. Partial Classification using
Association Rules. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery
and Data Mining, pages 115— 118, Newport Beach, CA, August 1997.

[328] D. Barbará, J. Couto, S. Jajodia, and N. Wu. ADAM: A Testbed for
Exploring the Use of Data Mining in Intrusion Detection. SIGMOD Record,
30(4):15–24, 2001.

[329] S. D. Bay and M. Pazzani. Detecting Group Differences: Mining Contrast
Sets. Data Mining and Knowledge Discovery, 5(3):213–246, 2001.

[330] R. Bayardo. Efficiently Mining Long Patterns from Databases. In Proc. of
1998 ACM-SIGMOD Intl. Conf. on Management of Data, pages 85–93,
Seattle, WA, June 1998.

[331] R. Bayardo and R. Agrawal. Mining the Most Interesting Rules. In Proc.
of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 145–
153, San Diego, CA, August 1999.

[332] Y. Benjamini and Y. Hochberg. Controlling the False Discovery Rate: A
Practical and Powerful Approach to Multiple Testing. Journal Royal
Statistical Society B, 57 (1):289–300, 1995.

[333] T. Bernecker, H. Kriegel, M. Renz, F. Verhein, and A. Züle. Probabilistic
frequent itemset mining in uncertain databases. In Proceedings of the 15th
ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, pages 119–128, Paris, France, 2009.

[334] R. Bhaskar, S. Laxman, A. D. Smith, and A. Thakurta. Discovering
frequent patterns in sensitive data. In Proceedings of the 16th ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 503–512, Washington, DC, 2010.

[335] R. J. Bolton, D. J. Hand, and N. M. Adams. Determining Hit Rate in
Pattern Search. In Proc. of the ESF Exploratory Workshop on Pattern
Detection and Discovery in Data Mining, pages 36–48, London, UK,
September 2002.

[336] S. Brin, R. Motwani, and C. Silverstein. Beyond market baskets:
Generalizing association rules to correlations. In Proc. ACM SIGMOD Intl.
Conf. Management of Data, pages 265–276, Tucson, AZ, 1997.

[337] S. Brin, R. Motwani, J. Ullman, and S. Tsur. Dynamic Itemset Counting
and Implication Rules for market basket data. In Proc. of 1997 ACM-
SIGMOD Intl. Conf. on Management of Data, pages 255–264, Tucson, AZ,
June 1997.

[338] C. H. Cai, A. Fu, C. H. Cheng, and W. W. Kwong. Mining Association
Rules with Weighted Items. In Proc. of IEEE Intl. Database Engineering
and Applications Symp., pages 68–77, Cardiff, Wales, 1998.

[339] H. Cao, N. Mamoulis, and D. W. Cheung. Mining Frequent Spatio-
Temporal Sequential Patterns. In Proceedings of the 5th IEEE International
Conference on Data Mining, pages 82–89, Houston, TX, 2005.

[340] R. Chan, Q. Yang, and Y. Shen. Mining High Utility Itemsets. In
Proceedings of the 3rd IEEE International Conference on Data Mining,
pages 19–26, Melbourne, FL, 2003.

[341] Q. Chen, U. Dayal, and M. Hsu. A Distributed OLAP infrastructure for E-
Commerce. In Proc. of the 4th IFCIS Intl. Conf. on Cooperative Information
Systems, pages 209— 220, Edinburgh, Scotland, 1999.

[342] D. C. Cheung, S. D. Lee, and B. Kao. A General Incremental Technique
for Maintaining Discovered Association Rules. In Proc. of the 5th Intl. Conf.

on Database Systems for Advanced Applications, pages 185–194,
Melbourne, Australia, 1997.

[343] C. K. Chui, B. Kao, and E. Hung. Mining Frequent Itemsets from
Uncertain Data. In Proceedings of the 11th Pacific-Asia Conference on
Knowledge Discovery and Data Mining, pages 47–58, Nanjing, China,

[344] R. Cooley, P. N. Tan, and J. Srivastava. Discovery of Interesting Usage
Patterns from Web Data. In M. Spiliopoulou and B. Masand, editors,
Advances in Web Usage Analysis and User Profiling, volume 1836, pages
163–182. Lecture Notes in Computer Science, 2000.

[345] P. Dokas, L. Ertöz, V. Kumar, A. Lazarevic, J. Srivastava, and P. N. Tan.
Data Mining for Network Intrusion Detection. In Proc. NSF Workshop on
Next Generation Data Mining, Baltimore, MD, 2002.

[346] G. Dong and J. Li. Interestingness of discovered association rules in
terms of neighborhood-based unexpectedness. In Proc. of the 2nd Pacific-
Asia Conf. on Knowledge Discovery and Data Mining, pages 72–86,
Melbourne, Australia, April 1998.

[347] G. Dong and J. Li. Efficient Mining of Emerging Patterns: Discovering
Trends and Differences. In Proc. of the 5th Intl. Conf. on Knowledge
Discovery and Data Mining, pages 43–52, San Diego, CA, August 1999.

[348] W. DuMouchel and D. Pregibon. Empirical Bayes Screening for Multi-
Item Associations. In Proc. of the 7th Intl. Conf. on Knowledge Discovery

and Data Mining, pages 67–76, San Francisco, CA, August 2001.

[349] B. Dunkel and N. Soparkar. Data Organization and Access for Efficient
Data Mining. In Proc. of the 15th Intl. Conf. on Data Engineering, pages
522–529, Sydney, Australia, March 1999.

[350] A. V. Evfimievski, R. Srikant, R. Agrawal, and J. Gehrke. Privacy
preserving mining of association rules. In Proceedings of the Eighth ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 217–228, Edmonton, Canada, 2002.

[351] C. C. Fabris and A. A. Freitas. Discovering surprising patterns by
detecting occurrences of Simpson’s paradox. In Proc. of the 19th SGES
Intl. Conf. on Knowledge-Based Systems and Applied Artificial
Intelligence), pages 148–160, Cambridge, UK, December 1999.

