Modeling the Internet and the Web

This Page Intentionally Left Blank

Modeling the Internet and the Web
Probabilistic Methods and Algorithms

Pierre Baldi
School of Information and Computer Science,
University of California, Irvine, USA
Paolo Frasconi
Department of Systems and Computer Science,
University of Florence, Italy
Padhraic Smyth
School of Information and Computer Science,
University of California, Irvine, USA
Copyright 2003 Pierre Baldi, Paolo Frasconi and Padhraic Smyth
Published by John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester,
West Sussex PO19 8SQ, England
Phone (+44) 1243 779777
Email (for orders and customer service enquiries): [email protected]
Visit our Home Page on www.wileyeurope.com or www.wiley.com

All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system or
transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning or
otherwise, except under the terms of the Copyright, Designs and Patents Act 1988 or under the terms of
a licence issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1T 4LP,
UK, without the permission in writing of the Publisher. Requests to the Publisher should be addressed
to the Permissions Department, John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West
Sussex PO19 8SQ, England, or emailed to [email protected], or faxed to (+44) 1243 770620.

This publication is designed to provide accurate and authoritative information in regard to the subject matter
covered. It is sold on the understanding that the Publisher is not engaged in rendering professional services.
If professional advice or other expert assistance is required, the services of a competent professional should
be sought.

Other Wiley Editorial Ofces

John Wiley & Sons Inc., 111 River Street, Hoboken, NJ 07030, USA
Jossey-Bass, 989 Market Street, San Francisco, CA 94103-1741, USA
Wiley-VCH Verlag GmbH, Boschstr. 12, D-69469 Weinheim, Germany
John Wiley & Sons Australia Ltd, 33 Park Road, Milton, Queensland 4064, Australia
John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809
John Wiley & Sons Canada Ltd, 22 Worcester Road, Etobicoke, Ontario, Canada M9W 1L1

Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may
not be available in electronic books.

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

ISBN 0-470-84906-1

Typeset in 10/12pt Times by T&T Productions Ltd, London.

Printed and bound in Great Britain by Biddles Ltd, Guildford, Surrey.
This book is printed on acid-free paper responsibly manufactured from sustainable forestry
in which at least two trees are planted for each one used for paper production.
To Ezio and Jos (P.B.), to Neda (P.F.) and
to Seosamh and Brd ine (P.S.)
This Page Intentionally Left Blank
Contents

Preface xiii

1 Mathematical Background 1
1.1 Probability and Learning from a Bayesian Perspective 1
1.2 Parameter Estimation from Data 4
1.2.1 Basic principles 4
1.2.2 A simple die example 6
1.3 Mixture Models and the Expectation Maximization Algorithm 10
1.4 Graphical Models 13
1.4.1 Bayesian networks 13
1.4.2 Belief propagation 15
1.4.3 Learning directed graphical models from data 16
1.5 Classication 17
1.6 Clustering 20
1.7 Power-Law Distributions 22
1.7.1 Denition 22
1.7.2 Scale-free properties (80/20 rule) 24
1.7.3 Applications to Languages: Zipfs and Heaps Laws 24
1.7.4 Origin of power-law distributions and Fermis model 26
1.8 Exercises 27

2 Basic WWW Technologies 29

2.1 Web Documents 30
2.1.1 SGML and HTML 30
2.1.2 General structure of an HTML document 31
2.1.3 Links 32
2.2 Resource Identiers: URI, URL, and URN 33
2.3 Protocols 36
2.3.1 Reference models and TCP/IP 36
2.3.2 The domain name system 37
2.3.3 The Hypertext Transfer Protocol 38
2.3.4 Programming examples 40
viii CONTENTS

2.4 Log Files 41

2.5 Search Engines 44
2.5.1 Overview 44
2.5.2 Coverage 45
2.5.3 Basic crawling 46
2.6 Exercises 49

3 Web Graphs 51
3.1 Internet and Web Graphs 51
3.1.1 Power-law size 53
3.1.2 Power-law connectivity 53
3.1.3 Small-world networks 56
3.1.4 Power law of PageRank 57
3.1.5 The bow-tie structure 58
3.2 Generative Models for the Web Graph and Other Networks 60
3.2.1 Web page growth 60
3.2.2 Lattice perturbation models: between order and disorder 61
3.2.3 Preferential attachment models, or the rich get richer 63
3.2.4 Copy models 66
3.2.5 PageRank models 67
3.3 Applications 68
3.3.1 Distributed search algorithms 68
3.3.2 Subgraph patterns and communities 70
3.3.3 Robustness and vulnerability 72
3.4 Notes and Additional Technical References 73
3.5 Exercises 74

4 Text Analysis 77
4.1 Indexing 77
4.1.1 Basic concepts 77
4.1.2 Compression techniques 79
4.2 Lexical Processing 80
4.2.1 Tokenization 80
4.2.2 Text conation and vocabulary reduction 82
4.3 Content-Based Ranking 82
4.3.1 The vector-space model 82
4.3.2 Document similarity 83
4.3.3 Retrieval and evaluation measures 85
4.4 Probabilistic Retrieval 86
4.5 Latent Semantic Analysis 88
4.5.1 LSI and text documents 89
4.5.2 Probabilistic LSA 89
4.6 Text Categorization 93
CONTENTS ix

4.6.1 k nearest neighbors 93

4.6.2 The Naive Bayes classier 94
4.6.3 Support vector classiers 97
4.6.4 Feature selection 102
4.6.5 Measures of performance 104
4.6.6 Applications 106
4.6.7 Supervised learning with unlabeled data 111
4.7 Exploiting Hyperlinks 114
4.7.1 Co-training 114
4.7.2 Relational learning 115
4.8 Document Clustering 116
4.8.1 Background and examples 116
4.8.2 Clustering algorithms for documents 117
4.8.3 Related approaches 119
4.9 Information Extraction 120
4.10 Exercises 122

5 Link Analysis 125

5.1 Early Approaches to Link Analysis 126
5.2 Nonnegative Matrices and Dominant Eigenvectors 128
5.3 Hubs and Authorities: HITS 131
5.4 PageRank 134
5.5 Stability 138
5.5.1 Stability of HITS 139
5.5.2 Stability of PageRank 139
5.6 Probabilistic Link Analysis 140
5.6.1 SALSA 140
5.6.2 PHITS 142
5.7 Limitations of Link Analysis 143

6 Advanced Crawling Techniques 149

6.1 Selective Crawling 149
6.2 Focused Crawling 152
6.2.1 Focused crawling by relevance prediction 152
6.2.2 Context graphs 154
6.2.3 Reinforcement learning 155
6.2.4 Related intelligent Web agents 157
6.3 Distributed Crawling 158
6.4 Web Dynamics 160
6.4.1 Lifetime and aging of documents 161
6.4.2 Other measures of recency 167
6.4.3 Recency and synchronization policies 167
x CONTENTS

7 Modeling and Understanding Human Behavior on the

Web 171
7.1 Introduction 171
7.2 Web Data and Measurement Issues 172
7.2.1 Background 172
7.2.2 Server-side data 174
7.2.3 Client-side data 177
7.3 Empirical Client-Side Studies of Browsing Behavior 179
7.3.1 Early studies from 1995 to 1997 180
7.3.2 The Cockburn and McKenzie study from 2002 181
7.4 Probabilistic Models of Browsing Behavior 184
7.4.1 Markov models for page prediction 184
7.4.2 Fitting Markov models to observed page-request data 186
7.4.3 Bayesian parameter estimation for Markov models 187
7.4.4 Predicting page requests with Markov models 189
7.4.5 Modeling runlengths within states 193
7.4.6 Modeling session lengths 194
7.4.7 A decision-theoretic surng model 198
7.4.8 Predicting page requests using additional variables 199
7.5 Modeling and Understanding Search Engine Querying 201
7.5.1 Empirical studies of search behavior 202
7.5.2 Models for search strategies 207
7.6 Exercises 208

8 Commerce on the Web: Models and Applications 211

8.1 Introduction 211
8.2 Customer Data on the Web 212
8.3 Automated Recommender Systems 212
8.3.1 Evaluating recommender systems 214
8.3.2 Nearest-neighbor collaborative ltering 215
8.3.3 Model-based collaborative ltering 218
8.3.4 Model-based combining of votes and content 223
8.4 Networks and Recommendations 224
8.4.1 Email-based product recommendations 224
8.4.2 A diffusion model 226
8.5 Web Path Analysis for Purchase Prediction 228
8.6 Exercises 232

Appendix A Mathematical Complements 235

A.1 Graph Theory 235
A.1.1 Basic denitions 235
A.1.2 Connectivity 236
A.1.3 Random graphs 236
CONTENTS xi

A.2 Distributions 237

A.2.1 Expectation, variance, and covariance 237
A.2.2 Discrete distributions 237
A.2.3 Continuous distributions 238
A.2.4 Weibull distribution 240
A.2.5 Exponential family 240
A.2.6 Extreme value distribution 241
A.3 Singular Value Decomposition 241
A.4 Markov Chains 243
A.5 Information Theory 243
A.5.1 Mathematical background 244
A.5.2 Information, surprise, and relevance 247

Appendix B List of Main Symbols and Abbreviations 253

References 257

Index 277
This Page Intentionally Left Blank
Preface

Since its early ARPANET inception during the Cold War, the Internet has grown
by a staggering nine orders of magnitude. Today, the Internet and the World Wide
Web pervade our lives, having fundamentally altered the way we seek, exchange,
distribute, and process information. The Internet has become a powerful social force,
transforming communication, entertainment, commerce, politics, medicine, science,
and more. It mediates an ever growing fraction of human knowledge, forming both
the largest library and the largest marketplace on planet Earth.
Unlike the invention of earlier media such as the press, photography, or even the
radio, which created specialized passive media, the Internet and the Web impact all
information, converting it to a uniform digital format of bits and packets. In addition,
the Internet and the Web form a dynamic medium, allowing software applications
to control, search, modify, and lter information without human intervention. For
example, email messages can carry programs that affect the behavior of the receiving
computer. This active medium also promotes human intervention in sharing, updating,
linking, embellishing, critiquing, corrupting, etc., information to a degree that far
exceeds what could be achieved with printed documents.
In common usage, the words Internet and Web (or World Wide Web or WWW)
are often used interchangeably. Although they are intimately related, there are of
course some nuances which we have tried to respect. Internet, in particular, is the
more general term and implicitly includes physical aspects of the underlying networks
as well as mechanisms such as email and peer-to-peer activities that are not directly
associated with the Web. The term Web, on the other hand, is associated with the
information stored and available on the Internet. It is also a term that points to other
complex networks of information, such as webs of scientic citations, social relations,
or even protein interactions. In this sense, it is fair to say that a predominant fraction
of our book is about the Web and the information aspects of the Internet. We use Web
every time we refer to the World Wide Web and web when we refer to a broader
class of networks or other kinds of networks, i.e. web of citations.
As the Internet and the Web continue to expand at an exponential rate, it also
evolves in terms of the devices and processors connected to it, e.g. wireless devices
and appliances. Ever more human domains and activities are ensnared by the Web,
thus creating challenging problems of ownership, security, and privacy. For instance,
xiv PREFACE

we are quite far from having solved the security, privacy, and authentication problems
that would allow us to hold national Internet elections.
As scientists, the Web has also become a tool we use on a daily basis for tasks
ranging from the mundane to the intractable, to search and disseminate information,
to exchange views and collaborate, to post job listings, to retrieve and quote (by
Uniform Resource Locator (URL)) bibliographic information, to build Web servers,
and even to compute. There is hardly a branch of computer science that is not affected
by the Internet: not only the most obvious areas such as networking and protocols,
but also security and cryptography; scientic computing; human interfaces, graphics,
and visualization; information retrieval, data mining, machine learning, language/text
modeling and articial intelligence, to name just a few.
What is perhaps less obvious and central to this book is that not only have the
Web and the Internet become essential tools of scientic enterprise, but they have
also themselves become the objects of active scientic investigation. And not only for
computer scientists and engineers, but also for mathematicians, economists, social
scientists, and even biologists.
There are many reasons why the Internet and the Web are exciting, albeit young,
topics for scientic investigation. These reasons go beyond the need to improve the
underlying technology and to harness the Web for commercial applications. Because
the Internet and the Web can be viewed as dynamic constellations of interconnected
processors and Web pages, respectively, they can be monitored in many ways and
at many different levels of granularity, ranging from packet trafc, to user behavior,
to the graphical structure of Web pages and their hyperlinks. These measurements
provide new types of large-scale data sets that can be scientically analyzed and
mined at different levels. Thus researchers enjoy unprecedented opportunities to,
for instance:
gather, communicate, and exchange ideas, documents, and information;
monitor a large dynamic network with billions of nodes and one order of mag-
nitude more connections;
gather large training sets of textual or activity data, for the purposes of modeling
and predicting the behavior of millions of users;
analyze and understand interests and relationships within society.
The Web, for instance, can be viewed as an example of a very large distributed and
dynamic system with billions of pages resulting from the uncoordinated actions of
millions of individuals. After all, anyone can post a Web page on the Internet and link
it to any other page. In spite of this complete lack of central control, the graphical
structure of the Web is far from random and possesses emergent properties shared
with other complex graphs found in social, technological, and biological systems.
Examples of properties include the power-law distribution of vertex connectivities
and the small-world property any two Web pages are usually only a few clicks
away from each other. Similarly, predictable patterns of congestion (e.g. trafc jams)
PREFACE xv

have also been observed in Internet trafc. While the exploitation of these regularities
may be benecial to providers and consumers, their mere existence and discovery has
become a topic of basic research.

Why Probabilistic Modeling?

By its very nature, a very large distributed, decentralized, self-organized, and evolving
system necessarily yields uncertain and incomplete measurements and data. Probabil-
ity and statistics are the fundamental mathematical tools that allow us to model, rea-
son and proceed with inference in uncertain environments. Not only are probabilistic
methods needed to deal with noisy measurements, but many of the underlying phe-
nomena, including the dynamic evolution of the Internet and the Web, are themselves
probabilistic in nature. As in the systems studied in statistical mechanics, regularities
may emerge from the more or less random interactions of myriads of small factors.
Aggregation can only be captured probabilistically. Furthermore, and not unlike bio-
logical systems, the Internet is a very high-dimensional system, where measurement
of all relevant variables becomes impossible. Most variables remain hidden and must
be factored out by probabilistic methods.
There is one more important reason why probabilistic modeling is central to this
book. At a fundamental level the Web is concerned with information retrieval and the
semantics, or meaning, of that information. While the modeling of semantics remains
largely an open research problem, probabilistic methods have achieved remarkable
successes and are widely used in information retrieval, machine translation, and more.
Although these probabilistic methods bypass or fake semantic understanding, they are,
for instance, at the core of the search engines we use every day. As it happens, the
Internet and the Web themselves have greatly aided the development of such methods
by making available large corpora of data from which statistical regularities can be
extracted.
Thus, probabilistic methods pervasively apply to diverse areas of Internet and
Web modeling and analysis, such as network trafc, graphical structure, informa-
tion retrieval engines, and customer behavior.

Audience and Prerequisites

Our aim has been to write an interdisciplinary textbook about the Internet both to ll
a specic niche at the undergraduate and graduate level and to serve as a reference
for researchers and practitioners from academia, industry, and government. Thus, it
is aimed at a relatively broad audience including both students and more advanced
researchers, with diverse backgrounds. We have tried to provide a succinct and self-
contained description of the main concepts and methods in a manner accessible to
computer scientists and engineers and also to those whose primary background is in
other disciplines touched by the Internet and the Web. We hope that the book will be
xvi PREFACE

of interest to students, postdoctoral fellows, faculty members and researchers from a

variety of disciplines including Computer Science, Engineering, Statistics, Applied
Mathematics, Economics and Business, and Social Sciences.
The topic is quite broad. On the surface the Web could appear to be a limited sub-
discipline of computer science, but in reality it is impossible for a single researcher
to have an in-depth knowledge and understanding of all the areas of science and
technology touched by the Internet and the Web. While we do not claim to cover all
aspects of the Internet for instance, we do not look in any detail at the physical
layer we do try to cover the most important aspects of the Web at the information
level and provide pointers to the reader for topics that are left out. We propose a
unied treatment based on mathematical and probabilistic modeling that emphasizes
the unity of the eld, as well as its connections to other areas such as machine learning,
data mining, graph theory, information retrieval, and bioinformatics.
The prerequisites include an understanding of several basic mathematical con-
cepts at an undergraduate level, including probabilistic concepts and methods, basic
calculus, and matrix algebra, as well as elementary concepts in data structures and
algorithms. Some additional knowledge of graph theory and combinatorics is helpful
but not required. Mathematical proofs are usually short and mathematical details that
can be found in the cited literature are sometimes left out in favor of a more intuitive
treatment. We expect the typical reader to be able to gather complementary informa-
tion from the references, as needed. For instance we refer to, but do not provide the
details of, the algorithm for nding the shortest path between two vertices in a graph,
since this is readily found in other textbooks.
We have included many concrete examples, such as examples of pseudocode and
analyses of specic data sets, as well as exercises of varying difculty at the end of each
chapter. Some are meant to encourage basic thinking about the Internet and the Web,
and the corresponding probabilistic models. Other exercises are more suited for class
projects and require computer measurements and simulations, such as constructing
the graph of pages and hyperlinks associated with ones own institution. These are
complemented by more mathematical exercises of varying levels of difculty.
While the book can be used in a course about the Web, or as complementary reading
material in, for instance, an information retrieval class, we are also in the process of
using it to teach a course on the application of probability, statistics, and information
theory in computer science by providing a unied theme, set of methods, and a variety
of close-to-home examples and problems aimed at developing both mathematical
intuition and computer simulation skills.