[352] G. Fang, G. Pandey, W. Wang, M. Gupta, M. Steinbach, and V. Kumar.
Mining Low-Support Discriminative Patterns from Dense and High-
Dimensional Data. IEEE Trans. Knowl. Data Eng., 24(2):279–294, 2012.

[353] L. Feng, H. J. Lu, J. X. Yu, and J. Han. Mining inter-transaction
associations with templates. In Proc. of the 8th Intl. Conf. on Information
and Knowledge Management, pages 225–233, Kansas City, Missouri, Nov

[354] A. A. Freitas. Understanding the crucial differences between
classification and discovery of association rules—a position paper.
SIGKDD Explorations, 2(1):65–69, 2000.

[355] J. H. Friedman and N. I. Fisher. Bump hunting in high-dimensional data.
Statistics and Computing, 9(2):123–143, April 1999.

[356] T. Fukuda, Y. Morimoto, S. Morishita, and T. Tokuyama. Mining
Optimized Association Rules for Numeric Attributes. In Proc. of the 15th
Symp. on Principles of Database Systems, pages 182–191, Montreal,
Canada, June 1996.

[357] D. Gunopulos, R. Khardon, H. Mannila, and H. Toivonen. Data Mining,
Hypergraph Transversals, and Machine Learning. In Proc. of the 16th
Symp. on Principles of Database Systems, pages 209–216, Tucson, AZ,
May 1997.

[358] R. Gupta, G. Fang, B. Field, M. Steinbach, and V. Kumar. Quantitative
evaluation of approximate frequent pattern mining algorithms. In
Proceedings of the 14th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pages 301–309, Las Vegas, NV,

[359] E. Han, G. Karypis, and V. Kumar. Min-apriori: An algorithm for finding
association rules in data with continuous attributes. Department of
Computer Science and Engineering, University of Minnesota, Tech. Rep,

[360] E.-H. Han, G. Karypis, and V. Kumar. Scalable Parallel Data Mining for
Association Rules. In Proc. of 1997 ACM-SIGMOD Intl. Conf. on
Management of Data, pages 277–288, Tucson, AZ, May 1997.

[361] E.-H. Han, G. Karypis, V. Kumar, and B. Mobasher. Clustering Based on
Association Rule Hypergraphs. In Proc. of the 1997 ACM SIGMOD
Workshop on Research Issues in Data Mining and Knowledge Discovery,
Tucson, AZ, 1997.

[362] J. Han, H. Cheng, D. Xin, and X. Yan. Frequent pattern mining: current
status and future directions. Data Mining and Knowledge Discovery,
15(1):55–86, 2007.

[363] J. Han, Y. Fu, K. Koperski, W. Wang, and O. R. Zaïane. DMQL: A data
mining query language for relational databases. In Proc. of the 1996 ACM
SIGMOD Workshop on Research Issues in Data Mining and Knowledge
Discovery, Montreal, Canada, June 1996.

[364] J. Han, J. Pei, and Y. Yin. Mining Frequent Patterns without Candidate
Generation. In Proc. ACM-SIGMOD Int. Conf. on Management of Data
(SIGMOD’00), pages 1–12, Dallas, TX, May 2000.

[365] C. Hidber. Online Association Rule Mining. In Proc. of 1999 ACM-
SIGMOD Intl. Conf. on Management of Data, pages 145–156, Philadelphia,
PA, 1999.

[366] R. J. Hilderman and H. J. Hamilton. Knowledge Discovery and
Measures of Interest. Kluwer Academic Publishers, 2001.

[367] J. Hipp, U. Guntzer, and G. Nakhaeizadeh. Algorithms for Association
Rule Mining— A General Survey. SigKDD Explorations, 2(1):58–64, June


[368] H. Hofmann, A. P. J. M. Siebes, and A. F. X. Wilhelm. Visualizing
Association Rules with Interactive Mosaic Plots. In Proc. of the 6th Intl.
Conf. on Knowledge Discovery and Data Mining, pages 227–235, Boston,
MA, August 2000.

[369] J. D. Holt and S. M. Chung. Efficient Mining of Association Rules in Text
Databases. In Proc. of the 8th Intl. Conf. on Information and Knowledge
Management, pages 234–242, Kansas City, Missouri, 1999.

[370] M. Houtsma and A. Swami. Set-oriented Mining for Association Rules in
Relational Databases. In Proc. of the 11th Intl. Conf. on Data Engineering,
pages 25–33, Taipei, Taiwan, 1995.

[371] Y. Huang, S. Shekhar, and H. Xiong. Discovering Co-location Patterns
from SpatialDatasets: A General Approach. IEEE Trans. on Knowledge
and Data Engineering, 16 (12):1472–1485, December 2004.

[372] S. Hwang, Y. Liu, J. Chiu, and E. Lim. Mining Mobile Group Patterns: A
Trajectory-Based Approach. In Proceedings of the 9th Pacific-Asia
Conference on Knowledge Discovery and Data Mining, pages 713–718,
Hanoi, Vietnam, 2005.

[373] T. Imielinski, A. Virmani, and A. Abdulghani. DataMine: Application
Programming Interface and Query Language for Database Mining. In Proc.
of the 2nd Intl. Conf. on Knowledge Discovery and Data Mining, pages
256–262, Portland, Oregon, 1996.

[374] A. Inokuchi, T. Washio, and H. Motoda. An Apriori-based Algorithm for
Mining Frequent Substructures from Graph Data. In Proc. of the 4th
European Conf. of Principles and Practice of Knowledge Discovery in
Databases, pages 13–23, Lyon, France, 2000.

[375] S. Jaroszewicz and D. Simovici. Interestingness of Frequent Itemsets
Using Bayesian Networks as Background Knowledge. In Proc. of the 10th
Intl. Conf. on Knowledge Discovery and Data Mining, pages 178–186,
Seattle, WA, August 2004.

[376] M. Kamber and R. Shinghal. Evaluating the Interestingness of
Characteristic Rules. In Proc. of the 2nd Intl. Conf. on Knowledge
Discovery and Data Mining, pages 263–266, Portland, Oregon, 1996.

[377] M. Klemettinen. A Knowledge Discovery Methodology for
Telecommunication Network Alarm Databases. PhD thesis, University of
Helsinki, 1999.

[378] W. A. Kosters, E. Marchiori, and A. Oerlemans. Mining Clusters with
Association Rules. In The 3rd Symp. on Intelligent Data Analysis (IDA99),
pages 39–50, Amsterdam, August 1999.

[379] C. M. Kuok, A. Fu, and M. H. Wong. Mining Fuzzy Association Rules in
Databases. ACM SIGMOD Record, 27(1):41–46, March 1998.

[380] M. Kuramochi and G. Karypis. Frequent Subgraph Discovery. In Proc. of
the 2001 IEEE Intl. Conf. on Data Mining, pages 313–320, San Jose, CA,

November 2001.

[381] W. Lee, S. J. Stolfo, and K. W. Mok. Adaptive Intrusion Detection: A
Data Mining Approach. Artificial Intelligence Review, 14(6):533–567, 2000.

[382] N. Li, L. Zeng, Q. He, and Z. Shi. Parallel Implementation of Apriori
Algorithm Based on MapReduce. In Proceedings of the 13th ACIS
International Conference on Software Engineering, Artificial Intelligence,
Networking and Parallel/Distributed Computing, pages 236–241, Kyoto,
Japan, 2012.

[383] W. Li, J. Han, and J. Pei. CMAR: Accurate and Efficient Classification
Based on Multiple Class-association Rules. In Proc. of the 2001 IEEE Intl.
Conf. on Data Mining, pages 369–376, San Jose, CA, 2001.

[384] M. Lin, P. Lee, and S. Hsueh. Apriori-based frequent itemset mining
algorithms on MapReduce. In Proceedings of the 6th International
Conference on Ubiquitous Information Management and Communication,
pages 26–30, Kuala Lumpur, Malaysia, 2012.

[385] B. Liu, W. Hsu, and S. Chen. Using General Impressions to Analyze
Discovered Classification Rules. In Proc. of the 3rd Intl. Conf. on
Knowledge Discovery and Data Mining, pages 31–36, Newport Beach, CA,
August 1997.

[386] B. Liu, W. Hsu, and Y. Ma. Integrating Classification and Association
Rule Mining. In Proc. of the 4th Intl. Conf. on Knowledge Discovery and
Data Mining, pages 80–86, New York, NY, August 1998.

[387] B. Liu, W. Hsu, and Y. Ma. Mining association rules with multiple
minimum supports. In Proc. of the 5th Intl. Conf. on Knowledge Discovery
and Data Mining, pages 125— 134, San Diego, CA, August 1999.

[388] B. Liu, W. Hsu, and Y. Ma. Pruning and Summarizing the Discovered
Associations. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and
Data Mining, pages 125–134, San Diego, CA, August 1999.

[389] J. Liu, S. Paulsen, W. Wang, A. B. Nobel, and J. Prins. Mining
Approximate Frequent Itemsets from Noisy Data. In Proceedings of the 5th
IEEE International Conference on Data Mining, pages 721–724, Houston,
TX, 2005.

[390] Y. Liu, W.-K. Liao, and A. Choudhary. A two-phase algorithm for fast
discovery of high utility itemsets. In Proceedings of the 9th Pacific-Asia
Conference on Knowledge Discovery and Data Mining, pages 689–695,
Hanoi, Vietnam, 2005.

[391] F. Llinares-López, M. Sugiyama, L. Papaxanthos, and K. M. Borgwardt.
Fast and Memory-Efficient Significant Pattern Mining via Permutation
Testing. In Proceedings of the 21th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pages 725–734,
Sydney, Australia, 2015.

[392] A. Marcus, J. I. Maletic, and K.-I. Lin. Ordinal association rules for error
identification in data sets. In Proc. of the 10th Intl. Conf. on Information and
Knowledge Management, pages 589–591, Atlanta, GA, October 2001.

[393] N. Megiddo and R. Srikant. Discovering Predictive Association Rules. In
Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining,
pages 274–278, New York, August 1998.

[394] R. Meo, G. Psaila, and S. Ceri. A New SQL-like Operator for Mining
Association Rules. In Proc. of the 22nd VLDB Conf., pages 122–133,
Bombay, India, 1996.

[395] R. J. Miller and Y. Yang. Association Rules over Interval Data. In Proc. of
1997 ACM-SIGMOD Intl. Conf. on Management of Data, pages 452–461,
Tucson, AZ, May 1997.

[396] S. Moens, E. Aksehirli, and B. Goethals. Frequent Itemset Mining for Big
Data. In Proceedings of the 2013 IEEE International Conference on Big
Data, pages 111–118, Santa Clara, CA, 2013.

[397] Y. Morimoto, T. Fukuda, H. Matsuzawa, T. Tokuyama, and K. Yoda.
Algorithms for mining association rules for binary segmentations of huge
categorical databases. In Proc. of the 24th VLDB Conf., pages 380–391,
New York, August 1998.

[398] F. Mosteller. Association and Estimation in Contingency Tables. JASA,
63:1–28, 1968.

[399] A. Mueller. Fast sequential and parallel algorithms for association rule
mining: A comparison. Technical Report CS-TR-3515, University of
Maryland, August 1995.

[400] R. T. Ng, L. V. S. Lakshmanan, J. Han, and A. Pang. Exploratory Mining
and Pruning Optimizations of Constrained Association Rules. In Proc. of
1998 ACM-SIGMOD Intl. Conf. on Management of Data, pages 13–24,
Seattle, WA, June 1998.