Content and General Outline of the Book

We have strived to write a comprehensive but reasonably concise introductory book
that is self-contained and summarizes a wide range of results that are scattered
throughout the literature. A portion of the book is built on material taken from articles
we have written over the years, as well as talks, courses, and tutorials. Our main focus
PREFACE xvii

is not on the history of a rapidly evolving eld, but rather on what we believe are
the primary relevant methods and algorithms, and a general way of thinking about
modeling of the Web that we hope will prove useful.
Chapter 1 covers in succinct form most of the mathematical background needed
for the following chapters and can be skipped by those with a good familiarity with its
material. It contains an introduction to basic concepts in probabilistic modeling and
machine learning from the Bayesian framework and the theory of graphical models
to mixtures, classication, and clustering these are all used throughout various
chapters and form a substantial part of the glue of this book.
Chapter 2 provides an introduction to the Internet and the Web and the foundations
of the WWW technologies that are necessary to understand the rest of the book,
including the structure of Web documents, the basics of Internet protocols, Web server
log les, and so forth. Server log les, for instance, are important to thoroughly
understand the analysis of human behavior on the Web in Chapter 7. The chapter also
deals with the basic principles of Web crawlers. Web crawling is essential to gather
information about the Web and in this sense is a prerequisite for the study of the Web
graph in Chapter 3.
Chapter 3 studies the Internet and the Web as large graphs. It describes, models,
and analyzes the power-law distribution of Web sizes, connectivity, PageRank, and
the small-world properties of the underlying graphs. Applications of graphical prop-
erties, for instance to improve search engines, are also covered in this chapter and
further studied in later chapters.
Chapter 4 deals with text analysis in terms of indexing, content-based ranking,
latent semantic analysis, and text categorization, providing the basic components
(together with link analysis) for understanding how to efciently retrieve information
over the Web.
Chapter 5 builds upon the graphical results of Chapter 4 and deals with link analysis,
inferring page relevance from patterns of connectivity, Web communities, and the
stability and evolution of these concepts with time.
Chapter 6 covers advanced crawling techniques selective, focused, and distributed
crawling and Web dynamics. It is essential material in order to understand how to
build a new search engine for instance.
Chapter 7 studies human behavior on the Web. In particular it builds and stud-
ies several probabilistic models of human browsing behavior and also analyzes the
statistical properties of search engine queries.
Finally, Chapter 8 covers various aspects of commerce on the Web, including
analysis of customer Web data, automated recommendation systems, and Web path
analysis for purchase prediction.
Appendix A contains a number of technical sections that are important for reference
and for a thorough understanding of the material, including an informal introduction
to basic concepts in graph theory, a list of standard probability densities, a short
section on Singular Value Decomposition, a short section on Markov chains, and a
brief, critical, overview of information theory.
xviii PREFACE

What Is New and What Is Omitted

On several occasions we present new material, or old material but from a somewhat
new perspective. Examples include the notion of surprise in Appendix A, as well as a
simple model for power-law distributions originally due to Enrico Fermi that seems to
have been forgotten, which is described in Chapter 2. The material in this book and its
treatment reect our personal biases. Many relevant topics had to be omitted in order to
stay within reasonable size limits. In particular, the book contains little material about
the physical layer, about any hardware, or about Internet protocols. Other important
topics that we would have liked to cover but had to be left out include the aspects
of the Web related to security and cryptography, human interfaces and design. We
do cover many aspects of text analysis and information retrieval, but not all of them,
since a more exhaustive treatment of any of these topics would require a book by
itself. Thus, in short, the main focus of the book is on the information aspects of the
Web, its emerging properties, and some of its applications.

Notation
In terms of notation, most of the symbols used are listed at the end of the book, in
Appendix B. A symbol such as D represents the data, regardless of the amount or
complexity. Boldface letters are usually reserved for matrices and vectors. Capital
letters are typically used for matrices and random variables, lowercase letters for
scalars and random variable realizations. Greek letters such as typically denote
the parameters of a model. Throughout the book P and E are used for probability
and expectation. If X is a random variable, we often write P (x) for P (X = x),
or sometimes just P (X) if no confusion is possible. E[X], var[X], and cov[X, Y ],
respectively, denote the expectation, variance, and covariance associated with the
random variables X and Y with respect to the probability distributions P (X) and
P (X, Y ).
We use the standard notation f (n) = o(g(n)) to denote a function f (n) that satises
f (n)/g(n) 0 as n , and f (n) = O(g(n)) when there exists a constant C > 0
such that f (n)
Cg(n) when n . Similarly, we use f (n) = (g(n)) to denote
a function f (n) such that asymptotically there are two constants C1 and C2 with
C1 g(n)
f (n)
C2 g(n). Calligraphic style is reserved for particular functions,
such as error or energy (E ), entropy and relative entropy (H ). Finally, we often deal
with quantities characterized by many indices. Within a given context, only the most
relevant indices are indicated.

Acknowledgements
Over the years, this book has been supported directly or indirectly by grants and
awards from the US National Science Foundation, the National Institutes of Health,
PREFACE xix

NASA and the Jet Propulsion Laboratory, the Department of Energy and Lawrence
Livermore National Laboratory, IBM Research, Microsoft Research, Sun Microsys-
tems, HNC Software, the University of California MICRO Program, and a Laurel
Wilkening Faculty Innovation Award. Part of the book was written while P.F. was
visiting the School of Information and Computer Science (ICS) at UCI, with partial
funding provided by the University of California. We also would like to acknowledge
the general support we have received from the Institute for Genomics and Bioinfor-
matics (IGB) at UCI and the California Institute for Telecommunications and Infor-
mation Technology (Cal(IT)2 ). Within IGB, special thanks go the staff, in particular
Suzanne Knight, Janet Ko, Michele McCrea, and Ann Marie Walker. We would like
to acknowledge feedback and general support from many members of Cal(IT)2 and
thank in particular its directors Bill Parker, Larry Smarr, Peter Rentzepis, and Ramesh
Rao, and staff members Catherine Hammond, Ashley Larsen, Doug Ramsey, Stuart
Ross, and Stephanie Sides. We thank a number of colleagues for discussions and
feedback on various aspects of the Web and probabilistic modeling: David Eppstein,
Chen Li, Sharad Mehrotra, and Mike Pazzani at UC Irvine, as well as Albert-Lszl
Barabsi, Nicol Cesa-Bianchi, Monica Bianchini, C. Lee Giles, Marco Gori, David
Heckerman, David Madigan, Marco Maggini, Heikki Mannila, Chris Meek, Amnon
Meyers, Ion Muslea, Franco Scarselli, Giovanni Soda, and Steven Scott. We thank all
of the people who have helped with simulations or provided feedback on the various
versions of this manuscript, especially our students Gianluca Pollastri, Alessandro
Vullo, Igor Cadez, Jianlin Chen, Xianping Ge, Joshua OMadadhain, Scott White,
Alessio Ceroni, Fabrizio Costa, Michelangelo Diligenti, Sauro Menchetti, and Andrea
Passerini. We also acknowledge Jean-Pierre Nadal, who brought to our attention the
Fermi model of power laws. We thank Xinglian Yie and David OHallaron for pro-
viding their data on search engine queries in Chapter 8. We also thank the staff from
John Wiley & Sons, Ltd, in particular Senior Editor Sian Jones and Robert Calver,
and Emma Dain at T&T Productions Ltd. Finally, we acknowledge our families and
friends for their support in the writing of this book.

Pierre Baldi, Paolo Frasconi and Padhraic Smyth

October 2002, Irvine, CA
This Page Intentionally Left Blank
1

Mathematical Background

In this chapter we review a number of basic concepts in probabilistic modeling and

data analysis that are used throughout the book, including parameter estimation, mix-
ture models, graphical models, classication, clustering, and power-law distributions.
Each of these topics is worthy of an entire chapter (or even a whole book) by itself, so
our treatment is necessarily brief. Nonetheless, the goal of this chapter is to provide
the introductory foundations for models and techniques that will be widely used in
the remainder of the book. Readers who are already familiar with these concepts, or
who want to avoid mathematical details during a rst reading, can safely skip to the
next chapter. More specic technical mathematical concepts or mathematical com-
plements that concern only a specic chapter rather than the entire book are given in
Appendix A.

1.1 Probability and Learning from a Bayesian

Perspective
Throughout this book we will make frequent use of probabilistic models to charac-
terize various phenomena related to the Web. Both theory and experience have shown
that probability is by far the most useful framework currently available for modeling
uncertainty. Probability allows us to reason in a coherent manner about events and
make inferences about such events given observed data. More specically, an event e
is a proposition or statement about the world at large. For example, let e be the propo-
sition that the number of Web pages in existence on 1 January 2003 was greater than
ve billion. A well-dened proposition e is either true or false by some reasonable
denition of what constitutes a Web page, the total number that existed in January
2003 was either greater than ve billion or not. There is, however, considerable uncer-
tainty about what this number was back in January 2003 since, as we will discuss later
in Chapters 2 and 3, accurately estimating the size of the Web is a quite challenging
problem. Consequently there is uncertainty about whether the proposition e is true or
not.

Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
2 PROBABILITY AND LEARNING FROM A BAYESIAN PERSPECTIVE

A probability, P (e), can be viewed as a number that reects our uncertainty about
whether e is true or false in the real world, given whatever information we have avail-
able. This is known as the degree of belief or Bayesian interpretation of probability
(see, for instance, Berger 1985; Box and Tiao 1992; Cox 1964; Gelman et al. 1995;
Jaynes 2003) and is the one that we will use by default throughout this text. In fact,
to be more precise, we should use a conditional probability P (e | I) in general to
represent degree of belief, where I is the background information on which our belief
is based. For simplicity of notation we will often omit this conditioning on I, but it
may be useful to keep in mind that everywhere you see a P (e) for some proposition
e, there is usually some implicit background information I that is known or assumed
to be true.
The Bayesian interpretation of a probability P (e) is a generalization of the more
classic interpretation of a probability as the relative frequency of successes to total
trials, estimated over an innite number of hypothetical repeated trials (the so-called
frequentist interpretation). The Bayesian interpretation is more useful in general,
since it allows us to make statements about propositions such as the number of Web
pages in existence where a repeated trials interpretation would not necessarily apply.
It can be shown that, under a small set of reasonable axioms, degrees of belief can
be represented by real numbers and that when rescaled to the [0, 1] interval these
degrees of condence must obey the rules of probability and, in particular, Bayes
theorem (Cox 1964; Jaynes 1986, 2003; Savage 1972). This is reassuring, since it
means that the standard rules of probability still apply whether we are using the degree
of belief interpretation or the frequentist interpretation. In other words, the rules for
manipulating probabilities such as conditioning or the law of total probability remain
the same no matter what semantics we attach to the probabilities.
The Bayesian approach also allows us to think about probability as being a dynamic
entity that is updated as more data arrive as we receive more data we may naturally
change our degree of belief in certain propositions given these new data. Thus, for
example, we will frequently refer to terms such as P (e | D) where D is some data.
In fact, by Bayes theorem,
P (D | e)P (e)
P (e | D) = . (1.1)
P (D)
The interpretation of each of the terms in this equation is worth discussing. P (e) is your
belief in the event e before you see any data at all, referred to as your prior probability
for e or prior degree of belief in e. For example, letting e again be the statement that
the number of Web pages in existence on 1 January 2003 was greater than ve billion,
P (e) reects your degree of belief that this statement is true. Suppose you now receive
some data D which is the number of pages indexed by various search engines as of
1 January 2003. To a reasonable approximation we can view these numbers as lower
bounds on the true number and lets say for the sake of argument that all the numbers
are considerably less than ve billion. P (e | D) now reects your updated posterior
belief in e given the observed data and it can be calculated by using Bayestheorem via
MATHEMATICAL BACKGROUND 3

Equation (1.1). The right-hand side of Equation (1.1) includes the prior, so naturally
enough the posterior is proportional to the prior.
The right-hand side also includes P (D | e), which is known as the likelihood of the
data, i.e. the probability of the data under the assumption that e is true. To calculate the
likelihood we must have a probabilistic model that connects the proposition e we are
interested in with the observed data D this is the essence of probabilistic learning.
For our Web page example, this could be a model that puts a probability distribution
on the number of Web pages that each search engine may nd if the conditioning event
is true, i.e. if there are in fact more than ve billion Web pages in existence. This could
be a complex model of how the search engines actually work, taking into account all
the various reasons that many pages will not be found, or it might be a very simple
approximate model that says that each search engine has some conditional distribution
on the number of pages that will be indexed, as a function of the total number that
exist. Appendix A provides examples of several standard probability models these
are in essence the building blocks for probabilistic modeling and can be used as
components either in likelihood models P (D | e) or as priors P (e).
Continuing with Equation (1.1), the likelihood expression reects how likely the
observed data are, given e and given some model connecting e and the data. If P (D | e)
is very low, this means that the model is assigning a low probability to the observed
data. This might happen, for example, if the search engines hypothetically all reported
numbers of indexed pages in the range of a few million rather than in the billion range.
Of course we have to factor in the alternative hypothesis, e, here and we must ensure
that both P (e) + P (e) = 1 and P (e | D) + P (e | D) = 1 to satisfy the basic axioms
of probability. The normalization constant in the denominator of Equation (1.1) can
be calculated by noting that P (D) = P (D | e)P (e) + P (D | e)P (e). It is easy to see
that P (e | D) depends both on the prior and the likelihood in terms of competing
with the alternative hypothesis e the larger they are relative to the prior for e and
the likelihood for e, then the larger our posterior belief in e will be.
Because probabilities can be very small quantities and addition is often easier to
work with than multiplication, it is common to take logarithms of both sides, so that

log P (e | D) = log P (D | e) + log P (e) log P (D). (1.2)

To apply Equation (1.1) or (1.2) to any class of models, we only need to specify the
prior P (e) and the data likelihood P (D | e).
Having updated our degree of belief in e, from P (e) to P (e | D), we can continue
this process and incorporate more data as they become available. For example, we
might later obtain more data on the size of the Web from a different study call this
second data set D2 . We can use Bayes rule to write
P (D2 | e, D)P (e | D)
P (e | D, D2 ) = . (1.3)
P (D2 | D)
Comparing Equations (1.3) and (1.2) we see that the old posterior P (e | D) plays the
role of the new prior when data set D2 arrives.
4 PARAMETER ESTIMATION FROM DATA

The use of priors is a strength of the Bayesian approach, since it allows the incor-
poration of prior knowledge and constraints into the modeling process. In general,
the effects of priors diminish as the number of data points increases. Formally, this is
because the log-likelihood log P (D | e) typically increases linearly with the number
of data points in D, while the prior log P (e) remains constant. Finally, and most impor-
tantly, the effects of different priors, as well as different models and model classes,
can be assessed within the Bayesian framework by comparing the corresponding
probabilities.
The computation of the likelihood is of course model dependent and is not addressed
here in its full generality. Later in the chapter we will briey look at a variety of
graphical model and mixture model techniques that act as components in a exible
toolbox for the construction of different types of likelihood functions.

1.2 Parameter Estimation from Data

1.2.1 Basic principles
We can now turn to the main type of inference that will be used throughout this book,
namely estimation of parameters under the assumption of a particular functional
form for a model M. For example, if our model M is the Gaussian distribution, then
we have two parameters: the mean and standard deviation of this Gaussian.
In what follows below we will refer to priors, likelihoods, and posteriors relating
to sets of parameters . Typically our parameters are a set of real-valued numbers.
Thus, both the prior P () and the posterior P ( | D) are dening probability density
functions over this set of real-valued numbers. For example, our prior density might
assert that a particular parameter (such as the standard deviation) must be positive, or
that the mean is likely to lie within a certain range on the real line. From a Bayesian
viewpoint our prior and posterior reect our degree of belief in terms of where lies,
given the relevant evidence. For example, for a single parameter , P ( > 0 | D) is
the posterior probability that is greater than zero, given the evidence provided by the
data. For simplicity of notation we will not always explicitly include the conditioning
term for the model M, P ( | M) or P ( | D, M), but instead write expressions such
as P ( ) which implicitly assume the presence of a model M with some particular
functional form.
The general objective of parameter estimation is to nd or approximate the best
set of parameters for a model that is, to nd the set of parameters maximizing the
posterior P ( | D), or log P ( | D). This is called maximum a posteriori (MAP)
estimation.
In order to deal with positive quantities, we can minimize log P ( | D):
E ( ) = log P ( | D) = log P (D | ) log P ( ) + log P (D). (1.4)
From an optimization standpoint, the logarithm of the prior P ( ) plays the role of a
regularizer, that is, of an additional penalty term that can be used to enforce additional
MATHEMATICAL BACKGROUND 5

constraints such as smoothness. Note that the term P (D) in (1.4) plays the role of
a normalizing constant that does not depend on the parameters , and is therefore
irrelevant for this optimization. If the prior P () is uniform over parameter space,
then the problem reduces to nding the maximum of P (D | ), or log P (D | ). This
is known as maximum-likelihood (ML) estimation.
A substantial portion of this book and of machine-learning practice in general is
based on MAP estimation, that is, the minimization of
E () = log P (D | ) log P (), (1.5)
or the simpler ML estimation procedure, that is, the minimization of
E () = log P (D | ). (1.6)
In many useful and interesting models the function being optimized is complex and
its modes cannot be found analytically. Thus, one must resort to iterative and pos-
sibly stochastic methods such as gradient descent, expectation maximization (EM),
or simulated annealing. In addition, one may also have to settle for approximate or
sub-optimal solutions, since nding global optima of these functions may be computa-
tionally infeasible. Finally it is worth mentioning that mean posterior (MP) estimation
is also used, and can be more robust than the mode (MAP) in certain respects. The
MP is found by estimating by its expectation E[ ] with respect to the posterior
P ( | D), rather than the mode of this distribution.
Whereas nding the optimal model, i.e. the optimal set of parameters, is common
practice, it is essential to note that this is really useful only if the distribution P ( | D)
is sharply peaked around a unique optimum. In situations characterized by a high
degree of uncertainty and relatively small amounts of available data, this is often not
the case. Thus, a full Bayesian approach focuses on the function P ( | D) over the
entire parameter space rather than at a single point. Discussion of the full Bayesian
approach is somewhat beyond the scope of this book (see Gelman et al. (1995) for
a comprehensive introductory treatment). In most of the cases we will consider, the
simpler ML or MAP estimation approaches are sufcient to yield useful results this
is particularly true in the case of Web data sets which are often large enough that the
posterior distribution can be reasonably expected to be concentrated relatively tightly
about the posterior mode.
The reader ought also to be aware that whatever criterion is used to measure the
discrepancy between a model and data (often described in terms of an error or energy
function), such a criterion can always be dened in terms of an underlying probabilistic
model that is amenable to Bayesian analysis. Indeed, if the t of a model M = M()
with parameters is measured by some error function f (, D)  0 to be minimized,
one can always dene the associated likelihood as
ef ( ,D)
P (D | M()) = , (1.7)
Z


where Z = ef (,D) d is a normalizing factor (the partition function in statisti-
cal mechanics) that ensures the probabilities integrate to unity. As a result, minimizing
6 PARAMETER ESTIMATION FROM DATA

the error function is equivalent to ML estimation or, more generally, MAP estimation,
since Z is a constant and maximizing log P (D | M( )) is the same as minimizing
f ( , D) as a function of . For example, when the sum of squared differences is used
as an error function (a rather common practice), this implies an underlying Gaus-
sian model on the errors. Thus, the Bayesian point of view claries the probabilistic
assumptions that underlie any criteria for matching models with data.

1.2.2 A simple die example

Consider an m-ary alphabet of symbols, A = {a1 , . . . , am }. Assume that we have
a data set D that consists of a set of n observations from this alphabet, where D =
X1 , . . . , Xn . We can convert the raw data D into counts for each symbol {n1 , . . . , nm }.
For example, we could write a program to crawl the Web to nd Web pages containing
English text and parse the resulting set of pages to remove HTML and other non-
English words. In this case the alphabet A would consist of a set of English words
and the counts would be the number of times each word was observed.
The simplest probabilistic model for such a bag of words is a memoryless model
consisting of a die with m sides, one for each symbol in the alphabet, and where
we assume that the set D has been generated by n successive independent tosses of
this die. Because the tosses are independent and there is a unique underlying die, for
likelihood considerations it does not matter whether the data were generated from
many Web pages or from a single Web page. Nor does it even matter what the order
of the words on any page is, since we are not modeling the sequential order of the
words in our very simple model. In Chapter 4 we will see that this corresponds to the
widely used bag of words model for text analysis.
Our model M has m parameters  i , 1
i
m, corresponding to the probability
of producing each symbol, with i i = 1. Assuming independent repeated draws
from this model M, this denes a multinomial model with a likelihood given by

m
P (D | ) = ini , (1.8)
i=1

where ni is the number of times symbol i appears in D. The negative log-posterior is

then 
log P ( | D) = ni log i log P () + log P (D). (1.9)
i
If we assume a uniform prior distribution over the parameters , then the MAP
parameter estimation problem is identical to the ML parameter estimation problem
and can be solved by optimizing the Lagrangian
   
L= ni log i 1 i (1.10)
i i
associated with the negative log-likelihood and augmented by the normalizing con-
straint (see in addition the exercises at the end of Chapter 7). Setting the partial
MATHEMATICAL BACKGROUND 7

derivatives L/i to zero immediately yields i = ni /. Using the normalization

constraint gives = n, so that nally, as expected, we get the estimates
ni
iML = for all ai A. (1.11)
n
Note that the value of the negative log-likelihood per letter, for the optimal parameter
set ML , approaches the entropy (see Appendix A) H ( ML ) of ML as n :
1 ni 
lim ni log = iML log iML = H ( ML ). (1.12)
n n n
i i

Another way of looking at these results is to say that, except for a constant entropy
term, the negative log-likelihood is essentially the relative entropy between the xed
die probabilities pi and the observed frequencies ni /n. In Chapter 7 we will see how
this idea can be used to quantify how well various sequential models can predict
which Web page a Web surfer will request next.
The observed frequency estimate iML = ni /n is of course intuitive when n is
large. The strong law of large numbers tells us that for large enough values of n, the
observed frequency will almost surely be very close to the true value of i . But what
happens if n is small relative to m and some of the symbols (or words) are never
observed in D? Do we want to set the corresponding probability to zero? In general
this is not a good idea, since it would lead us to assign probability zero to any new
data set containing one or more symbols that were not observed in D.
This is a problem that is most elegantly solved by introducing a Dirichlet prior
on the space of parameters (Berger 1985; MacKay and Peto 1995a). This approach
is used again in Chapter 7 for modeling Web surng patterns with Markov chains
and the basic denitions are described in Appendix A. A Dirichlet distribution on a
probability vector = (1 , . . . , m ) with parameters and q = (q1 , . . . , qm ) has
the form
q 1
()  qi 1  i i
m m
Dq ( ) =  i = , (1.13)
i (qi ) i=1 i=1
Z(i)
 
with , i , qi  0 and i = qi = 1. Alternatively, it can be parameterized by
a single vector , with i = qi . When m = 2, it is also called a Beta distribution
(Figure 1.1). For a Dirichlet distribution, E(i ) = qi , var(i ) = qi (1 qi )/( + 1),
and cov(i j ) = qi qj /( + 1). Thus, q is the mean of the distribution, and
determines how peaked the distribution is around its mean. Dirichlet priors are the
natural conjugate priors for multinomial distributions. In general, we say that a prior
is conjugate with respect to a likelihood when the functional form of the posterior is
identical to the functional form of the prior. Indeed, the likelihood in Equation (1.8) and
the Dirichlet prior have the same functional form. Therefore the posterior associated
with the product of the likelihood with the prior has also the same functional form
and must be a Dirichlet distribution itself.
8 PARAMETER ESTIMATION FROM DATA
10 10

8 8

6 6
1 = 0.5, 2 = 0.8 1 = 1, 2 = 1
4 4

2 2

0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

10 10

8 8

6 6
1 = 8, 2 = 8 1 = 20, 2 = 3
4 4

2 2

0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

Figure 1.1 Beta distribution or Dirichlet distribution with m = 2. Different shapes can be
obtained by varying the parameters 1 and 2 . For instance, 1 = 2 = 1 corresponds to
the uniform distribution. Likewise, 1 = 2  = 1 corresponds to a bell-shaped distribution
centered on 0.5, with height and width controlled by 1 + 2 .

Indeed, with a Dirichlet prior Dq ( ) the negative log-posterior becomes


log P ( | D) = [ni + qi 1] log i + log Z + log P (D). (1.14)
i

Z is the normalization constant of the Dirichlet distribution and does not depend
on the parameters i . Thus the MAP optimization problem is very similar to the
one previously solved, except that the counts ni are replaced by ni + qi 1. We
immediately get the estimates

ni + qi 1
iMAP = for all ai a, (1.15)
n+m
provided this estimate is positive. In particular, the effect of the Dirichlet prior is
equivalent to adding pseudocounts to the observed counts. When q is uniform, we
say that the Dirichlet prior is symmetric. Notice that the uniform distribution over
is a special case of a symmetric Dirichlet prior, with qi = 1/ = 1/m. It is also clear
from (1.14) that the posterior distribution P (M | D) is a Dirichlet distribution Dr
with = n + and ri = (ni + qi )/(n + ).
The expectation of the posterior is the vector ri , which is slightly different from the
MAP estimate (1.8). This suggests using an alternative estimate for i , the predictive
MATHEMATICAL BACKGROUND 9
2

1.5 PRIOR:
= 2, q1 = 0.5
1

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

LIKELIHOOD:
0.04
n = 10, n1 = 1
0.03
ML = 0.1
0.02
0.01

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

POSTERIOR:
4 = 12, r1 = 0.167
MAP = 0.1
2
MP = 0.167

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 1.2 An illustration of parameter estimation with m = 2 and a at prior.

distribution or MP (mean posterior) estimate

ni + qi
iMP = . (1.16)
n+
This is often a better choice. Here, in particular, the MP estimate minimizes the
expected relative entropy distance (see Appendix A) f ( MP ) = E(H ( , MP )),
where the expectation is taken with respect to the posterior P ( | D).
Figure 1.2 illustrates a simple example of parameter estimation for this model for
an alphabet with only m = 2 symbols. In this example we have a at Beta prior
( = 0.5) and our data consist of n = 10 data points in total, with only n1 = 1
observations for the rst of the two words. Since the prior is at, the posterior Beta
distribution has the same shape as the likelihood function. Figure 1.3 shows the same
inference problem with the same data, but with a different prior now the prior is
stronger and favors a parameter that is around 0.5. In this case the likelihood and
the posterior have different shapes and the posterior in effect lies between the shapes
of the likelihood and the prior.
The simple die model, where there is no memorybetween the events, is also called
a Markov model of order zero. Naturally if we could somehow model the sequential
dependence between symbols, assuming such dependence exists, we could get a
10 MIXTURE MODELS AND THE EM ALGORITHM
3
PRIOR:
2 = 10, q1 = 0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.04 LIKELIHOOD:
n = 10, n1 = 1
0.03
ML = 0.1
0.02
0.01

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

POSTERIOR:
4
= 20, r1 = 0.3
3
MAP = 0.278
2
MP = 0.3
1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 1.3 An illustration of parameter estimation with m = 2 and a nonat prior.

better model and make more accurate predictions on future data. A standard class of
models for sequence data is the nite-state Markov chain, described in Appendix A.
The application of Markov chains to the problem of inferring the importance of Web
pages from the Web graph structure is described in the context of the PageRank
algorithm in Chapter 5. A second application of Markov chains to the problem of
predicting the page that a Web surfer will request next is treated in detail in Chapter 7.

1.3 Mixture Models and the Expectation

Maximization Algorithm
One way to build complex probabilistic models out of simpler models is the concept
of a mixture. In mixture models, a complex distribution P is parameterized as a linear
convex combination of simpler or canonical distributions in the form


K
P = l Pl , (1.17)
l=1
MATHEMATICAL BACKGROUND 11

where the i  0 are called the mixture coefcients and satisfy i i = 1. The dis-
tributions Pi are called the components of the mixture and have their own parameters
(means, standard deviations, etc.). Mixture distributions provide a exible way for
modeling complex distributions, combining together simple building-blocks, such as
Gaussian distributions. A review of mixture models can be found in Everitt (1984),
Titterington et al. (1985), McLachlan and Peel (2000), and Hand et al. (2001). Mix-
ture models are used, for instance, in clustering problems, where each component in
the mixture corresponds to a different cluster from which the data can be generated
and the mixture coefcients represent the frequency of each cluster.
To be more precise, imagine a data set D = (d1 , . . . , dn ) and an underlying mixture
model with K components of the form


K 
K
P (di | M) = P (Ml )P (di | Ml ) = l P (di | Ml ), (1.18)
l=1 l=1

where l  0, l l = 1, and Ml is the model for mixture l. Assuming that the data
points are conditionally independent given the model, we have

n
P (D | M) = P (di | M).
i=i

The Lagrangian associated with the log-likelihood and with the normalization con-
straints on the mixing coefcients is given by


n 
K   
K 
L= log l P (di | Ml ) + 1 l (1.19)
i=1 l=1 l=1

with the corresponding critical equation

L  P (di | Ml )
n
= = 0. (1.20)
l P (di )
i=1

Multiplying each critical equation by l and summing over l immediately yields the
value of the Lagrange multiplier = n. Multiplying the critical equation again by
P (Ml ) = l , and using Bayes theorem in the form
P (Ml )
P (Ml | di ) = P (di | Ml ) (1.21)
P (di )
yields
1
n
ML
l = P (Ml | di ). (1.22)
n
i=1

Thus, the ML estimate of the mixing coefcients for class l is the sample mean of
the conditional probabilities that di comes from model l. Observe that we could have
12 MIXTURE MODELS AND THE EM ALGORITHM

estimated these mixing coefcients using the MAP framework and, for example, a
Dirichlet prior on the mixing coefcients.
Consider now that each model Ml has its own vector of parameters l . Differenti-
ating the Lagrangian with respect to each parameter of each component lj gives

L  l P (di | Ml )
n
= . (1.23)
lj P (di ) lj
i=1

Substituting Equation (1.21) into Equation (1.23) nally provides the critical equation

n
log P (di | Ml )
P (Ml | di ) =0 (1.24)
lj
i=1

for each l and j . The ML equations for estimating the parameters are weighted aver-
ages of the ML equations
log P (di | Ml )
=0 (1.25)
lj
arising from each point separately. As in Equation (1.22), the weights are the proba-
bilities of di being generated from model l.
The ML Equations (1.22) and (1.24) can be used iteratively to search for ML
estimates. This is a special case of a more general algorithm known as the Expectation
Maximization (EM) algorithm (Dempster et al. 1977). In its most general form the EM
algorithm is used for inference in the presence of missing data given a probabilistic
model for both the observed and missing data. For mixture models, the missing data
are considered to be a set of n labels for the n data points, where each label indicates
which of the K components generated the corresponding data point.
The EM algorithm proceeds in an iterative manner from a starting point. The starting
point could be, for example, a randomly chosen setting of the model parameters. Each
subsequent iteration consists of an E step and M step. In the E step, the membership
probabilities p(Ml | di ) of each data point are estimated for each mixture component.
The M step is equivalent to K separate estimation problems with each data point
contributing to the log-likelihood associated with each of the K components with
a weight given by the estimated membership probabilities. Variations of the M step
are possible depending, for instance, on whether the parameters lj are estimated by
gradient descent or by solving Equation (1.24) exactly.
A different avor of this basic EM algorithm can be derived depending on whether
the membership probabilities P (Ml | di ) are estimated in hard (binary) or soft (actual
posterior probabilities) fashion during the E step. In a clustering context, the hard
version of this algorithm is also known as K-means, which we discuss later in the
section on clustering.
Another generalization occurs when we can specify priors on the parameters of the
mixture model. In this case we can generalize EM to the MAP setting by using MAP
equations for parameter estimates in the M step of the EM algorithm, rather than ML
equations.
MATHEMATICAL BACKGROUND 13

2 5
Figure 1.4 A simple Bayesian network with ve nodes and ve random variables and
the global factorization property, P (X1 , X2 , X3 , X4 , X5 ) = P (X1 )P (X2 )P (X3 | X1 , X2 )
P (X4 | X3 )P (X5 | X3 , X4 ), associated with the Markov independence assumptions. For
instance, conditioned on X3 , X5 is independent of X1 and X2 .

1.4 Graphical Models

Probabilistic modeling in complex domains leads to high-dimensional probability
distributions over data, model parameters, and other hidden variables. These high-
dimensional probability distributions are typically intractable to compute with. The
theory of graphical models (Lauritzen 1996; Whittaker 1990) provides a principled
graph-theoretic approach to modeling or approximating high-dimensional distribu-
tions in terms of simpler, more local, lower-dimensional distributions. Probabilistic
graphical models can be developed using both directed and undirected graphs, each
with different probabilistic semantics. Here, for simplicity, we concentrate on the
directed approach which corresponds to the Bayesian or belief network class of mod-
els (Buntine 1996; Charniak 1991; Frey 1998; Heckerman 1997; Jensen 1996; Jordan
1999; Pearl 1988; Whittaker 1990).

1.4.1 Bayesian networks

A Bayesian network model M consists of a set of random variables X1 , . . . , Xn and an
underlying directed acyclic graph (DAG) G = (V , E), such that each random variable
is uniquely associated with a vertex of the DAG. Thus, in what follows we will use
variables and nodes interchangeably. The parameters of the model are the numbers
that specify the local conditional probability distributions P (Xi | Xpa[i] ), 1
i
n,
where Xpa[i] denotes the parents of node i in the graph (Figure 1.4). The fundamental
property of the model M is that the global probability distribution must be equal to
the product of the local conditional distributions

P (X1 , . . . , Xn ) = P (Xi | Xpa[i] ). (1.26)
i
14 GRAPHICAL MODELS
Yt2 Yt1 Y Yt+1
t

Xt2 X X X
t1 t t+1

Figure 1.5 The underlying Bayesian network structure for both hidden Markov models and
Kalman lter models. The two independence assumptions underlying both models are that
(1) the current state Xt only depends on the past state Xt1 and (2) the current state Yt
only depends on the current state Xt . In the hidden Markov model the state variables Xt are
discrete-valued, while in the Kalman lter model the state variables are continuous.

The local conditional probabilities can be specied in terms of lookup tables (for
categorical variables). This is often impractical, due to the size of the tables, requiring
in general O(k p+1 ) values if all the variables take k values and have p parents. A
number of more compact but also less general representations are often used, such
as noisy OR models (Pearl 1988) or neural-network-style representations such as
sigmoidal belief networks (Neal 1992). In these neural-network representations, the
local conditional probabilities are dened by local connection weights and sigmoidal
functions for the binary case, or normalized exponentials for the general multivariate
case. Another useful representation is a decision tree approximation (Chickering et
al. 1997), which will be discussed in more detail in Chapter 8 in the context of
probabilistic models for recommender systems.
It is easy to see why the graph must be acyclic. This is because in general it is not
possible to consistently dene the joint probability of the variables in a cycle from
the product of the local conditioning probabilities. That is, in general, the product
P (X2 | X1 )P (X3 | X2 )P (X1 | X3 ) does not consistently dene a distribution on
X1 , X2 , X3 .
The direction of the edges can represent causality or time course if this interpretation
is natural, or the direction can be chosen more for convenience if an obvious causal
ordering of the variables is not present.
The factorization in Equation (1.26) is a generalization of the factorization in sim-
ple Markov chain models, and it is equivalent to any one of a set of independence
properties which generalize the independence properties of rst-order Markov chains,
where the future depends on the past only through the present (see Appendix A). For
instance, conditional on its parents, a variable Xi is independent of all other nodes,
except for its descendants. Another equivalent statement is that, conditional on a set
of nodes I , Xi is independent of Xj if and only if i and j are d-separated, that is,
if there is no d-connecting path from i to j with respect to I (Charniak 1991; Pearl
1988).
A variety of well-known probabilistic models can be represented as Bayesian net-
works, including nite-state Markov models, mixture models, hidden Markov models,
Kalman lter models, and so forth (Baldi and Brunak 2001; Bengio and Frasconi 1995;
Ghahramani and Jordan 1997; Jordan et al. 1997; Smyth et al. 1997). Representing
MATHEMATICAL BACKGROUND 15

these models as Bayesian networks provides a unifying language and framework for
what would otherwise appear to be rather different models. For example, both hid-
den Markov and Kalman lter models share the same underlying Bayesian network
structure as depicted in Figure 1.5.