[401] P. K. Novak, N. Lavrač, and G. I. Webb. Supervised descriptive rule
discovery: A unifying survey of contrast set, emerging pattern and
subgroup mining. Journal of Machine Learning Research, 10(Feb):377–
403, 2009.

[402] E. Omiecinski. Alternative Interest Measures for Mining Associations in
Databases. IEEE Trans. on Knowledge and Data Engineering, 15(1):57–
69, January/February 2003.

[403] B. Ozden, S. Ramaswamy, and A. Silberschatz. Cyclic Association
Rules. In Proc. of the 14th Intl. Conf. on Data Eng., pages 412–421,
Orlando, FL, February 1998.

[404] A. Ozgur, P. N. Tan, and V. Kumar. RBA: An Integrated Framework for
Regression based on Association Rules. In Proc. of the SIAM Intl. Conf. on
Data Mining, pages 210–221, Orlando, FL, April 2004.

[405] J. S. Park, M.-S. Chen, and P. S. Yu. An effective hash-based algorithm
for mining association rules. SIGMOD Record, 25(2):175–186, 1995.

[406] S. Parthasarathy and M. Coatney. Efficient Discovery of Common
Substructures in Macromolecules. In Proc. of the 2002 IEEE Intl. Conf. on

Data Mining, pages 362— 369, Maebashi City, Japan, December 2002.

[407] N. Pasquier, Y. Bastide, R. Taouil, and L. Lakhal. Discovering frequent
closed itemsets for association rules. In Proc. of the 7th Intl. Conf. on
Database Theory (ICDT’99), pages 398–416, Jerusalem, Israel, January

[408] J. Pei, J. Han, H. J. Lu, S. Nishio, and S. Tang. H-Mine: Hyper-Structure
Mining of Frequent Patterns in Large Databases. In Proc. of the 2001 IEEE
Intl. Conf. on Data Mining, pages 441–448, San Jose, CA, November

[409] J. Pei, J. Han, B. Mortazavi-Asl, and H. Zhu. Mining Access Patterns
Efficiently from Web Logs. In Proc. of the 4th Pacific-Asia Conf. on
Knowledge Discovery and Data Mining, pages 396–407, Kyoto, Japan,
April 2000.

[410] G. Piatetsky-Shapiro. Discovery, Analysis and Presentation of Strong
Rules. In G. Piatetsky-Shapiro and W. Frawley, editors, Knowledge
Discovery in Databases, pages 229–248. MIT Press, Cambridge, MA,

[411] C. Potter, S. Klooster, M. Steinbach, P. N. Tan, V. Kumar, S. Shekhar,
and C. Carvalho. Understanding Global Teleconnections of Climate to
Regional Model Estimates of Amazon Ecosystem Carbon Fluxes. Global
Change Biology, 10(5):693— 703, 2004.

[412] C. Potter, S. Klooster, M. Steinbach, P. N. Tan, V. Kumar, S. Shekhar, R.
Myneni, and R. Nemani. Global Teleconnections of Ocean Climate to
Terrestrial Carbon Flux. Journal of Geophysical Research, 108(D17),

[413] G. D. Ramkumar, S. Ranka, and S. Tsur. Weighted association rules:
Model and algorithm. In Proc. ACM SIGKDD, 1998.

[414] M. Riondato and F. Vandin. Finding the True Frequent Itemsets. In
Proceedings of the 2014 SIAM International Conference on Data Mining,
pages 497–505, Philadelphia, PA, 2014.

[415] S. Sarawagi, S. Thomas, and R. Agrawal. Integrating Mining with
Relational Database Systems: Alternatives and Implications. In Proc. of
1998 ACM-SIGMOD Intl. Conf. on Management of Data, pages 343–354,
Seattle, WA, 1998.

[416] K. Satou, G. Shibayama, T. Ono, Y. Yamamura, E. Furuichi, S. Kuhara,
and T. Takagi. Finding Association Rules on Heterogeneous Genome
Data. In Proc. of the Pacific Symp. on Biocomputing, pages 397–408,
Hawaii, January 1997.

[417] A. Savasere, E. Omiecinski, and S. Navathe. An efficient algorithm for
mining association rules in large databases. In Proc. of the 21st Int. Conf.
on Very Large Databases (VLDB’95), pages 432–444, Zurich, Switzerland,
September 1995.

[418] A. Savasere, E. Omiecinski, and S. Navathe. Mining for Strong Negative
Associations in a Large Database of Customer Transactions. In Proc. of
the 14th Intl. Conf. on Data Engineering, pages 494–502, Orlando, Florida,
February 1998.

[419] M. Seno and G. Karypis. LPMiner: An Algorithm for Finding Frequent
Itemsets Using Length-Decreasing Support Constraint. In Proc. of the
2001 IEEE Intl. Conf. on Data Mining, pages 505–512, San Jose, CA,
November 2001.

[420] T. Shintani and M. Kitsuregawa. Hash based parallel algorithms for
mining association rules. In Proc of the 4th Intl. Conf. on Parallel and
Distributed Info. Systems, pages 19–30, Miami Beach, FL, December

[421] A. Silberschatz and A. Tuzhilin. What makes patterns interesting in
knowledge discovery systems. IEEE Trans. on Knowledge and Data
Engineering, 8(6):970–974, 1996.

[422] C. Silverstein, S. Brin, and R. Motwani. Beyond market baskets:
Generalizing association rules to dependence rules. Data Mining and
Knowledge Discovery, 2(1): 39–68, 1998.

[423] E.-H. Simpson. The Interpretation of Interaction in Contingency Tables.
Journal of the Royal Statistical Society, B(13):238–241, 1951.

[424] L. Singh, B. Chen, R. Haight, and P. Scheuermann. An Algorithm for
Constrained Association Rule Mining in Semi-structured Data. In Proc. of

the 3rd Pacific-Asia Conf. on Knowledge Discovery and Data Mining,
pages 148–158, Beijing, China, April 1999.

[425] R. Srikant and R. Agrawal. Mining Quantitative Association Rules in
Large Relational Tables. In Proc. of 1996 ACM-SIGMOD Intl. Conf. on
Management of Data, pages 1–12, Montreal, Canada, 1996.

[426] R. Srikant and R. Agrawal. Mining Sequential Patterns: Generalizations
and Performance Improvements. In Proc. of the 5th Intl Conf. on Extending
Database Technology (EDBT’96), pages 18–32, Avignon, France, 1996.

[427] R. Srikant, Q. Vu, and R. Agrawal. Mining Association Rules with Item
Constraints. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and
Data Mining, pages 67–73, Newport Beach, CA, August 1997.

[428] M. Steinbach, P. N. Tan, and V. Kumar. Support Envelopes: A Technique
for Exploring the Structure of Association Patterns. In Proc. of the 10th Intl.
Conf. on Knowledge Discovery and Data Mining, pages 296–305, Seattle,
WA, August 2004.

[429] M. Steinbach, P. N. Tan, H. Xiong, and V. Kumar. Extending the Notion
of Support. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and
Data Mining, pages 689–694, Seattle, WA, August 2004.

[430] M. Steinbach, H. Yu, G. Fang, and V. Kumar. Using constraints to
generate and explore higher order discriminative patterns. Advances in
Knowledge Discovery and Data Mining, pages 338–350, 2011.

[431] E. Suzuki. Autonomous Discovery of Reliable Exception Rules. In Proc.
of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 259–
262, Newport Beach, CA, August 1997.

[432] P. N. Tan and V. Kumar. Mining Association Patterns in Web Usage
Data. In Proc. of the Intl. Conf. on Advances in Infrastructure for e-
Business, e-Education, e-Science and e-Medicine on the Internet,
L’Aquila, Italy, January 2002.

[433] P. N. Tan, V. Kumar, and J. Srivastava. Selecting the Right
Interestingness Measure for Association Patterns. In Proc. of the 8th Intl.
Conf. on Knowledge Discovery and Data Mining, pages 32–41, Edmonton,
Canada, July 2002.

[434] P. N. Tan, M. Steinbach, V. Kumar, S. Klooster, C. Potter, and A.
Torregrosa. FindingSpatio-Temporal Patterns in Earth Science Data. In
KDD 2001 Workshop on Temporal Data Mining, San Francisco, CA, 2001.

[435] N. Tatti. Probably the best itemsets. In Proceedings of the 16th ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 293–302, Washington, DC, 2010.

[436] H. Toivonen. Sampling Large Databases for Association Rules. In Proc.
of the 22nd VLDB Conf., pages 134–145, Bombay, India, 1996.

[437] H. Toivonen, M. Klemettinen, P. Ronkainen, K. Hatonen, and H. Mannila.
Pruning and Grouping Discovered Association Rules. In ECML-95

Workshop on Statistics, Machine Learning and Knowledge Discovery in
Databases, pages 47–52, Heraklion, Greece, April 1995.

[438] I. Tsoukatos and D. Gunopulos. Efficient mining of spatiotemporal
patterns. In Proceedings of the 7th International Symposium on Advances
in Spatial and Temporal Databases, pages 425–442, 2001.

[439] S. Tsur, J. Ullman, S. Abiteboul, C. Clifton, R. Motwani, S. Nestorov, and
A. Rosenthal. Query Flocks: A Generalization of Association Rule Mining.
In Proc. of 1998 ACM-SIGMOD Intl. Conf. on Management of Data, pages
1–12, Seattle, WA, June 1998.

[440] A. Tung, H. J. Lu, J. Han, and L. Feng. Breaking the Barrier of
Transactions: Mining Inter-Transaction Association Rules. In Proc. of the
5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 297–301,
San Diego, CA, August 1999.

[441] J. Vaidya and C. Clifton. Privacy preserving association rule mining in
vertically partitioned data. In Proceedings of the Eighth ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pages
639–644, Edmonton, Canada, 2002.

[442] K. Wang, Y. He, and J. Han. Mining Frequent Itemsets Using Support
Constraints. In Proc. of the 26th VLDB Conf., pages 43–52, Cairo, Egypt,
September 2000.

[443] K. Wang, S. H. Tay, and B. Liu. Interestingness-Based Interval Merger
for Numeric Association Rules. In Proc. of the 4th Intl. Conf. on Knowledge

Discovery and Data Mining, pages 121–128, New York, NY, August 1998.

[444] L. Wang, R. Cheng, S. D. Lee, and D. W. Cheung. Accelerating
probabilistic frequent itemset mining: a model-based approach. In
Proceedings of the 19th ACM Conference on Information and Knowledge
Management, pages 429–438, 2010.

[445] G. I. Webb. Preliminary investigations into statistically valid exploratory
rule discovery. In Proc. of the Australasian Data Mining Workshop
(AusDM03), Canberra, Australia, December 2003.

[446] H. Xiong, X. He, C. Ding, Y. Zhang, V. Kumar, and S. R. Holbrook.
Identification of Functional Modules in Protein Complexes via Hyperclique
Pattern Discovery. In Proc. of the Pacific Symposium on Biocomputing,
(PSB 2005), Maui, January 2005.

[447] H. Xiong, S. Shekhar, P. N. Tan, and V. Kumar. Exploiting a Support-
based Upper Bound of Pearson’s Correlation Coefficient for Efficiently
Identifying Strongly Correlated Pairs. In Proc. of the 10th Intl. Conf. on
Knowledge Discovery and Data Mining, pages 334–343, Seattle, WA,
August 2004.

[448] H. Xiong, M. Steinbach, P. N. Tan, and V. Kumar. HICAP: Hierarchial
Clustering with Pattern Preservation. In Proc. of the SIAM Intl. Conf. on
Data Mining, pages 279–290, Orlando, FL, April 2004.