1.4.2 Belief propagation

A fundamental operation in Bayesian networks is the propagation of evidence, which

consists of updating the probabilities of sets of variables conditioned on other vari-
ables whose values are observed or assumed. An example would be calculating
P (X1 | x4 , x42 ), where the x-values represent specic values of the observed vari-
ables and where we could have a model involving (say) 50 variables in total thus,
there are 47 other variables whose values we are not interested in here and whose
values must be averaged over (or marginalized out) if these values are unknown.
This process is also known as belief propagation or inference. Belief propagation
is NP-complete in the general case (Cooper 1990). But for singly connected graphs
(no more than one path between any two nodes in the underlying undirected graph),
propagation can be executed in time linear in n, the number of nodes, using a sim-
ple message-passing approach (Aji and McEliece 2000; Pearl 1988). In the general
case, all known exact algorithms for multiply connected networks rely on the con-
struction of an equivalent singly connected network, the junction tree. The junction
tree is constructed by clustering the original variables according to the cliques of the
corresponding triangulated moral graph as described in Pearl (1988), Lauritzen and
Spiegelhalter (1988), and Shachter (1988) with renements in Jensen et al. (1990),
and Dechter (1999). A similar algorithm for the estimation of the most probable con-
guration of a set of variables, given observed or assumed values of other variables,
is given in Dawid (1992). Shachter et al. (1994) show that all of the known exact
inference algorithms are equivalent in some sense to the algorithms in Jensen et al.
(1990) and Dawid (1992).
In practice, belief propagation is often tractable in sparsely connected graphs.
Although it is known to be NP-complete in the general case, approximate propa-
gation algorithms have been derived using a variety of methods including Monte
Carlo methods such as Gibbs sampling (Gilks et al. 1994; York 1992), mean eld
methods, and variational methods (Ghahramani 1998; Jaakkola and Jordan 1997;
Saul and Jordan 1996). An important observation, supported by both empirical evi-
dence and results in coding theory, is that the simple message-passing algorithm
of Pearl (1988) yields reasonably good approximations for certain classes of net-
works in the multiply connected case (see McEliece et al. 1997 for details). More
recent results on belief propagation, and approximation methods for graphical models
with cycles, can be found in Weiss (2000), Yedidia et al. (2000), Aji and McEliece
(2000), Kask and Dechter (1999), McEliece and Yildirim (2002) and references
therein.
16 GRAPHICAL MODELS

Table 1.1 Four basic Bayesian network learning problems depending on whether the structure
of the network is known in advance or not, and whether there are hidden (unobserved) variables
or not.

No hidden Hidden

Known structure easy doable (via EM)

Unknown structure doable hard

1.4.3 Learning directed graphical models from data

There are several levels of learning in graphical models in general and Bayesian
networks in particular. These range from learning the entire graph structure (the
edges in the model) to learning the local conditional distributions when the structure is
known. To a rst approximation, four different situations can be considered, depending
on whether the structure of the network is known or not, and whether the network
contains unobserved data, such as hidden nodes (variables), which are completely
unobserved in the model (Table 1.1).
When the structure is known and there are no hidden variables, the problem is a rel-
atively straightforward statistical question of estimating probabilities from observed
frequencies. The die example in this chapter is an example of this situation, and we
discussed how ML, MAP, and MP ideas can be applied to this problem. At the other
end of the spectrum, learning both the structure and parameters of a network that con-
tains unobserved variables can be a very difcult task. Reasonable approaches exist
in the two remaining intermediary cases. When the structure is known but contains
hidden variables, algorithms such as EM can be applied, as we have described earlier
for mixture models. Considerably more details and other pointers can be found in
Buntine (1996), Heckerman (1997, 1998), and Jordan (1999).
When the structure is unknown, but no hidden variables are assumed, a variety
of search algorithms can be formulated to search for the structure (and parameters)
that optimize some particular objective function. Typically these algorithms operate
by greedily searching the space of possible structures, adding and deleting edges to
effect the greatest change in the objective function. When the complexity (number
of parameters) of a model M is allowed to vary, using the likelihood as the objective
function will inevitably lead to the model with the greatest number of parameters,
since it is this model that can t the training data D the best. In the case of searching
through a space of Bayesian networks, this means that the highest likelihood network
will always be the one with the maximal number of edges. While such a network
may provide a good t to the training data, it may in effect overt the data. Thus,
it may not generalize well to new data and will often be outperformed by more
parsimonious (sparse) networks. Rather than selecting the model that maximizes
the likelihood P (D | M), the Bayesian approach is to select the model with the
maximum posterior probability given the data, P (M | D), where we average over
MATHEMATICAL BACKGROUND 17

parameter uncertainty

P (M | D) P (D | , M)P ( | M) d ,

where P (D | , M) is the likelihood and P ( | M) is the prior (here we explicitly

include both and M). There is an implicit penalty in effect here. By averaging over
parameter space, instead of picking the most likely parameter, more complex models
will in effect be penalized for having a higher-dimensional parameter space and will
only win if their predictive power on the training data (measured via the likelihood)
can overcome the inherent penalty that arises from having to integrate over a higher-
dimensional space. Heckerman (1998) provides a full discussion of how this Bayesian
estimation approach can be used to automatically construct Bayesian networks from
data.

1.5 Classication
Classication consists of learning a mapping that can classify a set of measurements
on an object, such as a d-dimensional vector of attributes, into one of a nite number
K of classes or categories. This mapping is referred to as a classier and it is typically
learned from training data. Each training instance (or data point) consists of two parts:
an input part x and an associated output class target c, where c {1, 2, . . . , K}. The
classier mapping can be thought of as a function of the inputs, g(x), which maps
any possible input x into an output class c.
For example, email ltering can be formulated as a classication problem. In this
case the input x is a set of attributes dened on an email message, and c = 1 if and
only if the message is spam. We can construct a training data set of pairs of email
messages and labels:
D = {[x1 , c1 ], . . . , [xn , cn ]}.
The class labels in the training data c can be obtained by having a human manually
label each email message as spam or non-spam.
The goal of classication learning is to take a training data set D and estimate
the parameters of a classication function g(x). Typically we seek the best function
g, from some set of candidate functions, that minimizes an empirical loss function,
namely
n
E= l(ci , g(xi )),
i=1

where l(ci , g(xi )) is dened as the loss that is incurred when our predicted class label
is g(xi ), given input xi , and the true class label is ci . A widely used loss function for
classication is the so-called 0-1 loss function, where l(a, b) is zero if a = b and
one otherwise, or in other words we incur a loss of zero when our prediction matches
the true class label and a loss of one otherwise. Other loss functions may be more
18 CLASSIFICATION

appropriate in certain situations. For example, in email ltering we might want to

assign a higher penalty or loss to classifying as spam an email message that is really
non-spam, versus the other type of error, namely classifying an email message as
non-spam that is really spam.
Generally speaking, there are two broad approaches to classication, the proba-
bilistic approach and the discriminative or decision-boundary approach. In the prob-
abilistic approach we can learn a probability distribution P (x | c) for each of the
K classes, as well as the marginal probability for each class P (c). This can be done
straightforwardly by dividing the training data D into K different subsets according
to class labels, assuming some functional form for P (x | c) for each class, and then
using ML or Bayesian estimation techniques to estimate the parameters of P (x | c)
for each of the K classes. Once these are known we can then use Bayes rule to
calculate the posterior probability of each of the classes, given an input x:
P (x | c = k)P (c = k)
P (c = k | x) = K , 1
k
K. (1.27)
j =1 P (x | c = j )P (c = j )

To make a class label prediction for a new input that is not in the training data x, we
calculate P (c = k | x) for each of the K classes. If we are using the 0-1 loss function,
then the optimal decision is to choose the most likely class, i.e.
c = arg max{P (c = k | x)}.
k

An example of this general approach is the so-called Naive Bayes classier, to be

discussed in more detail in Chapter 4 in the context of document classication, where
the assumption is made that each of the individual attributes in x are conditionally
independent given the class label

m
P (x | c = k) = p(xj | c = k), 1
k
K,
j =1

where xj is the j th attribute and m is the total number of attributes in the input x.
A limitation of this general approach is that by modeling P (x | c = k) directly,
we may be doing much more work than is necessary to discriminate between the
classes. For example, say the number of attributes is m = 100 but only two of these
attributes carry any discriminatory information the other 98 are irrelevant from the
point of view of making a classication decision. A good classier would ignore these
98 features. Yet the full probabilistic approach we have prescribed here will build a
full 100-dimensional distribution model to solve this problem. Another way to state
this is that by using Bayes rule, we are solving the problem somewhat indirectly: we
are using a generative model of the inputs and then inverting this via Bayes rule to
get P (c = k | x).
A probabilistic solution to this problem is to instead focus on learning the posterior
(conditional) probabilities P (c = k | x) directly, and to bypass Bayes rule. Con-
ceptually, this is somewhat more difcult to do than the previous approach, since the
MATHEMATICAL BACKGROUND 19

training data provide class labels but do not typically provide examples of values
of P (c = k | x) directly. One well-known approach in this category is to assume a
logistic functional form for P (c | x),
1
P (c = 1 | x) = ,
1 + ew
T xw
0

where for simplicity we assume a two-class problem and where w is a weight vector
of dimension d, wT is the transpose of this vector, and w T x is the scalar inner product
of w and x. Equivalently, we can represent this equation in log-odds form
P (c = 1 | x)
log = w T x + w0 ,
1 P (c = 1 | x)
where now the role of the weights in the vector w is clearer: a large positive (negative)
weight wj for attribute xj means that, as xj gets larger, the probability of class c1
increases (decreases), assuming all other attribute values are xed. Estimation of
the weights or parameters of this logistic model from labeled data D can be carried
out using iterative algorithms that maximize ML or Bayesian objective functions.
Multilayer perceptrons, or neural networks, can also be interpreted as logistic models
where multiple logistic functions are combined in various layers.
The other alternative to the probabilistic approach is to simply seek a function
f that optimizes the relevant empirical loss function, with no direct reference to
probability models. If x is a d-dimensional vector where each attribute is numerical,
these models can often be interpreted as explicitly searching for (and dening) deci-
sion regions in the d-dimensional input space x. There are many non-probabilistic
classication methods available, including perceptrons, support vector machines and
kernel approaches, and classication trees. In Chapter 4, in the context of document
classication, we will discuss one such method, support vector machines, in detail.
The advantage of this approach to classication is that it seeks to directly maximize
the chosen loss function for the problem, and no more. In this sense, if the loss function
is well dened, the direct approach can in a certain sense be optimal. However, in many
cases the loss function is not known precisely ahead of time, or it may be desirable to
have posterior class probabilities available for various reasons, such as for ranking or
for passing probabilistic information on to another decision making algorithm. For
example, if we are classifying documents, we might not want the classier to make
any decision on documents whose maximum class probability is considered too low
(e.g. less than 0.9), but instead to pass such documents to a human decision-maker
for closer scrutiny and a nal decision.
Finally, it is important to note that our ultimate goal in building a classier is to
be able to do well on predicting the class labels of new items, not just the items in
the training data. For example, consider two classiers where the rst one is very
simple with only d parameters, one per attribute, and the second is very complex with
100 parameters per attribute. Further assume that the functional form of the second
model includes the rst one as a special case. Clearly the second model can in theory
always do as well, if not better than, the rst model, in terms of tting to the training
20 CLUSTERING

data. But the second model might be hopelessly overtting the data and on new unseen
data it might produce less accurate predictions than the simpler model.
In this sense, minimizing the empirical loss function on the training data is only a
surrogate for what we would really like to minimize, namely the expected loss over all
future unseen data points. Of course this future loss is impossible to know. Nonetheless
there is a large body of work in machine learning and statistics on various methods
that try to estimate how well we will do on future data using only the available training
data D and that then use this information to construct classiers that generalize more
accurately. A full discussion of these techniques is well beyond the scope of this book,
but an excellent treatment for the interested reader can be found in Hastie et al. (2001)
and Devroye et al. (1996).

1.6 Clustering
Clustering is very similar to classication, except that we are not provided with
class labels in the training data. For this reason classication is often referred to as
supervised learning and clustering as unsupervised learning. Clustering essentially
means that we are trying to nd natural classes that are suggested by the data.
The problem is that this is a rather vague prescription and as a result there are many
different ways to dene what precisely is meant by a cluster, the quality of a particular
clustering of the data, and algorithms to optimize a particular cluster qualityfunction
or objective function given a set of data. As a consequence there are a vast number of
different clustering algorithms available. In this section we briey introduce two of
the more well-known methodologies and refer the reader to other sources for more
complete discussions (Hand et al. 2001; Hastie et al. 2001).
One of the simplest and best known clustering algorithms is the K-means algorithm.
In a typical implementation the number of clusters is xed a priori to some value K.
K representative points or centers are initially chosen for each cluster more or less at
random. These points are also called centroids or prototypes. Then at each step:
(i) each point in the data is assigned to the cluster associated with the closest
representative;
(ii) after the assignment, new representative points are computed for instance by
averaging or by taking the center of gravity of each computed cluster;
(iii) the two procedures above are repeated until the system converges or uctuations
remain small.
Notice that the K-means algorithm requires choosing the number of clusters, being
able to compute a distance or similarity between points, and computing a representa-
tive for each cluster given its members.
From this general version of K-means one can dene different algorithmic varia-
tions, depending on how the initial centroids are chosen, how symmetries are broken,
whether points are assigned to clusters in a hard or soft way, and so forth. A good
MATHEMATICAL BACKGROUND 21

implementation ought to run the algorithm multiple times with different initial con-
ditions, since any individual run may converge to a local extremum of the objective
function.
We can interpret the K-means algorithm as a special case of the EM algorithm for
mixture models described earlier in this chapter. More specically, the description
of K-means given above corresponds to a hard version of EM for mixture models,
where the membership probabilities are either zero or one, each point being assigned
to only one cluster. It is well known that the center of gravity of a set of points
minimizes its average quadratic distance to any xed point. Therefore, in the case of a
mixture of spherical Gaussians, the M step of the K-means algorithm described above
maximizes the corresponding quadratic log-likelihood and provides an ML estimate
for the center of each Gaussian component.
When the objective function corresponds to an underlying probabilistic mixture
model (Everitt 1984; McLachlan and Peel 2000; Titterington et al. 1985), K-means
is an approximation to the classical EM algorithm (Baldi and Brunak 2001; Dempster
et al. 1977), and as such it typically converges toward a solution that is at least a local
ML or maximum posterior solution. A classical case is when Euclidean distances are
used in conjunction with a mixture of Gaussian models.
We can also use the EM algorithm for mixture models in its original form with
probabilistic membership weights to perform clustering. This is sometimes referred
to as model-based clustering. The general procedure is again to x K in advance,
select specic parametric models for each of the K clusters, and then use EM to
learn the parameters of this mixture model. The resulting component models provide
a generative probabilistic model for the data in each cluster. For example, if the data
points are real-valued vectors in a d-dimensional space, then each component model
can be a d-dimensional Gaussian or any other appropriate multivariate distribution.
More generally, we can use mixtures of graphical models where each mixture com-
ponent is encoded as a particular graphical model. A particularly simple and widely
used graphical model is the so-called Naive Bayes model, described in the previ-
ous section on classication, where the attributes are assumed to be conditionally
independent given the component.
A useful feature of model-based clustering (and mixture models in general) is that
the data need not be in vector form to dene a clustering algorithm. For example,
we can dene mixtures of Markov chains to cluster sequences, where the underlying
model assumes that the observed data are being generated by K different Markov
chains and the goal is to learn these Markov models without knowing which sequence
came from which model. In Chapter 7 we will describe this idea in more detail in an
application involving clustering of Web users based on their Web page requests.
The main limitation of model-based clustering is the requirement to specify a
parametric functional form for the probability distribution of each component in
some applications this may be difcult to do. However, if we are able to assume some
functional form for each cluster, the gains can be substantial. For example, we can
include a component in the model to act as a background cluster to absorbdata points
that are outliers or that do not t any of the other components very well. We can also
22 POWER-LAW DISTRIBUTIONS

estimate the posterior distribution P (K | D) for different numbers of clusters K to

see what value of K is most likely given the data (see Fraley and Raftery (2002) for a
variety of such extensions).Another drawback of the standard mixture model approach
to clustering, and indeed of many other approaches to clustering, is the assumption
that each data point was generated by one and only one cluster. In clustering text
documents on the Web, for example, we might want to allow documents to belong to
multiple clusters, e.g. sportsand nance. Recent proposals by Hofmann (2001) and
Blei et al. (2002a) develop probabilistic models to specically address this issue we
will discuss such models in Chapter 4 for modeling text documents and in Chapter 8
in the context of recommender algorithms.

1.7 Power-Law Distributions

1.7.1 Denition
Power-law distributions arise naturally in a variety of domains. In the case of the
Web, power-law distributions appear to be a ubiquitous phenomenon and have been
reported, for instance, for
Web page sizes;
Web page connectivity;
Web connected components size;
Web page access statistics;
Web browsing behavior.
Each one of these examples will be studied in great detail in the coming chapters.
Continuous or discrete power-law distributions are characterized by the fact that
for large values of the independent variable x the distribution decays polynomially
as x , with > 1. This is in sharp contrast with many other standard distributions,
such as exponential, Poisson, Binomial, Gaussian, and Gamma, where the asymptotic
decay of the distribution is exponential. This is intuitively summarized by saying that
in a power-law distribution, rare events are not so rare.
Power-law distributions on ranked data are often called Zipf distributions, or
ParetoZipf distributions. In language applications, for instance, the distribution of
words ranked by their frequency in a large corpus of text is invariably a power-law
distribution, also known as Zipfs Law. There is a vast literature on power-law distri-
butions and the related log-normal distribution (e.g. Mihail and Papadimitriou 2002;
Mitzenmacher 2002).
A discrete power-law distribution with coefcient > 1 is a distribution of the
form
P (X = k) = Ck (1.28)
MATHEMATICAL BACKGROUND 23

for k = 1, 2, . . . . The corresponding density in the continuous case is

f (x) = Cx (1.29)

for x [1, +). In many real life situations associated with power-law distributions,
the distribution for small values of k or x may deviate from the expressions in Equa-
tions (1.28) and (1.29). Thus, a more exible denition is to say that Equations (1.28)
and (1.29) describe the behavior for sufciently large values of x or k.
In a loglog plot, the signature of a power-law distribution is a line with a slope
determined by the coefcient . This provides a simple means for testing power-law
behavior and for estimating the exponent . An example of ML estimation of is
provided in Chapter 7.
In both the discrete and continuous cases, moments of order m  0 are nite if and
only if > m + 1. In particular, the expectation is nite if and only if > 2, and the
variance is nite if and only if > 3. In the discrete case,


E[Xm ] = Ck m = C ( m), (1.30)
k=1

where is Riemanns zeta function ( (s) = k 1/k s ). In the continuous case,

C
E[X m ] = Cx m dx = . (1.31)
1 m1
In particular, C = 1, E[X] = 1/( 2) and
 
1 1 2
var[X] = .
3 2
A simple consequence is that in a power-law distribution the average behavior is
not the most frequent or the most typical, in sharp contrast with what is observed
with, for instance, a Gaussian distribution. Furthermore, the total mass of the points
to the right of the average is greater than the total mass of points to the left. Thus, in
a random sample, the majority of points are to the right of the average.
Another interesting property of power-law distributions is that the cumulative tail,
that is the area under the tail of the density,

also has a power-law behavior with
exponent 1. This is obvious since x u du = x +1 /( 1).
There is also a natural connection between power-law distributions and log-normal
distributions (see Appendix A). X has a log-normal density if log X has a normal
density with mean and variance 2 . This implies that the density of X is
1
e(log x)
2 /2 2
f (x) = (1.32)
2 x
and thus log f (x) = C log x (log x )2 /2 2 . In the range where is large
compared with |log x |, we have log f (x) C log x, which corresponds to a
straight line. Very basic generative models can lead to log-normal or to power-law
24 POWER-LAW DISTRIBUTIONS

distributions depending on small variations and therefore it is not surprising that in a

number of situations both models have been proposed to capture the same phenomena,
such as the distributions of computer le sizes (see, for instance, Perline 1996 and
Mitzenmacher 2002).

1.7.2 Scale-free properties (80/20 rule)

Power-law distributions result in scale-free properties in the sense that, for 1
a
b,


dx  1
b C/i b Cx b

= . (1.33)
a
a C/i a Cx dx

In other words, the ratio of the sums depends only on the ratio b/a and not on the
absolute scale of a or b things look the same at all length scales. This scale-free
property is also known in more folkloristic terms as the 80/20 rule, which basically
states that 80% of your business comes from 20% of your clients. Of course it is
not the particular value 80/20 = 4/1 ratio that matters, since this may vary from one
phenomenon to the next, it is the existence of such a ratio across different scales, or
business sizes (and business types). This property is also referred to as self-scaling
or self-similar, since the ratio of the areas associated with a and b remains constant,
regardless of their actual values or frequencies. One important consequence of this
behavior is that observations made at one scale can be used to infer behavior at other
scales.

1.7.3 Applications to Languages: Zipfs and Heaps Laws

As briey mentioned above, a power-law distribution on ranked data is also called
a Zipf Law. If, for instance, one plots the frequency rank of words contained in a
large corpus of text data versus the number of occurrences or actual frequencies,
one invariably obtains a power-law distribution, with exponent close to one (but
see Gelbukh and Sidorov 2001). An example is plotted in Figure 1.6. Under simple
assumptions, this is also observed in random text generated by a Markov model (Li
1992).
A second related law for textual data is Heaps Law. Heaps Law is the observation
that asymptotically, a corpus of text containing N words typically contains on the order
of CN distinct words, with 0 < < 1. Empirically, was found to be between 0.4
and 0.6 by some authors (Arajo et al. 1997; Moura et al. 1997). However, the value
of the exponent varies somewhat with the language, the type of data set, and so forth.
Using the Web data set from the Web KB (World Wide Knowledge Base) project
at Carnegie Mellon University, consisting of several thousand Web pages collected
from Computer Science departments of various universities, we nd a characteristic
exponent equal to 0.76 (Figure 1.7).
MATHEMATICAL BACKGROUND 25
Frequency

100000

10000

1000

100

10

Rank
1
1 10 100 1000 10000 100000

Figure 1.6 Zipfs Law for data from the Web KB project at Carnegie Mellon Univer-
sity consisting of 8282 Web pages collected from Computer Science departments of various
universities. Tokenization was carried out by removing all HTML tags and by replacing each
non-alphanumeric character by a space. Loglog plot of word ranks versus frequencies.

Vocabulary size

Empirical
100 K
0.76
|N|

80 K

60 K

40 K

20 K

Text length N
0
0 500 K 1M 1.5 M 2M 2.5 M 3M 3.5 M 4M

Figure 1.7 Heaps Law for data from Web KB project at Carnegie Mellon University,
consisting of 8282 Web pages collected from Computer Science departments of various uni-
versities. Tokenization was carried out by removing all HTML tags and by replacing each
non-alphanumeric character by a space. Plot of total number of words (text length) versus the
number of distinct words (vocabulary size). The characteristic exponent for this Web data set
is 0.76.
26 POWER-LAW DISTRIBUTIONS

1.7.4 Origin of power-law distributions and Fermis model

It is in general easy to detect power-law behavior in data: after all a power-law curve
yields a straight line in a loglog plot. Explaining the origin of the power-law behavior,
however, is a different story. Because power laws are so ubiquitous, several attempts
have been made to try to reduce their origin to a single, or perhaps a few, canonical
models. This is not completely unreasonable in the same way that the Gaussian distri-
bution is often explained by some kind of central limit theorem, i.e. it is the distribution
that emerges from the sum of a large number of more or less independent small effects.
As in the case of the Gaussian, however, it does not seem likely that a single
model can account for all occurrences of power-law distributions, and certainly not
with any precision, since it is easy to introduce small perturbations in a power-law
distribution in a large number of different ways. In Chapter 4, we will look at a simple
generative model that tries to provide an explanation for the power law observed in the
connectivity of the Web. Here we describe a very simple model that yields power-law
distributions of objects submitted to two exponential laws. The model was originally
proposed by Fermi (1949) while modeling particles and cosmic radiation.
Imagine that particles are being produced continuously at a constant rate and that
particles age, so that the current age distribution is an exponential of the form et .
As it ages, a particle gains energy exponentially, so that the energy of a particle of
age t is w(t) = Cet . By combining the distribution of particle ages with the energy
associated with each age, it is easy to nd the distribution of energies at time t. In
particular,

P (w0
w
w0 + dw) = P (w0
Cet
w0 + dw)

 w0 +dw
= (w) dw = (w0 ) dw. (1.34)
w0

This is equivalent to
   /  
1 w0 1 w0 + dw w0 w0 + dw /
P log
t
log = (1.35)
C C C C
by using the exponential distribution of t. This nally yields the density
 
w / 1
(w) = , (1.36)
C w
with power-law exponent = 1 + /. Thus a power-law density results naturally
from the competition of two exponential phenomena: a positive exponential control-
ling the growth of the energy and a negative exponential controlling the age distribu-
tion.
A different generative model of power-law behavior based on the notion of prefer-
ential attachment will be given in Chapter 3. A third and more indirect model based
on information theory is due to Mandelbrot (see Mandelbrot 1977 and references
MATHEMATICAL BACKGROUND 27

therein; Carlson and Doyle 1999; Zhu et al. 2001) and is given in one of the exercises
below.

1.8 Exercises
Exercise 1.1. Prove that Equations (1.15) and (1.16) are correct.
Exercise 1.2. In the die model, compute the posterior when the prior is a mixture of
Dirichlet distributions. Can you infer a general principle?
Exercise 1.3. The die model with a single Dirichlet prior is simple enough that one
can proceed with higher levels of Bayesian inference, beyond the computation of the
posterior or its mode. For example, explicitly compute the normalizing factor (also
called the evidence) P (D). Use the evidence to derive a strategy for optimizing the
parameters of the Dirichlet prior.
Exercise 1.4. Consider a mixture model with a Dirichlet prior on the mixing coef-
cients. Study the MAP and MP estimates for the mixing coefcients.
Exercise 1.5. Compute the differential entropy (see Appendix A) of the continuous
power-law distribution f (x) = Cx , x  1.
Exercise 1.6. Simulate a rst-order Markov language (by selecting an alphabet con-
taining the space symbol, dening a die probability vector over the alphabet, tossing
the die many times, and computing word statistics) whether Zipfs and Heaps Laws
are true or not in this case. Experiment with the size of the alphabet, the letter proba-
bilities, the order of the Markov model, and the length of the generated sequence.
Exercise 1.7. Find a large text data set (for instance from Web KB) and compute
the corresponding Zipf and Heaps curves, e.g. word rank versus word frequency and
text size versus vocabulary, or the respective logarithms.
Exercise 1.8. Study the connection between Zipfs and Heaps Laws. In particular,
is Heaps Law a consequence of Zipfs Law?
Exercise 1.9. Collect statistics about the size of cities, or computer les, and deter-
mine whether there is an underlying power-law distribution. Provide an explanation
for your observation.
Exercise 1.10. Mandelbrots information theoretic generative model for power-law
distributions considers random text generated using a set of M different words
with associated probabilities p1 , . . . , pM , ranked in decreasing order, and associated
transmission costs C1 , . . . , CM . A possible cost function is to take Ci proportional
to log i. Why?
Now suppose that the goal is to design the probabilities pi in order to optimize
the average amount
 of information per unit transmission cost. The average cost per
word is E[C] = i pi C i and the average information, or entropy (see Appendix A),
per word is H(p) = i pi log pi . Compute the values of the pi to maximize the
quantity H/C and show that, with logarithmic costs, it corresponds to a power law.
28 EXERCISES

Exercise 1.11. Show that the product of two independent log-normal random vari-
ables is also log-normal. Use this simple fact to derive a generative model for log-
normal distributions.
2

Basic WWW Technologies

Providing a complete description of the technology by which the Web is implemented

is beyond the scope of this book. However, a basic understanding of Web technology
is essential for the rest of this book and for implementing Web algorithms.
In simple terms, the World Wide Web is implemented by combining three essential
components: resources, resource identiers, and transfer protocols. Resources are
dened in a very general way as conceptual mappings to concrete or abstract entities
(Berners-Lee et al. 1998). In the most common case these entities correspond to, but
are not limited to, Web pages and other kinds of les, chosen from a large assortment
of proprietary and public domain formats. Note the importance of dening resources
in a conceptual rather than physical way. For example, we may say that the site of the
World Wide Web Consortium (W3C) is a resource. This conceptual entity does not
change over time, at least in the short term. The concrete entity, on the other hand,
could include sequences of bits that represent Web pages, for example stored as les
in some computer disk. This level of detail, however, is unnecessary. Moreover, the
actual les may change at any time without affecting the conceptual entity.
Resource identiers are string of characters representing generalized addresses that
may also implicitly contain precise instructions for accessing the identied resource.
For example, the string http://www.w3.org/ is used to identify the W3C web-
site. Protocols are essentially conventions that regulate the communication between
two connected entities such as a Web user agent (e.g. a browser) and a server. The
key property of the Web is that any resource may contain several references to other
resources (encoded as identiers), and it is these hyperlinks that give rise to an inter-
connected network. Protocols provide a convenient way for an agent to navigate this
network by following the referenced resources.
As pointed out in the Preface, people often tend to confuse the Internet and the
Web. On one hand, because the common use of Web services involves only a subset
of Internet protocols at the application level (see below), the Internet encompasses
something larger than just the Web. On the other hand, according to the above deni-
tion, several webs may be conceived, not just those that use Internet protocols for
fast communication. For example, the whole body of scientic literature forms a web
through the mechanism of citations, but the access protocol, in general, still implies a
Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
30 WEB DOCUMENTS

visit to the library, whether an electronic or a brick and mortar one. Hence, in turn,
under an extended set of protocols the Web encompasses more than just the Internet.

2.1 Web Documents

A vast proportion of existing Web resources are HTML documents. HTML (Hyper-
Text Markup Language) is a non-proprietary format that provides basic tagging sup-
port for structuring and laying out text and images as we commonly experience them
when using a Web browser. Briey, the language allows one to specify the formatting
of headings, paragraphs, lists, tables, and so forth, to create llable forms for elec-
tronic transactions, to include procedural scripts, and to combine text with several
types of multimedia objects such as video or audio clips. The original language was
developed by Tim Berners-Lee at CERN but was extended in several ways during the
1990s until the present version 4.01 (Raggett et al. 1999).

2.1.1 SGML and HTML

HTML is actually one application of the Standard Generalized Markup Language
(SGML), an international standard that denes how descriptive markup should be
embedded in a document (ISO 1986). Application in this context has a special mean-
ing: SGML is in fact a metalanguage, i.e. a formalism that permits us to specify lan-
guages, and HTML is a language that can be specied in SGML. In general, markup
refers to any extra information that is distinguished from the document contents and
that characterizes the structure and/or the style of the document. Descriptive markup
is used to describe different elements of a document, like chapters, sections, and
paragraphs in a book. For example, LATEX, the typesetting system that was employed
to prepare this book, extensively uses descriptive elements for dening the structure
of a document, although it is not derived from SGML. Procedural markup, on the
other hand, is a lower-level specication of the sequence of processing steps that
are necessary to render the document in its nal form. Common WYSIWYG (what
you see is what you get) applications for document editing tend to stress procedural
markup, i.e. specication of how to display the elements, while the main focus of
descriptive markup in SGML is about the structure of the document, i.e. specication
of what the elements are. More details on SGML can be found in Sperberg-McQueen
and Burnard (2002) and Raggett et al. (1999). The reader should also be aware that,
in 1996, a subset of SGML, called the eXtensible Markup Language (XML), was
developed to simplify the distribution and the processing of SGML documents on the
Web. XML is released as a W3C recommendation.
As an SGML application, HTML needs to specify a valid character set, and the
syntax and semantics of tags. In SGML, valid element types and their relations are
formally specied in a special set of declarations called the Document Type Denition
(DTD). As a loose analogy, the relationship between an SGML application and its
DTD is similar to the relationship between a programming language (like C or Java)
BASIC WWW TECHNOLOGIES 31

and its formal grammar. In other words, the DTD of HTML species the metalevel
of a generic HTML document. For example, the following declaration in the DTD
of HTML 4.01 species that an unordered list is a valid element and that it should
consist of one or more list items,
<!ELEMENT UL - - (LI)+ -- unordered list -->

where ELEMENT is a keyword that introduces a new element type, UL is the new
element type, which represents an unordered list being dened and LI is the ele-
ment type that represents a list item. Here, metasymbols (), for grouping and +, for
repetition, are similar to the corresponding metasymbols in regular expressions. In
particular, + means that the preceding symbol (i.e. LI) should appear one or more
times. The declaration of an element in a document is generally obtained by enclosing
some content within a pair of matching tags. A start tag is realized by the element
name enclosed in angle brackets, as in <LI>. The end tag is distinguished by adding a
slash before the element name, as in </LI>. The semantics in this case says that any
text enclosed between <LI> and </LI> should be treated as a list item. Note that,
besides enriching the document with structure, markup may affect the appearance of
the document in the window of a visual browser (e.g. the text of a list item should be
rendered with a certain right indentation and preceded by a bullet).
Some elements may be enriched with attributes. For example, the following is a
portion of the declaration of the element IMG, used to embed inline images in HTML
documents:
<!ELEMENT IMG - O EMPTY -- Embedded image -->
<!ATTLIST IMG
src %URI; #REQUIRED -- URI of image to embed --
alt %Text; #REQUIRED -- short description --
...
The attribute src is a required resource identier (see Section 2.2 below) that allows
us to access the le containing the image to be displayed. In the simplest case this
is just a local lename. The attribute alt is text that should be used in place of the
image, for example when the document is rendered in a terminal-based browser. An
actual image would be then inserted as
<img src="Web.png" alt="a Web">

This is an example of an element that must not use a closing tag (as implied by the
keywords O and EMPTY in the element declaration).

2.1.2 General structure of an HTML document

HTML 4.01 is a language with a fairly large set of elements which, for lack of space,
cannot be described here in complete detail. The interested reader is urged to consult
the W3C recommendation of 24 December 1999 (Raggett et al. 1999) for a complete
specication. Here we limit our discussion to the analysis of the general structure of
a document and we present a rather simple example.
32 WEB DOCUMENTS

A valid HTML document is comprised of the following three parts.

Version information. This is a declarative section that species which DTD is used
in the document (see lines 12 in Figure 2.1). The W3C recommends three possible
DTDs: strict, transitional, or frameset. Strict contains all the elements that are not
deprecated (i.e. outdated by new elements) and are not related to frames, transitional
allows deprecated elements, and frameset contains all of the above plus frames.

Header. This is a declarative section that contains metadata, i.e. data about the docu-
ment. This section is enclosed within <head> and </head> in Figure 2.1 (lines 4
8). In our example, line 5 species the character set (ISO-8859-1, aka Latin-1) and
line 6 describes the authors. Metadata in this case are specied by setting attributes
of the element meta. The title is instead an element (line 7).

Body. This is the actual document content. This element is enclosed within <body>
and </body> in Figure 2.1 (lines 918).

In Figure 2.1 we see some other examples of HTML elements. <h1> and <h2> are
used for headings, <ul> consists of an unordered list, whose (bulleted) items are
tagged by <li>, and so on.

2.1.3 Links

The <A> element (see, for example, lines 15 and 16) is used in HTML to implement
links, an essential feature of hypertexts. As we will see later in this book, a link
is rather similar to an edge in a directed graph. It connects two objects referred to
as anchors. In our example, the source anchors are the elements of the document
delimited by <A> and </A>. In the World Wide Web, the target anchor is a resource
that may be physically stored in the same server as the source document, or may be
in a different server, possibly located in another country or continent (see Figure 2.2).
The target resource is identied by means of a special string called the Uniform
Resource Identier (URI). If the attribute href of the element <A> is set, its value
is the URI of the target anchor. In our example there are two URIs, namely the two
strings http://www.w3.org/ and toc.html.
Note that links can be implemented in several other ways. For HTML documents,
source anchors can be associated with images, forms, or active elements imple-
mented using scripting languages or applets. But linking is not limited to HTML
documents, since many document formats now support mechanisms for encoding
URIs.
In 2000, HTML was refomulated as an XML application. The resulting language
was called XHTML and released as a W3C recommendation. Details can be found at
http://www.w3.org/TR/html/.
BASIC WWW TECHNOLOGIES 33

1: <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"

2: "http://www.w3.org/TR/html4/strict.dtd">
3: <html>
4: <head>
5: <meta http-equiv="Content-Type" content="text/html; charset=ISO-8859-1">
6: <meta name="Author" content="Pierre Baldi; Paolo Frasconi; Padhraic Smyth">
7: <title>Modeling The Internet and the World Wide Web</title>
8: </head>
9: <body>
10: <h1> <img src="Web.png" alt="a Web" height=64 width=64>
11: Modeling the Internet and the Web</h1>
12: <h2>Probabilistic Methods and Algorithms</h2>
13: <p>A short list of bookmarks:</p>
14: <ul>
15: <li><a href="http://www.w3.org/">The W3 Consortium</a></li>
16: <li><a href="toc.html">Table of Contents</a></li>
17: </ul>
18: </body>
19: </html>

Figure 2.1 Simple HTML document (bottom) demonstrating the use of

a few common tags. The rendered document is shown on the top.