[449] H. Xiong, P. N. Tan, and V. Kumar. Mining Strong Affinity Association
Patterns in Data Sets with Skewed Support Distribution. In Proc. of the

2003 IEEE Intl. Conf. on Data Mining, pages 387–394, Melbourne, FL,

[450] X. Yan and J. Han. gSpan: Graph-based Substructure Pattern Mining. In
Proc. of the 2002 IEEE Intl. Conf. on Data Mining, pages 721–724,
Maebashi City, Japan, December 2002.

[451] C. Yang, U. M. Fayyad, and P. S. Bradley. Efficient discovery of error-
tolerant frequent itemsets in high dimensions. In Proceedings of the
seventh ACM SIGKDD international conference on Knowledge discovery
and data mining, pages 194–203, , San Francisco, CA, 2001.

[452] C. Yang, U. M. Fayyad, and P. S. Bradley. Efficient discovery of error-
tolerant frequent itemsets in high dimensions. In Proc. of the 7th Intl. Conf.
on Knowledge Discovery and Data Mining, pages 194–203, San
Francisco, CA, August 2001.

[453] M. J. Zaki. Parallel and Distributed Association Mining: A Survey. IEEE
Concurrency, special issue on Parallel Mechanisms for Data Mining,
7(4):14–25, December 1999.

[454] M. J. Zaki. Generating Non-Redundant Association Rules. In Proc. of
the 6th Intl. Conf. on Knowledge Discovery and Data Mining, pages 34–43,
Boston, MA, August 2000.

[455] M. J. Zaki. Efficiently mining frequent trees in a forest. In Proc. of the 8th
Intl. Conf. on Knowledge Discovery and Data Mining, pages 71–80,
Edmonton, Canada, July 2002.

[456] M. J. Zaki and M. Orihara. Theoretical foundations of association rules.
In Proc. of the 1998 ACM SIGMOD Workshop on Research Issues in Data
Mining and Knowledge Discovery, Seattle, WA, June 1998.

[457] M. J. Zaki, S. Parthasarathy, M. Ogihara, and W. Li. New Algorithms for
Fast Discovery of Association Rules. In Proc. of the 3rd Intl. Conf. on
Knowledge Discovery and Data Mining, pages 283–286, Newport Beach,
CA, August 1997.

[458] C. Zeng, J. F. Naughton, and J. Cai. On differentially private frequent
itemset mining. Proceedings of the VLDB Endowment, 6(1):25–36, 2012.

[459] F. Zhang, Y. Zhang, and J. Bakos. GPApriori: GPU-Accelerated
Frequent Itemset Mining. In Proceedings of the 2011 IEEE International
Conference on Cluster Computing, pages 590–594, Austin, TX, 2011.

[460] H. Zhang, B. Padmanabhan, and A. Tuzhilin. On the Discovery of
Significant Statistical Quantitative Rules. In Proc. of the 10th Intl. Conf. on
Knowledge Discovery and Data Mining, pages 374–383, Seattle, WA,
August 2004.

[461] Z. Zhang, Y. Lu, and B. Zhang. An Effective Partioning-Combining
Algorithm for Discovering Quantitative Association Rules. In Proc. of the
1st Pacific-Asia Conf. on Knowledge Discovery and Data Mining,
Singapore, 1997.

[462] N. Zhong, Y. Y. Yao, and S. Ohsuga. Peculiarity Oriented Multi-database
Mining. InProc. of the 3rd European Conf. of Principles and Practice of
Knowledge Discovery in Databases, pages 136–146, Prague, Czech
Republic, 1999.

5.10 Exercises
1. For each of the following questions, provide an example of an association
rule from the market basket domain that satisfies the following conditions.
Also, describe whether such rules are subjectively interesting.

a. A rule that has high support and high confidence.

b. A rule that has reasonably high support but low confidence.

c. A rule that has low support and low confidence.

d. A rule that has low support and high confidence.

2. Consider the data set shown in Table 5.20 .

Table 5.20. Example of market basket transactions.

Customer ID Transaction ID Items Bought

1 0001 {a, d, e}

1 0024 {a, b, c, e}

2 0012 {a, b, d, e}

2 0031 {a, c, d, e}

3 0015 {b, c e}

3 0022 {b, d, e}

4 0029 {c d}

4 0040 {a, b, c}

5 0033 {a, d, e}

5 0038 {a, b, e}

a. Compute the support for itemsets {e}, {b, d}, and {b, d, e} by treating each
transaction ID as a market basket.

b. Use the results in part (a) to compute the confidence for the association
rules and . Is confidence a symmetric measure?

c. Repeat part (a) by treating each customer ID as a market basket. Each
item should be treated as a binary variable (1 if an item appears in at
least one transaction bought by the customer, and 0 otherwise).

d. Use the results in part (c) to compute the confidence for the association
rules and .

e. Suppose and are the support and confidence values of an
association rule r when treating each transaction ID as a market basket.
Also, let and be the support and confidence values of r when
treating each customer ID as a market basket. Discuss whether there are
any relationships between and or and .


a. What is the confidence for the rules and ?

b. Let , and be the confidence values of the rules
, and , respectively. If we assume that , and

have different values, what are the possible relationships that may exist
among , and ? Which rule has the lowest confidence?

c. Repeat the analysis in part (b) assuming that the rules have identical
support. Which rule has the highest confidence?

{b, d}→{e} {e}→{b, d}

{b, d}→{e} {e}→{b, d}

s1 c1

s2 c2

s1 s2 c1 c2

∅→A A→∅
c1, c2 c3 {p}→{q}, {p}

→{q, r} {p, r}→{q} c1, c2 c3

c1, c2 c3

d. Transitivity: Suppose the confidence of the rules and are
larger than some threshold, minconf. Is it possible that has a
confidence less than minconf?

4. For each of the following measures, determine whether it is monotone, anti-
monotone, or non-monotone (i.e., neither monotone nor anti-monotone).

Example: Support, is anti-monotone because whenever

a. A characteristic rule is a rule of the form , where the
rule antecedent contains only a single item. An itemset of size k can
produce up to k characteristic rules. Let be the minimum confidence of
all characteristic rules generated from a given itemset:

Is monotone, anti-monotone, or non-monotone?

b. A discriminant rule is a rule of the form , where the
rule consequent contains only a single item. An itemset of size k can
produce up to k discriminant rules. Let be the minimum confidence of all
discriminant rules generated from a given itemset:

Is monotone, anti-monotone, or non-monotone?

c. Repeat the analysis in parts (a) and (b) by replacing the min function with
a max function.


s=σ(x)|T| s(X)≥s(Y)

{p}→{q1, q2, …, qn}


ζ({p1, p2, …, pk})=min[c({p1}→{p2, p3, …, pk}), …c({pk}→{p1, p2, …, pk


{p1, p2, …, pn}→{q}


η({p1, p2, …, pk})=min[c({p2, p3, …, pk}→{p1}), …c({p1, p2, …, pk−1}


5. Prove Equation 5.3 . (Hint: First, count the number of ways to create an
itemset that forms the left-hand side of the rule. Next, for each size k itemset
selected for the left-hand side, count the number of ways to choose the
remaining items to form the right-hand side of the rule.) Assume that
neither of the itemsets of a rule are empty.

6. Consider the market basket transactions shown in Table 5.21 .

a. What is the maximum number of association rules that can be extracted
from this data (including rules that have zero support)?

b. What is the maximum size of frequent itemsets that can be extracted
(assuming )?

Table 5.21. Market basket transactions.

Transaction ID Items Bought

1 {Milk, Beer, Diapers}

2 {Bread, Butter, Milk}

3 {Milk, Diapers, Cookies}

4 {Bread, Butter, Cookies}

5 {Beer, Cookies, Diapers}

6 {Milk, Diapers, Bread, Butter}

7 {Bread, Butter, Diapers}

8 {Beer, Diapers}

9 {Milk, Diapers, Bread, Butter}

10 {Beer, Cookies}



c. Write an expression for the maximum number of size-3 itemsets that can
be derived from this data set.

d. Find an itemset (of size 2 or larger) that has the largest support.

e. Find a pair of items, a and b, such that the rules and
have the same confidence.

7. Show that if a candidate k-itemset X has a subset of size less than that
is infrequent, then at least one of the -size subsets of X is necessarily

8. Consider the following set of frequent 3-itemsets:

{1, 2, 3}, {1, 2, 4}, {1, 2, 5}, {1, 3, 4}, {1, 3, 5}, {2, 3, 4}, {2, 3, 5}, {3, 4, 5}.

Assume that there are only five items in the data set.

a. List all candidate 4-itemsets obtained by a candidate generation
procedure using the merging strategy.

b. List all candidate 4-itemsets obtained by the candidate generation
procedure in Apriori.

c. List all candidate 4-itemsets that survive the candidate pruning step of the
Apriori algorithm.

9. The Apriori algorithm uses a generate-and-count strategy for deriving
frequent itemsets. Candidate itemsets of size are created by joining a pair
of frequent itemsets of size k (this is known as the candidate generation step).
A candidate is discarded if any one of its subsets is found to be infrequent
during the candidate pruning step. Suppose the Apriori algorithm is applied to
the data set shown in Table 5.22 with , i.e., any itemset
occurring in less than 3 transactions is considered to be infrequent.

{a}→{b} {b}→{a}





Table 5.22. Example of market basket transactions.

Transaction ID Items Bought

1 {a, b, d, e}

2 {b, c d}

3 {a, b, d, e}

4 {a, c, d, e}

5 {b, c, d, e}

6 {b, d, e}

7 {c, d}

8 {a, b, c}

9 {a, d, e}

10 {b, d}

a. Draw an itemset lattice representing the data set given in Table 5.22 .
Label each node in the lattice with the following letter(s):

N: If the itemset is not considered to be a candidate itemset by the
Apriori algorithm. There are two reasons for an itemset not to be
considered as a candidate itemset: (1) it is not generated at all during
the candidate generation step, or (2) it is generated during the
candidate generation step but is subsequently removed during the
candidate pruning step because one of its subsets is found to be

F: If the candidate itemset is found to be frequent by the Apriori

I: If the candidate itemset is found to be infrequent after support

b. What is the percentage of frequent itemsets (with respect to all itemsets
in the lattice)?

c. What is the pruning ratio of the Apriori algorithm on this data set?
(Pruning ratio is defined as the percentage of itemsets not considered to
be a candidate because (1) they are not generated during candidate
generation or (2) they are pruned during the candidate pruning step.)

d. What is the false alarm rate (i.e., percentage of candidate itemsets that
are found to be infrequent after performing support counting)?

10. The Apriori algorithm uses a hash tree data structure to efficiently count
the support of candidate itemsets. Consider the hash tree for candidate 3-
itemsets shown in Figure 5.32 .

Figure 5.32.
An example of a hash tree structure.

a. Given a transaction that contains items {1, 3, 4, 5, 8}, which of the hash
tree leaf nodes will be visited when finding the candidates of the

b. Use the visited leaf nodes in part (a) to determine the candidate itemsets
that are contained in the transaction {1, 3, 4, 5, 8}.

11. Consider the following set of candidate 3-itemsets:

{1, 2, 3}, {1, 2, 6}, {1, 3, 4}, {2, 3, 4}, {2, 4, 5}, {3, 4, 6}, {4, 5, 6}

a. Construct a hash tree for the above candidate 3-itemsets. Assume the
tree uses a hash function where all odd-numbered items are hashed to
the left child of a node, while the even-numbered items are hashed to the
right child. A candidate k-itemset is inserted into the tree by hashing on
each successive item in the candidate and then following the appropriate
branch of the tree according to the hash value. Once a leaf node is
reached, the candidate is inserted based on one of the following

Condition 1: If the depth of the leaf node is equal to k (the root is
assumed to be at depth 0), then the candidate is inserted regardless of
the number of itemsets already stored at the node.