2.2 Resource Identiers: URI, URL, and URN

Resources are referred to by means of strings with a controlled syntax and semantics,
called the resource identiers. Uniform Resource Identiers (URI) are the most gen-
eral set of identiers. Uniformity refers to the generality of these identiers and to
the property that the interpretation of each string should be unambiguous, regardless
of the context. The precise denition of URIs has evolved from the informational
document released as RFC 1630 (Berners-Lee 1994) to the draft standard RFC 2396
(Berners-Lee et al. 1998) that describes in detail the general grammar of URIs. Since
the early formalizations, the design of URIs has focused on the concepts of exten-
sibility and completeness. Extensibility means that the specication should leave
room for adding new naming schemes in the future, perhaps for types of services
that have not been invented yet. Completeness means that it must be possible to
encode every possible naming scheme. This need is justied by the denition of
34 RESOURCE IDENTIFIERS: URI, URL, AND URN
http://ibook.ics.uci.edu/

Modeling the Internet and

the Web

Probabilistic Methods and

Algorithms

A short list of bookmarks

The W3 Consortium
Table of Contents
www.w3.org

ibook.ics.uci.edu
--------
--------
-------- ------------
-------- ------------
------------
------------

http://www.w3.org/
http://ibook.ics.uci.edu/toc.html

Figure 2.2 Anchors and Web links.

resources as conceptual mappings to entities. This denition is extremely general and

not limited to electronic documents that can be accessed in a networked environment.
Everything that has an identity can be considered as a resource, including persons
or abstract entities. Indeed, URIs have been dened that identify telephone or fax
numbers (e.g. fax:+39-055-2222222), short message services in GSM mobile
phones (e.g. sms:+393301234567), or television broadcasts (e.g. tv:rai.it).
URIs actually include two overlapping subsets of identiers: Uniform Resource
Locators (URLs) and Uniform Resource Names (URNs). URLs are identiers that
explicitly encode details about the algorithm used to access the resource. For exam-
ple, the URI ftp://ftp.ietf.org/rfc/rfc2396.txt is a URL because it
species that the document containing the text of RFC 2396 can be accessed via le
transfer protocol from the ftp server of the Internet Engineering Task Force. URNs,
on the other hand, are those URIs that must persist and remain unique even if the
resource becomes unavailable.
In order to maximize the ability of transcribing them, URIs are strings over a
minimal subset of the 7-bit ASCII character set. This allows us, for example, to
advertise URIs via traditional media such as radio or printed newspapers. A small
number of characters is reserved for special purposes. For example, the percent sign
(%) is used as an escape character forcing the next two characters to be interpreted
as hex digits. In this way a space can be represented in a URI by the substring %20
BASIC WWW TECHNOLOGIES 35

and a tilde by %7E. Similarly, the question mark ? is reserved in URIs that represent
queries to delimit the identier of a queryable object and the query itself.
An absolute URI consists of two portions separated by a colon: a scheme specier
followed by a second portion whose syntax and semantics depend on the particu-
lar scheme. In BNF (BackusNaur Form), the general syntax of absolute URIs is
expressed as

absoluteURI ::= scheme : ( hierarchicalPart | opaquePart )

As an example, three common schemes for identifying Web resources are

http, https and ftp. Incidentally, the names of these schemes correspond
to the TCP/IP application level protocols (see Section 2.3.1) that are used to
access the resource. In these three cases, the second part of the URI is hier-
archical and consists of a double slash // followed by a so-called author-
ity component (e.g. a host name) and optionally by a path and/or a query.
For example, the URI http://bsd.slashdot.org/article.pl?sid=
02/08/16/0041250 contains all three components: the authority is bsd.
slashdot.org, the absolute path is /article.pl, and the query is the sub-
string following the question mark. More generally, the authority can be omitted and
we have

hierarchicalPart ::= ( networkPath | absolutePath ) [ ? query ]

networkPath ::= // authority [ absolutePath ]
absolutePath ::= / pathSegments

The authority can be either a server with an IP, or a reference to a registry of a naming
authority:

authority ::= server | registryName

For Internet URIs we are interested in the rst case. While in the simplest case a
server is specied by its full hostname, in general we must allow the possibility of
password-protected sites and servers that do not listen to the default TCP/IP port for
a given protocol. Thus

server ::= [ [ userinfo @ ] hostport ]

hostport ::= host [ : port ]
host ::= hostname | IPv4address

For example, the string more:[email protected]:8081 species that the

resource can be accessed on the nonstandard port 8081 of the host www.gnacz.org
and that the access requires logging in as user more with password less. The host
can also be specied using the numerical IP address (see Section 2.3.1):

IPv4address ::= 1*digit . 1*digit . 1*digit . 1*digit

36 PROTOCOLS

By website we generally mean the collection of resources that share the same
authority in their URI. Anchors in actual HTML pages may also contain relative URIs.
For example, the absolute URI http://www.gnacz.org/foo/bar.html
may be simply referenced to as /foo/bar.html in any source anchor contained
in a page belonging to the site www.gnacz.org.
URI-reference ::= [ absoluteURI | relativeURI ] [ # fragment ]
relativeURI ::= ( networkPath | absolutePath
| relativePath ) [ ? query ]

The full grammar specication for URIs can be found in Berners-Lee et al. (1998).

2.3 Protocols
Protocols describe how messages are encoded and specify how messages should be
exchanged. Protocols that regulate the functioning of computer networks are layered
in a hierarchical fashion. The lower layer in the hierarchy is typically related to the
physical communication mechanisms (e.g. it may be concerned with optical bers
or wireless communication). Higher levels are related to the functioning of specic
applications such as email or le transfer. A hierarchical organization of protocols
allows one to employ different levels of abstraction when describing and implement-
ing the software and the hardware components of a networked environment. For
example, a host in the Internet is characterized by a 32 bit address, independently of
whether the physical connection goes through a home DSL cable, through an ofce
Ethernet LAN, or through an airport wireless network. Similarly, email protocols are
a lingua franca spoken by clients and servers. Sending email just involves connect-
ing to a server and using this language. All of the details about how the information
is actually transmitted to the recipient are hidden by the hierarchical mechanism of
encapsulation by which higher level messages are embedded into a format understood
by the lower level protocols.

2.3.1 Reference models and TCP/IP

A rather famous model of protocol hierarchy was proposed in the early 1980s by the
International Organization for Standardization (ISO) and was called the Open System
Interconnect (OSI), shortly known as the ISO/OSI reference model (Day 1995; Day
and Zimmerman 1983). It consists of the following seven layers.
Physical: concerned with the transmission of electrical or optical (possibly also
acoustic) signals.
Data link: provides error control and divides data into frames.
Network: provides routing of packets from source to destination.
BASIC WWW TECHNOLOGIES 37

HTTP FTP IMAP4 Telnet SMTP POP3 TFTP

Application layer
(Process )
DNS

LDAP

TCP UDP Transport layer

(Host-to-host )

Network layer
IP (Internet )

Ethernet ATM Modem Data link/physical layer

(Network access )

Figure 2.3 Layers in the TCP/IP reference model.

Transport: provides end-to-end reliability (guaranteeing all packets reach destina-

tion).
Session: primitive functions that coordinate the dialoging of applications.
Presentation: concerned about the transfer syntax used by applications.
Application: applications such as le transfer, email, browsing, etc.
In a sense, the value of the ISO model is more theoretical than practical, since different
protocols are actually implemented in the Internet. The two most important ones are
the Transmission Control Protocol (TCP) and the Internet Protocol (IP). While both
TCP and IP are acronyms for two specic protocols, the term TCP/IP commonly
refers to a large suite of different protocols that have been implemented since the late
1960s, forming one of the core technologies underpinning the Internet.

2.3.2 The domain name system

Each physical device actively connected to the Internet must have a unique address
that, according to the IPv4 protocol, is a 32 bit address commonly represented as a
string of four decimal numbers separated by dots, as in 128.200.84.11 (see Deering
and Hinden (1998) for a specication of the more recent IPv6). IP addresses can have
one or more domain names attached. For example, 128.200.222.100 corresponds to
the domain name ftp.uci.edu. Names are obviously more understandable to
humans than numbers. However, whenever a user agent needs to open a connection
38 PROTOCOLS

to an Internet host for which the name is known (e.g. ftp.uci.edu), it rst must
determine the IP number. The mapping from names to numbers, or resolution, is
realized by the domain name service (DNS).
Names are organized hierarchically. The rightmost component of a name corre-
sponds to the highest level in the hierarchy and is known as the top-level domain
(TLD). Some historical TLDs include .com, .edu, .org, and .gov. Other TLDs
are, for example, associated with ISO 3166 country codes, such as .de, .fr, .it,
or .uk. The management of the domain name system and the root server system
(that handles resolution for TLDs), as well as the allocation of the IP address space, is
currently the responsibility of the Internet Corporation for Assigned Names and Num-
bers (ICANN). Within each TLD, several subdomains are registered (e.g. under the
management of a local naming authority) and the process can be repeated recursively.
Resolution is carried out by consulting a hierarchical distributed database. We
illustrate the mechanism by using an example. Suppose the user agent running
on lnx32.abcd.net needs to translate the name ftp.uni-gnat.edu. The
agent may rst query a local database that holds the most commonly used host-
names (this is normally stored in the le /etc/hosts in Unix-like systems). If
the search is unsuccessful, or no local database is maintained, a recursive query
is typically sent to a local DNS server, say dns.abcd.net. This server may be
able to provide an answer by consulting its cache. This happens if the same query
has been answered before and the result has been stored in the servers cache.
If not, then dns.abcd.net will need to consult other DNS servers, becoming
itself a client. For example, it might be congured to send a non-recursive query to
the DNS server of its Internet service provider, say dns.abcdprov.com. Hav-
ing received a non-recursive query, if dns.abcdprov.com cannot provide an
answer, it does not forward the query to another server but rather it returns to
dns.abcd.net a referral to another DNS server that might be able to answer.
Suppose it returns a referral to a root server. dns.abcd.net will now repeat
the query to the root server, which replies with a referral to the DNS server of the
.edu domain. The iteration continues and dns.abcd.net sends the query to the
server for the .edu, which suggests that dns.uni-gnat.edu is the authority for
the zone to which ftp.uni-gnat.edu belongs. Eventually, dns.abcd.net
receives the answer from dns.uni-gnat.edu and returns it to the original client
lnx32.abcd.net.

2.3.3 The Hypertext Transfer Protocol

The World Wide Web is largely based on the Hypertext Transfer Protocol (HTTP),
specifying the format of messages exchanged by a client, called in this case the user
agent (it could be a user browser or a crawler), and a server, i.e. the application
program that accepts connections and provides the requested services. Communi-
cation is established via a TCP connection to the default port (number 80) of the
server machine, although the URI can specify a different port number. In the simplest
BASIC WWW TECHNOLOGIES 39

Table 2.1 Methods in HTTP 1.1.

Method
name Description

GET Retrieve an entity identied by a request URI

(e.g. fetch a Web page or a le from the server)
HEAD Identical to GET except that the server must not return a
message body in response (i.e. returns only the header)
POST Append the enclosed entity. The supplied URI will handle
the entity (e.g. used to post a message to a newsgroup or to
append a record to a database)
PUT Store an enclosed entity under the supplied URI
(e.g. store a Web page or a le with the server
in order to create or modify a site)
DELETE Delete the resource identied by the request URI

TRACE Server echoes the request for debugging purposes

OPTIONS Server returns a list of supported methods

(e.g. to query the servers capabilities)
CONNECT Reserved for dynamic proxies

case the user agent directly connects to the server but, more generally, intermediary
agents like proxies, tunnels, or gateways can actually be present.
Messages are exchanged in ASCII format and pertain to one of several possible
methods, as detailed in Table 2.1. HTTP is essentially a request/response protocol.
A method essentially corresponds to a request from the user agent to the server. The
server responds to the request with a response message.
The most common method is GET, which allows the user agent to fetch an HTML
Web page, or another document such as an image or a le, from the server. Its usage
is illustrated in Figure 2.4. In this example, the response of the server is simply
the HTML code of the Web page associated with the requested URI (http://
www.ics.uci.edu/). In HTTP 1.0 the communication would be released upon
completion of the request, since HTTP 1.1 (Fielding et al. 1999; Krishnamurthy
et al. 1999) connections are persistent, i.e. the link remains active after a request.
Persistence offers considerable savings in overhead when several requests are sent to
the server, for instance while downloading a page with several images.
While some entities are simply documents stored in the servers le system, it is
also possible to have requests for entities that are actively generated by the server. This
is normally the case for pages served as a result of user queries: the servers software
could query a database management system to retrieve the requested information,
create on-the-y HTML code for display, and return the resulting ASCII stream as
an HTTP response. The collection of documents behind query forms is generally
referred to as the hidden (or invisible) Web (Bergman 2000).
40 PROTOCOLS
telnet www.ics.uci.edu 80
Trying 128.195.1.77...
User agent's
request Connected to lolth.ics.uci.edu.
Escape character is ^].
Server's GET http://www.ics.uci.edu/ HTTP/1.1
response Host: www.ics.uci.edu

HTTP/1.1 200 OK
Date: Wed, 25 Sep 2002 19:43:12 GMT
Server: Apache/1.3.26 (Unix) PHP/4.1.2 mod_ssl/2.8.10 OpenSSL/0.9.6e
X-Powered-By: PHP/4.1.2
HTML code of Transfer-Encoding: chunked
returned Content-Type: text/html
webpage
f00
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<html>
<head>
<title>Information and Computer Science at the University of
California, Irvine</title>
...

Figure 2.4 Example of the use of the GET method in an HTTP 1.1 session.

2.3.4 Programming examples

The example shown in Figure 2.4 can be reproduced on almost any platform using a
terminal or a telnet application and is simply intended to demonstrate the mechanics of
the protocol. Figure 2.5, however, shows a simple programming example that requires
some basic knowledge of the Perl programming language (Wall et al. 1996). The
program fetches the header returned by the HTTP server and can be easily modied
to collect interesting pieces of information, such as the server software being run
(third line of output) or the timestamp (fourth line of output) that indicates the last
modication time of the page (in Chapter 6 we discuss how this information can
be useful to estimate the age distribution of Web pages). Lines 18 in Figure 2.5
are simply directives and variable declarations. A communication channel (socket)
is established in lines 1012 by invoking the constructor of the INET class with
three arguments: the host (either specied by an IP address or a domain name), the
connection port (80) and a connection timeout in seconds. If a connection cannot be
established, the program is terminated in line 13. Line 15 prints a string to the socket
(that behaves like a character output device) issuing the HEAD request to the server,
using HTTP 1.0 (the commented line below shows how to use HTTP 1.1). Finally,
the loop in lines 1820 is used to read and print (to the standard output), one line at
a time, the response returned by the server.
A similar program can be used to sample the IP address space. This technique,
explained in more detail in Chapter 3, can be used to estimate global properties of the
Web (e.g. the number of active sites). For example we may be interested in determining
the fraction of the addressable IPv4 space (232 = 4 294 967 296 hosts in total) with
running Web servers. Figure 2.6 shows a basic Perl script for sampling IPv4 addresses,
testing the presence of an HTTP server running on port 80. The program counts a
success if a socket can be established and, in addition, the server returns the status
BASIC WWW TECHNOLOGIES 41
1 #!/usr/bin/perl
2 # getHeader.pl
3 use strict;
4 use IO::Socket qw(:DEFAULT :crlf);
5 my $host = shift || "www.unifi.it";
6 my $path = shift || "/";
7 my $timeout= shift || 2;
8 my $socket;
9 # Create a socket for standard HTTP connection (port 80)
10 my $socket = IO::Socket::INET->new(PeerAddr => $host,
11 PeerPort => http(80),
12 Timeout => $timeout)
13 or die "Cannot connect\n";
14 # Sends a HEAD request by printing on the socket
15 print $socket "HEAD $path HTTP/1.0",CRLF,CRLF;
16 #print $socket "HEAD $path HTTP/1.1",CRLF,"host: $host",CRLF,CRLF;
17 my $head;
18 while ($head = <$socket>) {
19 print $head;
20 }

% ./getHeader.pl
HTTP/1.1 200 OK
Date: Tue, 03 Dec 2002 23:59:59 GMT
Server: Apache/1.3.27 (Unix) PHP/4.2.3
Last-Modified: Mon, 04 Nov 2002 08:16:25 GMT
ETag: "f209-f91-3dc62cd9"
Accept-Ranges: bytes
Content-Length: 3985
Connection: close
Content-Type: text/html

Figure 2.5 Sample Perl script demonstrating the usage of the HEAD method in HTTP 1.0
and 1.1. A sample terminal output obtained by running the script is shown at the bottom.

code 200, which indicates a successful connection (Fielding et al. 1999). Although
useful as a starting point, this basic program has several limitations (see Exercise 2.4).

2.4 Log Files

When users access a website, all transactions between the users browser and the
Web server software are logged in ASCII format in server log les. The main log
le is usually known as a server transfer log or access log and typically contains
elds such as the IP address from which a users request originated, the time of the
request, the method of the request (see Table 2.1), a numerical code called the status
code, indicating the response from the server, and the size in bytes of the transaction.
There are several other logs in addition to the server transfer log. These include the
referrer log, which contains information about the Web page from which the request
originated, indicating if the user came to the site via a hyperlink for example. The
agent log contains information about the browser software making the request and can
also indicate whether requests are originating from a robot or spider program if the
requesting software identies itself as such. The error log records error information
for any requests that result in errors, such as a request for a Web page that cannot
42 LOG FILES
1 #!/usr/bin/perl
2 # simpleSampleIP.pl
3 use strict;
4 use IO::Socket qw(:DEFAULT :crlf);
5 my $trials = shift || 42949; # number of hosts to probe (default 42949)
6 my $timeout = shift || 1; # connection timeout (default 1 sec)
7 my $found=0; # number of valid IP found
8 my $host; # generic IP address to be probed
9 my $socket;
10 for (my $i=1; $i<$trials; $i++) {
11 $host=sprintf("%d.%d.%d.%d",
12 int(rand(255)), int(rand(255)),
13 int(rand(255)), int(rand(255)));
14 print "trial $i: $host ... ";
15 select((select(STDOUT), $| = 1)[0]); # flush STDOUT
16 # Create a socket for standard HTTP connection (port 80)
17 my $failed=0;
18 my $socket = IO::Socket::INET->new(PeerAddr => $host,
19 PeerPort => http(80),
20 Timeout => $timeout)
21 or $failed=1;
22 if (!$failed) {
23 print $socket "HEAD / HTTP/1.0",CRLF,CRLF;
24 my $header = <$socket>; # read the header
25 my ($p, $r, $m) = split(/ /,$header); # extract status code
27 print "$p $r $m";
28 if ($r==200) { # accept only OK status
29 $found++;
30 }
31 }
32 print " ($found) \n";
33 }
34 print "$found out of $trials IP accept connections on port 80\n";

Figure 2.6 Basic Perl script for sampling the IPv4 address space.

Table 2.2 An example of an entry in the www.ics.uci.edu Web server log le.

Field name Value

Requested URL /frost/ics171/

Remote IP 68.5.116.149
Remote login name
Remote user
Time [16/Nov/2002:23:51:05 -0800]
Method GET /frost/ics171/HTTP 1.1
Status code 200
Bytes sent 9274
Referrer
User agent Mozilla/5.0 . . . Windows NT 5.0 . . .

be found (code 404) or a server unavailable response (code 503) (see Fielding et
al. (1999) for details). Server-side logging software can be congured so that these
various log les are combined into a single server log le that records all of these
elds for each request we will assume this to be the case in what follows.
BASIC WWW TECHNOLOGIES 43

Table 2.3 A second example of an entry in the www.ics.uci.edu Web server log le.

Field Name Value

Requested URL /pub/ietf/http/hypermail/1996q1/0264.html

Remote IP 65.65.193.164
Remote login name
Remote user
Time [16/Nov/2002:23:51:23 -0800]
Method GET /pub/ietf/http/hypermail/
1996q1/0264.html HTTP 1.0
Status code 302
Bytes sent 315
Referrer http://www.google.com/
search?q=%22maximum+length+of+a+url
%22&hl=en&lr=&ie=UTF-8&oe=UTF-8
&start=10&sa=N
User agent Mozilla/5.0 . . . Windows NT 5.0 . . .

Table 2.2 shows an example of a single transaction from the www.ics.uci.edu

server log le in November 2002. The rst eld is the requested URL, followed by the
Remote IP address that generated the page request. The login and user elds are blank
(indicated by ), since this is a publicly visible Web page that does not require any
specic login procedure. The time eld records the local time at the level of seconds.
Being able to tell when a request arrived at this granularity is quite valuable when
reconstructing a sequence of page requests for an individual IP address from a large
log we will discuss how such sessions are created in more detail in Chapter 7. The
method eld corresponds to one of the HTTP methods listed in Table 2.1 here the
method is GET, a request for a specic page. The response of the server is recorded
in the status code eld, which in this case is 200, indicating success. The next eld
records the total number of bytes sent from the server to the requester as a result of
this transaction, if the request is successful. The referrer eld contains information
about where the user came from before they requested the current page. Frequently,
as in this case, this referrer information is unavailable. This can happen, for example,
when a user types a URL directly into the browser rather than following a hyperlink.
However, when the referrer information is available, it can be very useful from a data
analysis point of view, since it indicates how users are arriving at the site. The nal
eld in the record is user agent, which in this case is a long string of information
about the type of browser software being used to generate the request only part of
this string is displayed here.
Table 2.3 shows another example of a record from the log le of www.ics.
uci.edu, with a timestamp that is 18 s after the previous example in fact, there
were several other requests in the intervening 18 s. This particular record is inter-
esting because, unlike the other example, it does contain information in the refer-
rer eld. In particular, the referrer eld tells us that the user that generated this
44 SEARCH ENGINES

request came via the Google search engine (www.google.com) by issuing the
query maximum+length+of+a+url to Google and then following a hyperlink
in the displayed results to the www.ics.uci.edu website. The particular page
being requested here is a page from the mailing list archives of the Internet Engineer-
ing Task Force (IETF) HTTP Working Group and is being provided via ftp.

2.5 Search Engines

2.5.1 Overview
We now introduce the basic design for a Web search engine. A user that browses the
Web starting manually from a set of entry points, such as personal bookmarks or hierar-
chical structures such as the Open Directory Project (ODP, http://dmoz.org/),
is unlikely to nd answers to a general query in a reasonable amount of time. In
particular, although links reported in hierarchies such as the ODP tend to be accurate
thanks to human editing, only a very limited coverage of the Web contents is possible.
According to a recent Pew Internet Project Report (2002), more than 25% of Internet
users locate information on the Web with the help of search engines, and more than
80% of Internet users have used a search engine at least once. Search engines allow
users to enter a query sentence (either using natural language or Boolean combina-
tions of keywords) and receive as output an ordered list of URLs, ranked according
to their relevance to the query.
To achieve these goals, modern search engines are complex aggregates of several
pieces of software. In Figure 2.7 we show a very simplied architecture that nonethe-
less claries the role of some of the major components commonly found in real world
implementations. Crawlers, also known as spiders, or Web robots, are computer pro-
grams that autonomously navigate the Web and download documents. Copies of the
documents fetched by the crawler are then stored into an internal repository and sub-
sequently indexed (see Section 4.1) to allow efcient searching. The link repository
is essentially a representation of the Web as a graph. This component has become
extremely important for improving the quality of ranking (to be further discussed
in Chapter 5). The search engine processes a users query, returning a set of match-
ing documents that are eventually ranked according to criteria programmed into the
scoring module.
The design of a search engines may vary signicantly, depending on many factors.
The most important distinction is probably between general-purpose engines (that are
meant to provide general answers to the largest possible set of users) and thematic or
specialized engines that focus on specic topics or types of documents.
Ideally, general-purpose engines should allow users to search the largest possi-
ble portion of the Web. However, various technological limitations impose trade-offs
among conicting objectives. Two of the most important measures of the performance
for a search engine are coverage and recency. Coverage is the fraction of the visible
Web that is searchable by the engine. Recency (or freshness) is the fraction of search-
BASIC WWW TECHNOLOGIES 45

Results

Crawlers Scoring results

Link repository Query

Query engine

End user

Crawling control
Document repository

Indexer

Inverted index

Figure 2.7 A high-level sketch of the architecture of a search engine.

able pages (or URLs) that are up-to-date in the internal repository data pages stored
by the engine. One approach for maintaining a fresh repository consists of crawl-
ing the Web periodically in order to replace the contents of the index with updated
versions of documents and possibly to add newly created documents. It is important
to stress that fetching a large portion of the Web can take several days because of
bandwidth limitations and that during the elapsed time between consecutive crawls
the index is not up to date. As a result, achieving high performance in terms of both
coverage and recency is a major challenge for general-purpose search engines, as
earlier noted by Lawrence and Giles (1998b). This issue is explained in more detail
in the following.

2.5.2 Coverage

Although several companies claimed by the mid 1990s that their search engines were
able to offer an almost complete coverage of Web content, it soon became clear that
only a relatively small fraction of existing Web pages could be fetched by a single
search engine. Lawrence and Giles (1998b, 1999) describe two experiments aimed
at measuring the performance of search engines in terms of coverage and freshness.
They use the following approach, known as overlap analysis, for estimating the size
of the indexable Web (see also Bharat and Broder 1998). Let W denote the set of Web
pages and let Wa W , Wb W be the pages crawled by two independent engines, a
and b. How big are Wa and Wb with respect to W ? Suppose we can uniformly sample
documents and test for membership in the two sets. Denote by P (Wa ) and P (Wb ) the
46 SEARCH ENGINES

probabilities that a page has been crawled by a or b, respectively. We know that

|Wa Wb |
P (Wa Wb | Wb ) = . (2.1)
|Wb |
Now if the two crawls are assumed to be independent, the left-hand side of Equa-
tion (2.1) reduces to P (Wa ), the coverage of engine a. So coverage can be easily
obtained from the size of the intersection of the two crawls. This quantity, how-
ever, cannot be computed exactly unless one has access to the data stored in their
repositories. In their rst study, Lawrence and Giles (1998b) used a controlled set of
575 queries in order to sample pages and count the number of times both engines
returned the same page. Assuming the resulting estimate P (Wa ) is correct, it is
straightforward to infer the size of the entire Web as |Wa |/P (Wa ). Using this approach,
it was estimated that the Web contained at least 320 million pages in November 1997,
that about 60% of the Web was covered by the six major engines at that time, and
that the maximum coverage from a single engine was only one third of the Web.
Actually, since this approach is based on observations, it yields an estimate of the size
of the so-called visible Web, excluding (for example) pages that are behind forms,
password protected, etc.
The experiment was repeated in February 1999, under slightly different conditions
and using 11 search engines (Lawrence and Giles 1999). This later study found an
estimated 42% collective coverage and a maximum 16% coverage by a single engine,
while the estimated size of the Web grew to roughly 800 million pages. At the time of
writing (November 2002), some of the major companies claim they can return results
for over three billion pages, while it is widely believed that the Web now contains
several billion pages. Fienberg et al. (1999) propose an interesting Bayesian approach
to the same problem in a later paper.
A strategy for improving coverage is the construction of meta-search engines that
dispatch a user query to several conventional engines simultaneously, collect the
result and merge them using some fusion policy before returning a list to the user.
SavvySearch (Dreilinger and Howe 1997) selects the search engine to which queries
should be dispatched. Inquirus (Lawrence and Giles 1998a) improves the general
meta-search approach by retrieving the pages returned by the search engines and
using a fusion policy that depends on the page contents. A Bayesian approach to the
meta-search problem is described in Aslam and Montague (2001).

2.5.3 Basic crawling

The algorithm Simple-Crawler listed below is a simplied version of a basic
general-purpose crawling process. The algorithm essentially performs a graph visit
(Cormen et al. 2001). It starts from a collection S0 of URLs (which are sometimes
referred to as the seeds) and stores the documents and links encountered into two
repositories, referenced by D and E, respectively. It maintains a list, Q, internally
that contains the fringe of the portion of the graph visited so far. The function Dequeue
BASIC WWW TECHNOLOGIES 47

extracts and returns one element out of the list. If Q is managed according to a rst-
in rst-out (FIFO) policy, then the crawler performs a breadth-rst search (BFS).
Assuming that the purpose is to download all pages reachable from S0 , this crawling
strategy may be reasonable, at least as a simple strategy to start with. However, we
will see later that the URLs in Q may be sorted according to different criteria in order
to focus the crawl in particular directions so as to improve the quality of the collected
documents with respect to an assigned goal. In line 4, we invoke the function Fetch,
which downloads the document addressed by the URL u. The fetched document is
then stored in the repository D. In line 6, the function Parse extracts all of the URLs
contained in the document (the set of children of u) and puts them in a temporary list
L. All the elements of L are in turn inserted into the crawling queue Q, unless they
have already been crawled or they are already enqueued for crawling. The latter test
is necessary to avoid loops, as the Web is not acyclic. Compared to standard graph
visit algorithms, storing reached and fetched URLs replaces traditional node coloring.
Note that in line 8 we also store the edges of the Web graph. We will see later how
recent scoring algorithms can take advantage of this information.
Simple-Crawler(S0 , D, E)
1 Q S0
2 while Q =
3 do u Dequeue(Q)
4 d(u) Fetch(u)
5 Store(D, (d(u), u))
6 L Parse(d(u))
7 for each v in L
8 do Store(E, (u, v))
9 if (v D v Q)
10 then Enqueue(Q, v).

While the above scheme is conceptually acceptable, in a real world application

several practical modications are also necessary.
Under the assumption of a static Web, the contents of Q and D completely
characterize the state of the crawler, in the sense that the process could be
restarted with no information loss if Q and D are preserved. Because of their
size and because of the need of persistent memory (to avoid data loss in the case
of a crash), these objects must be stored on disk in practical implementations.
In practice, it will be always necessary to halt Simple-Crawler before the loop
in lines 210 terminates with an empty Q, simply because storage resources are
exhausted or because the time slot assigned to the crawling process has expired.
In this case we are left with both fetched and discovered URLs (stored in D
and Q, respectively). The contents of pages whose URLs were just discovered
is unknown. Nonetheless, it is still possible to associate the text in the corre-
sponding anchors with these URLs and return them as query results. Indeed,
48 SEARCH ENGINES

several search engines can (and do) return links to documents they have never
downloaded.
The time required to complete the downloading of a document is unknown
because of many factors, including connection delays and network congestions.
Moreover, the bandwidth available at the crawling site(s) may be signicantly
larger than at the location where the document to be downloaded is located.
As a result, a single crawling process that just waits for the completion of
each download before moving to the next one would be a poor design choice.
Running concurrent fetching threads is thus the normal solution to maximize the
exploitation of available bandwidth. Note that a multithreaded implementation
with a single queue Q is still a single crawling process. Parallelization of the
crawling process is discussed in Section 6.3.
Crawlers should be respectful of servers and should not abuse the resources at
the target site otherwise administrators may decide to block access (Koster
1995). The Robots Exclusion Protocol (http://www.robotstxt.org/)
allows webmasters to dene which portions of a site are permissible for robots
to fetch from. Under this protocol, certain directories or dynamically generated
pages can be declared as off-limits for a crawler if they are listed in the le
robots.txt located in the root directory. A well-behaved crawler should
always avoid fetching excluded documents. Similarly, by inserting special a
META tag in the HTML code, a webmaster may indicate that the contents of
certain pages should not be indexed.
Overloading sites or network links should also be avoided. For example, multi-
ple threads should not fetch from the same server simultaneously or too often.
Unfortunately, outlinks often point within the same site (this happens in par-
ticular for dot com sites), leading to a relatively high degree of locality in
a single queue. For example, Mercator (Heydon and Najork 2001) is a Web
crawler that implements a sophisticated strategy for broadening as much as
possible the crawling fringe and increasing the elapsed time between two con-
secutive requests to the same server. It maintains a data structure formed by
several FIFO queues, each containing only URLs pointing to a unique server,
and by an index that associates a timestamp to each queue. Fetching threads
must check in the index if enough time has passed since the last access to the
same queue, and once they download a document they insert a timestamp in
the corresponding queue.
URLs that have already been fetched or discovered must not be inserted again
into the crawling queue Q. The test in line 9, however, is not straightforward
to implement. In particular, since Q and D are stored on disk, special care is
needed in order to avoid any overheads due to external memory management.
Heydon and Najork (2001) propose to use two caches. The rst cache contains
recently added URLs, exploiting locality during the crawl. The second cache
contains popular URLs, as determined during the crawl. Heydon and Najork
BASIC WWW TECHNOLOGIES 49

(2001) report high hit rate for the second cache and this can be explained by the
scale-free distribution for the number of incoming links of a given Web page
(see Chapter 3 for details on these graphical properties of the Web).
The real world is more complex than a mathematical model of the Web and in
practice several precautions must be enforced in a real crawler to avoid problems
like aliases and traps (Heydon and Najork 1999). Aliases occur if the same
document can be addressed by many distinct URLs, for example, if a site is
registered under multiple domain names. It is possible to cope with aliases by
using canonicalization, i.e. issuing a DNS request in order to get the canonical
name of the host (see Section 2.3.2). A related problem of multiple URLs
pointing to the same document occurs when the server embeds session IDs
into the URL. Traps can be generated by malicious common gateway interface
(CGI) scripts that keep generating on the y fake documents pointing to other
fake documents.
DNS lookup is necessary to map hostnames in URLs to IP addresses before
downloading, or to canonicalize URLs before inserting them in the discov-
ered queue. A DNS lookup can often take considerable time due to slow DNS
server responses. Efcient crawlers may need to rewrite DNS resolution in a
multithreaded fashion (Heydon and Najork 1999).

We should also bear in mind another important difference between graph visiting and
crawling. A graph is normally thought of as a static object during the execution of a
crawling algorithm. However, in reality, the Web is dynamic and changes continuously
in terms of both content and topology. Fetching a billion documents can take weeks.
Crawling is thus like taking a picture of a living scene with some objects remaining
xed and others moving at a very high speed. At the end of the process, D and E will
unavoidably contain stale information. The problem can be somewhat addressed by
more frequently downloading sites that are expected to change more often (e.g. portals,
press, etc.) and less frequent downloading of more static sites.Accounting for temporal
effects during a crawl is a relatively recent research topic in Section 6.4 we will
discuss specic crawling approaches that address Web dynamics.

2.6 Exercises
Exercise 2.1. Explain the differences amongst URI, URL, and URN using Venn
diagrams and provide examples for each intersection amongst these sets.
Exercise 2.2. Write a program that extracts all the URIs from a given HTML docu-
ment.
Exercise 2.3. Write a program to automatically query a given search engine. The
program should accept a query string as input and build up a list of URIs extracted
from the HTML pages returned by the search engine.
50 EXERCISES

Exercise 2.4. Improve the program of Figure 2.6 and perform some basic sampling
experiments. In particular:
write a multithreaded version so that there will be no problems waiting for a
longer timeout in the case of connection failure; and
correct the random generation of IP addresses to exclude special values that
are associated for example with default routers, or unassigned portions of the
address space (see ONeill et al. (1997) for details).
After completing your program collect some statistics from the Web, for example, the
distribution of usage of different HTTP server software.
3

Web Graphs

3.1 Internet and Web Graphs

The Internet can be viewed as a graph in many different ways. For instance, in one
representation, one could focus on the physical layer, and use a graph where nodes
correspond to computers, routers, and so forth, at xed locations, and edges corre-
spond to direct physical connections. This graph, however, is somewhat limited. It
misses the fast growing wireless boundary of the Internet and, more fundamentally,
it does not really capture the graphical connections associated with the information
available through the Internet. Thus, in the rest of this chapter, we focus on graphical
representations where the nodes represent Internet pages and the edges are associated
with hyperlinks that connect the information contained in different pages, the Web
graph. Remarkably, as we shall see later in the chapter, the Web graph shares several
interesting properties with the Internet hardware graph, as well as with many other
networks.
When focusing on the graph of hyperlinks, there are a number of additional factors
to be considered. First, the edges can be considered as directed, from the source page
to the target page, or undirected. For many of the results discussed in this chapter,
this distinction turns out not to be essential. Second, the graph itself is dynamic
and undergoing constant changes: nodes and connections are being added almost
every minute and, less frequently, old nodes or connections are being deleted. The
content of existing nodes, i.e. Web pages, is also subject to change, although these
changes may not alter the underlying graphical structure, unless new hyperlinks are
created. One consequence of these dynamical aspects is that it is virtually impossible
to know exactly what all the nodes and connections of the graph are at time t. Thus,
as discussed in Chapter 2, the exact size of the graph is not really a useful or even
well-dened quantity. Third, although most of the time we will focus on the primary
connected component of the graph (at least in the undirected version), additional small
disconnected components exist on the fringes. Finally, there are pages and hyperlinks
that are created dynamically on the y, for instance after entering a query in a search
engine. These have typically a short life span and therefore will be excluded from the
analysis.
Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
52 INTERNET AND WEB GRAPHS