Condition 2: If the depth of the leaf node is less than k, then the
candidate can be inserted as long as the number of itemsets stored at the
node is less than maxsize. Assume for this question.

Condition 3: If the depth of the leaf node is less than k and the number
of itemsets stored at the node is equal to maxsize, then the leaf node is
converted into an internal node. New leaf nodes are created as children


of the old leaf node. Candidate itemsets previously stored in the old leaf
node are distributed to the children based on their hash values. The new
candidate is also hashed to its appropriate leaf node.

b. How many leaf nodes are there in the candidate hash tree? How many
internal nodes are there?

c. Consider a transaction that contains the following items: {1, 2, 3, 5, 6}.
Using the hash tree constructed in part (a), which leaf nodes will be
checked against the transaction? What are the candidate 3-itemsets
contained in the transaction?

12. Given the lattice structure shown in Figure 5.33 and the transactions
given in Table 5.22 , label each node with the following letter(s):

Figure 5.33.
An itemset lattice

M if the node is a maximal frequent itemset,

C if it is a closed frequent itemset,

N if it is frequent but neither maximal nor closed, and

I if it is infrequent

Assume that the support threshold is equal to 30%.

13. The original association rule mining formulation uses the support and
confidence measures to prune uninteresting rules.

a. Draw a contingency table for each of the following rules using the
transactions shown in Table 5.23 .

Table 5.23. Example of market basket transactions.

Transaction ID Items Bought

1 {a, b, d, e}

2 {b, c, d}

3 {a, b, d, e}

4 {a, c, d, e}

5 {b, c, d, e}

6 {b, d, e}

7 {c, d}

8 {a, b, c}

9 {a, d, e}

10 {b, d}

Rules: .

b. Use the contingency tables in part (a) to compute and rank the rules in
decreasing order according to the following measures.

i. Support.

ii. Confidence.

iii. Interest


v. , where


14. Given the rankings you had obtained in Exercise 13, compute the
correlation between the rankings of confidence and the other five measures.
Which measure is most highly correlated with confidence? Which measure is
least correlated with confidence?

15. Answer the following questions using the data sets shown in Figure
5.34 . Note that each data set contains 1000 items and 10,000 transactions.
Dark cells indicate the presence of items and white cells indicate the absence
of items. We will apply the Apriori algorithm to extract frequent itemsets with

(i.e., itemsets must be contained in at least 1000 transactions).

{b}→{c}, {a}→{d}, {b}→{d}, {e}→{c}, {c}→{a}

(X→Y)=P(X, Y)P(X)P(Y).

IS(X→Y)=P(X, Y)P(X)P(Y).

Klosgen(X→Y)=P(X, Y )×max(P(Y|X))−P(Y), P(X|Y)−P(X))
P(Y|X)=P(X, Y)P(X)

Odds ratio(X→Y)=P(X, Y)P(X¯, Y¯)P(X, Y¯)P(X¯, Y).


Figure 5.34.
Figures for Exercise 15.

a. Which data set(s) will produce the most number of frequent itemsets?

b. Which data set(s) will produce the fewest number of frequent itemsets?

c. Which data set(s) will produce the longest frequent itemset?

d. Which data set(s) will produce frequent itemsets with highest maximum

e. Which data set(s) will produce frequent itemsets containing items with
wide-varying support levels (i.e., items with mixed support, ranging from
less than 20% to more than 70%)?


a. Prove that the coefficient is equal to 1 if and only if .

b. Show that if A and B are independent, then

c. Show that Yule’s Q and Y coefficients

are normalized versions of the odds ratio.

d. Write a simplified expression for the value of each measure shown in
Table 5.9 when the variables are statistically independent.

17. Consider the interestingness measure, , for an
association rule .

ϕ f11=f1+=f+1

P(A, B)×P(A¯, B¯)=P(A, B¯)×P(A¯, B)



a. What is the range of this measure? When does the measure attain its
maximum and minimum values?

b. How does M behave when P (A, B) is increased while P (A) and P (B)
remain unchanged?

c. How does M behave when P (A) is increased while P (A, B) and P (B)
remain unchanged?

d. How does M behave when P (B) is increased while P (A, B) and P (A)
remain unchanged?

e. Is the measure symmetric under variable permutation?

f. What is the value of the measure when A and B are statistically

g. Is the measure null-invariant?

h. Does the measure remain invariant under row or column scaling

i. How does the measure behave under the inversion operation?

18. Suppose we have market basket data consisting of 100 transactions and
20 items. Assume the support for item a is 25%, the support for item b is 90%
and the support for itemset {a, b} is 20%. Let the support and confidence
thresholds be 10% and 60%, respectively.

a. Compute the confidence of the association rule . Is the rule
interesting according to the confidence measure?

b. Compute the interest measure for the association pattern {a, b}. Describe
the nature of the relationship between item a and item b in terms of the
interest measure.


c. What conclusions can you draw from the results of parts (a) and (b)?

d. Prove that if the confidence of the rule is less than the support of
{b}, then:



where denote the rule confidence and denote the support of
an itemset.

19. Table 5.24 shows a contingency table for the binary variables A
and B at different values of the control variable C.

Table 5.24. A Contingency Table.


1 0

B 1 0 15

0 15 30

B 1 5 0

0 0 15

a. Compute the coefficient for A and B when and or 1.
Note that .

b. What conclusions can you draw from the above result?

20. Consider the contingency tables shown in Table 5.25 .




c(⋅) s(⋅)




ϕ C=0, C=1, C=0
ϕ=({A, B})=P(A, B)−P(A)P(B)P(A)P(B)(1−P(A))(1−P(B))

a. For table I, compute support, the interest measure, and the correlation
coefficient for the association pattern {A, B}. Also, compute the
confidence of rules and .

b. For table II, compute support, the interest measure, and the correlation
coefficient for the association pattern {A, B}. Also, compute the
confidence of rules and .