In short, we are dealing with a fuzzy graph that is constantly evolving on several
different time scales. For simplicity the main analysis focuses on a large and slowly
evolving subgraph of this graph. The dynamic aspect of course suggests other inter-
esting questions, related to the temporal evolution of relevant variables. For example,
how is the total number of nodes and links growing with time and what is the life
expectancy of a typical page? Some of these temporal issues will also be addressed
in this chapter and again later in Chapter 6.
There are many reasons, from mathematical, to social, to economic to study the
Web graph. It is an example of a large, dynamic, and distributed graph where we have
the privilege of being able to make a large number of measurements. In addition,
as we are about to see, the Web graph shares many properties with several other
complex graphs found in a variety of systems, ranging from social organizations to
biological systems. The graph may reect psychological and sociological properties
of how humans organize information and their societies. Perhaps on a more pragmatic
note, the structure of the graph may also provide insights regarding the robustness and
vulnerability of the Internet and help us develop more efcient search and navigation
algorithms. Indeed, one way to locate almost any kind of information today is to
navigate to the appropriate Web page by following a sequence of hyperlinks, i.e. a
path in the graph. Efcient navigation algorithms ought to be able to nd short paths,
at least most of the time. And nally, the behavior of users as they traverse the Web
graph is also of great interest and is the focus of Chapter 7.
Having set our focus on the graph of relatively stable Web pages and their hyper-
links, several different questions can be asked regarding the size and connectivity of
the graph, the number of connected components, the distribution of pages per site,
the distribution of incoming and outgoing connections per site, and the average and
maximal length of the shortest path between any two vertices (diameter).
Several empirical studies have been conducted to study the properties of the Web
graph. These are based on random sampling methods or, more often, on some form of
crawling applied to subgraphs of various sizes, ranging from a few thousand to a few
hundred million pages. To a rst-order approximation, these studies have revealed
consistent emerging properties of the Web graph, observable at different scales. One
particularly striking property is the fact that connectivity follows a power-law distri-
bution. Remarkably, these results have had a sizable impact in graph theory, with the
emergence of new problems and new classes of random graphs that are the focus of
active mathematical investigation (see Bollobs and Riordan (2003) and additional
references in the notes at the end of this chapter).
One preliminary but essential observation that holds for all the graphs to be con-
sidered in this chapter is that they are sparse, i.e. have a number of edges that is small
(|E| = O(n)) or at least o(n2 ) compared to the number O(n2 ) of edges in a complete
or dense graph. For the Web, this is intuitively obvious (at least in the foreseeable
future), because the average number of hyperlinks per page can be expected to be
roughly a constant, in part related to human information processing abilities. This
sparseness is already a departure from the uniform random graph model where the
probability of an edge p is a constant and the total number of edges |E| pn2 /2.
WEB GRAPHS 53

3.1.1 Power-law size

As discussed in Chapter 2, it is difcult to estimate the exact number of documents on
the Web at any given time and this number continues to grow at a fast pace. Simple
back-of-the-envelope estimates as well as more sophisticated ones (Lawrence and
Giles 1999) suggest a size of well over a billion nodes, and several billion edges. At
the time of this writing, major search engines such as Google claim an index size of
roughly three billion URLs, conrming that the actual order of magnitude of the Web
is in the vicinity of 1010 .
In Huberman and Adamic (1999), the distribution of site sizes measured by the
number of pages, as determined by crawls of common search engines including several
hundred thousand pages, is reported to follow a power-law distribution. This power-
law distribution is observed over several orders of magnitude with an exponent in
the range 1.61.9. This power law allows one, for instance, to estimate the expected
number of sites of any size, even if that size has not been observed in a crawling
experiment. This is achieved by using the scaling property of power laws seen in
Chapter 1 by writing P (m) = P (n) (m/n) .

3.1.2 Power-law connectivity

Several studies (Barabsi and Albert 1999; Broder et al. 2000; Kleinberg et al. 1999)
have consistently reported that the distribution of the number of connections per
node follows a power-law distribution, both in the directed and undirected cases.
These power-law results, derived from crawls of particular organizations or various
commercial crawls of the Web, are observed over several orders of magnitude and
at different scales. Thus, the general feel of the Web graph is the same for very
different subgraphs, such as those spanned by a university, a country, or the crawl of
a commercial search engine.
For instance, a study conducted by crawling the domain of the University of Notre
Dame, Indiana (Albert et al. 1999; Barabsi and Albert 1999) reported exponents
of = 2.45 and = 2.1 for the outdegree and indegree distributions, respectively
(Table 3.1). At the time of the study, the graph associated with the *.nd.edu domain
contained 325 729 documents and 1 469 680 links. A larger similar study is described
in Broder et al. (2000).
A power-law distribution for the connectivity is very different from the distribution
found in traditional uniform random graphs (Bollobs 1985; Erdos and Rnyi 1959,
1960), or in other related models discussed below (Watts and Strogatz 1998). In a
random uniform graph with random independent edges of xed probability p, with
p small, the number of edges with k neighbors has a Poisson distribution with mean
 
n1 k
=n p (1 p)n1k (3.1)
k
and decays exponentially fast to zero as k increases toward its maximum value n 1.
Thus the power-law distribution is a remarkable example of emerging order in a large
54 INTERNET AND WEB GRAPHS

graph created by many agents, each of whom is completely free to create documents
and hyperlinks.
The emergence of a power-law distribution is by itself intriguing enough to require
an explanation. But in addition, the need for an explanation is exacerbated by the fact
that a similar distribution has been observed for many other networks: business net-
works, social networks, transportation networks, telecommunication networks, bio-
logical networks (both molecular and neural) and so forth (Table 3.1). The following
list of examples is not exhaustive.
The Internet at the router and inter-domain level has a connectivity distribution
that falls off as a power-law distribution with 2.48 (Faloutsos et al. 1999).
Power-law connectivity has also been reported at the level of peer-to-peer net-
works (Ripeanu et al. 2002).
The call graphs associated with the calls handled by some subset of telephony
carriers over a certain time period (Abello et al. 1998).
The power grid of the western United States (Albert et al. 1999; Phadke and
Thorp 1988) where, for instance, nodes represent generators, transformers, and
substations and edges represent high-voltage transmission lines between them.
The citation network where nodes represent papers and links are associated
with citations (Redner 1998). A similar graph, smaller but famous in the math-
ematics community, is the graph of collaborators of Paul Erds (see http://
www.acs.oakland.edu/grossman/erdoshp.html).
The collaboration graph of actors (http://us.imdb.com), where nodes
correspond to actors and links to pairs of actors that have costarred in a lm
(Barabsi and Albert 1999).
The networks associated with metabolic pathways (Jeong et al. 2000) where
the probability that a given substrate participates in k reactions decays as k
in a representative variety of organisms. These reactions have directions and
both the indegrees and outdegrees follow a power-law distribution with similar
exponents in general. In Escherichia coli, for instance, the exponent is = 2.2
for both the indegrees and the outdegrees. In contrast with what happens for the
other networks, Jeong et al. (2000) report that in metabolic networks the diam-
eter, measured by the average distance between substrates, does not seem to
grow logarithmically with the number of molecular species, but rather appears
to be constant for all organisms and independent of the number of molecular
species. (They also observe that the ranking of the most connected substrates
is essentially the same across all organisms). A constant diameter may confer
increased exibility during evolution.
The networks formed by interacting genes and proteins as described in
Maslov and Sneppen (2002). These authors also report the existence of a
richer level of structure in the corresponding graphs. In particular, direct
WEB GRAPHS 55

Table 3.1 Power-law exponent, average degree, and average diameter (e.g. average minimal
distance between pairs of vertices). Data partly from Barabasi and Albert (1999) and Jeong et
al. (2000). Ex., entries left as an exercise; NA, not applicable.

Average Average
Exponent degree diameter

actors 2.3 28.78 Ex.

power grid 4 2.67 Ex.
citation 3 Ex. Ex.
protein interactions 2.5 Ex. Ex.
E. coli metabolic indegree 2.2 3.7 3.2
E. coli metabolic outdegree 2.2 3.7 3.2
Internet routers 2.48 Ex. Ex.
WWW (undirected) Ex. 57 19
WWW (indegree) 2.1 5.46 Ex.
WWW (outdegree) 2.45 Ex. Ex.
theoretical model (pref. attach.) 2.9 Ex. Ex.
WWW (pages per site) 1.8 NA NA
language NA NA

connections between proteins with high degrees are suppressed, whereas

direct connections between highly connected and low-connected pairs of pro-
teins are favored. This additional pattern of connectivity may reduce the
likelihood of cross talk between different functional modules and confer
additional robustness to the network by localizing the effect of perturba-
tions.

The graph of nervous connections in the model organism Caenorhabditis

elegans (Achacoso and Yamamoto 1992). C. elegans is a little transparent
worm that has a hardwired nervous system with a xed number of neu-
rons (vertices) and a genetically wired set of connections (edges) associated
with synapses or gap junctions between the corresponding neurons. How-
ever, because the entire network contains only 302 neuronal nodes (and about
5000 chemical synapses), the statistical signicance of this result is somewhat
weaker.
Reports of power-law distributions have on occasion been controversial, as in the
case of food webs (McCann et al. 1998; Paine 1992; Pimm et al. 1991), which are
another example of networks that are nonrandom but also nonregular. For biological
systems in particular, large-scale information about graphs of interacting elements is
not easy to derive and is only beginning to become available with high-throughput
technologies such as DNA and protein arrays. As more data become available, better
assessments of power-law connectivity, its origins, its uctuations, and how universal
it is should become possible.
56 INTERNET AND WEB GRAPHS

3.1.3 Small-world networks

A second empirical observation reported for the Web graph, which is also shared by
many of the networks above, is that the diameter of the Web graph is small relative to
the size of the overall system. This property is also often described in terms of small
world or six degrees of separation, originating in the study of sociological webs of
relationships by Milgram in the late 1960s and the idea that any two human beings on
the planet are typically connected by a short chain of acquaintance relationships (Korte
and Milgram 1978; Milgram 1967; Travers and Milgram 1969). In more mathematical
terms, small-world networks are characterized by a diameter that is exponentially
smaller than the size, i.e. where the diameter is bounded by a polynomial in log n.
Unlike the distribution of connectivity, an estimate of the diameter in terms of the
largest or average length of the shortest path between any two Web pages cannot
be derived by exhaustive analysis of n vertices. Finding the shortest path between a
source vertex and any other vertex in a graph takes O(n2 ) steps in general, which
can be reduced to O(n log n) in the case of a sparse graph like the Web, essentially
using dynamic programming (see Cormen et al. 2001 for more details). Even in a
sparse graph like the Web, this can be computationally demanding for large values of
n. Thus, diameter properties must be inferred from sampling, the average diameter
being easier to estimate than the maximal diameter.
By simulating a random graph with the same power-law connectivity as the one
found on the Web (Albert et al. 1999), it was found that the average distance between
two vertices is given by d = 0.35 + 2.06 log10 n. Thus the Web is clearly a small-
world network. For n = 109 , d = E[l] = 18.89, so that two randomly chosen doc-
uments are separated on average by 19 clicks (implicitly this implies that directed
paths are being considered, since current browsers, unlike search engines, can only
follow links in one direction). These authors also report that d has a roughly Gaussian
distribution. The logarithmic scaling of the diameter or the average distance between
documents on the Web is important to ensure that the future growth of information
on the Web remains manageable. Under these assumptions, a ten-fold increase in
overall Web size increases the average number of clicks along a minimal path by
only two. It should be noted, however, that even an increase of only two clicks on
average would probably be perceived as extremely annoying and better solutions are
needed indeed this is in part the role of search engines, which are studied in the
coming chapters.
As a more general but important side issue, the minimal distance between two
randomly selected nodes is interesting from a theoretical standpoint, but does not
characterize actual browsing for a variety of reasons. First, humans do not necessarily
take the shortest path available in the graph because, for instance:
(a) they may not know it, nor be able to effectively approximate it;
(b) they may get distracted by something on the way;
(c) they bookmark sites, so that future traversals may ignore the graph entirely.
WEB GRAPHS 57

Second, humans do not traverse links from one randomly selected site to another.
Exploratory browsing aside, they go to a given page in response to an informational
need, and generally start from a page that they believe will lead them to their target
quickly after all, this is what search engines are for.
One additional issue related to the power-law distribution, small network, and
sampling size, is the behavior of the degree distribution for very large values of k and
the difculty in assessing the behavior of the tail of the distribution with nite data.
Amaral et al. (2000) have analyzed several naturally occurring networks and reported
the existence of three different classes of small-world networks:
(a) scale-free networks characterized by a power-law distribution of connectivity
over the entire range of values;
(b) broad-scale networks, where the power-law distribution applies over a broad
range but is followed by a sharp cutoff; and
(c) single scale networks with a connectivity distribution that decays exponentially,
as in the case of standard random graph models.
The authors argue that in single and broad-scale networks there are in general addi-
tional constraints that limit the addition of new links, related for instance to the aging
of vertices, their limited capacity, or the cost of adding links. It is not clear, however,
that a given nite network can always be classied in a clean fashion into one of these
three categories. Furthermore, it is also not clear whether the very large-scale tail of
the degree distribution of the Web graph exhibits a cutoff or not, and if so whether it
is due to additional operating constraints.

3.1.4 Power law of PageRank

When searching information on the Web, it is important to be able to assess the
importance of a page relative to a query and to rank pages accordingly. A crude way
of measuring the importance of a page is to look at its indegree, i.e. the number of pages
that point to it. This index is not very reliable, however, because it is entirely local and
very easy to spam. PageRank may be a better and more global index, originally used in
the Google search engine (Brin and Page 1998). It will be described here informally
and studied in more detail in Chapter 5. Informally, the PageRank of a page can
be understood as the proportion of time a random surfer would spend on that page
at steady state. The idea is to imagine a random rst-order Markov surfer who, at
each time step, travels from one Web page to the next by randomly and uniformly
selecting an outgoing hyperlink on the current page. Thus, in this view, the Web graph
is the transition graph of a Markov chain and all outgoing edges from a given node
are assigned the same probability. Because this Markov chain could have isolated or
absorbing components, a general parameter  can be added so that, in addition, at
each time step there is a probability 1  of moving to any other page on the Web.
It is well known and easy to see that the steady-state distribution of such a Markov
58 INTERNET AND WEB GRAPHS

chain, representing the probability of being in any state, is obtained by looking at the
eigenvector and eigenvalue of the transition matrix. Specically, the page rank r(v)
of page v is the steady-state distribution obtained by solving the system of n linear
equations given by
1  r(v)
r(v) = + , (3.2)
n |ch[u]|
upa[v]

where pa[v] is the set of parent nodes, i.e. of pages that point to page v, and |ch[u]|
denotes the outdegree, i.e. the size of the set of children nodes.
Analysis of the distribution of PageRank values in Pandurangan et al. (2002) indi-
cates that PageRank also follows a power-law distribution with the same exponent
(namely 2.1) as the indegree distribution and, as in the case of degree distributions,
this distribution seems to be relatively insensitive to the particular snapshot of the
Web used for the measurements.

3.1.5 The bow-tie structure

A more careful and comprehensive analysis of the topology of the Web must take
into account the orientation of the edges and existence of disconnected components.
The small-world property applies to certain components and to the core of the Web
graph but is not true for the entire Web graph. There are signicant portions of the
Web that cannot be reached from other signicant portions, and there is a signicant
portion of pairs of nodes that can only be connected through fairly long paths.
A large-scale study using two Altavista crawls, each with over 200 million nodes
and 1.5 billion links (Broder et al. 2000), reported a macroscopic bow-tie structure
(Figure 3.1) in which over 90% of the nodes belong to a single connected component if
the hyperlinks are treated as undirected edges. Incidentally, the study also conrmed
the power-law distribution of connectivity both for page indegrees ( = 2.1) and
outdegrees ( = 2.72).
The bow-tie structure contains four different components. The rst component is
the core forming a giant strongly connected component (SCC), whose pages can
reach each other via directed paths. This component contains most of the prominent
sites. The second component (IN) consists of upstream pages that can reach the
core via a directed path of hyperlinks, but cannot be reached from the core in a
similar way. The third downstream component (OUT) consists of pages that can be
reached from the core via a directed path of hyperlinks, but cannot reach the core
in a similar way. Finally, the fourth component is more heterogeneous and contains
smaller disconnected components and tendrils consisting of pages that can neither
reach or be reached from the core. However, they can be reached via a directed
path originating from pages in IN, or pages that can project into OUT via a directed
path. Finally, using Broder et al.s terminology, tubes are directed paths involving
tendrils from a page in IN to a page in OUT. In the reported experiments, each of the
four components has roughly the same size and contains on the order of 50 million
WEB GRAPHS 59

TENDRILS

IN OUT
SCC

TUBES

DISCONNECTED COMPONENTS
Figure 3.1 Internet bow-tie mode, adapted from Broder et al. (2000), with four broad regions
of roughly equal size. A strongly connected component (SCC) within which each node can
reach any other node via a directed path. An IN component of nodes that can reach nodes in SCC
via a directed path but cannot be reached by nodes in SCC via a directed path. A similar OUT
component downstream of the SCC component. The fourth component is more heterogeneous
and contains (a) small disconnected components; and (b) tendrils associated with nodes that
are not in SCC and can be reached by a directed path originating in IN or terminating in OUT.
Tubes connect tendrils emanating from IN to tendrils projecting into OUT.

pages, although these proportions could vary in time and would need to be revisited
periodically.
Because this study contains disconnected components, it is clear that both the max-
imal and average diameter are innite. Thus, the author use the maximal minimal path
length and the average path length restricted to pairs of points that can be connected
to each other. With this caveat, they report in their study a maximal minimal diameter
of at least 28 for the central core, and over 500 for the entire graph. For any randomly
selected pair of points, the probability that a directed path exists from the rst to the
second is 0.24 and if a directed path exists, its average length is about 16, reasonably
close to the estimate of 19 obtained in the previous section. The average length of
an undirected path, when it exists, is close to 7. Thus in particular the core has a
small-world structure: the shortest directed path from any page in the core to any
other page in the core involves 1620 links on average.
The study in Broder et al. (2000) also reports a power-law distribution in the sizes
of the connected components with undirected edges, with an exponent = 2.5 over
roughly ve orders of magnitude. Bow-tie structure in subgraphs of the Webgraph
associated, for instance, with a particular topic are described in Dill et al. (2001).
60 GENERATIVE MODELS FOR THE WEB GRAPH

3.2 Generative Models for the Web Graph and

Other Networks
There have been several proposals for stochastic models that can explain, or at least
partly reproduce, the emergent properties observed for the distribution of site sizes and
the connectivity of Web-like graphs described in the previous sections. In particular,
there have been a number of such proposals for the power-law distributions and small-
world network properties. If the processes that drive the growth of the WWW are
highly decentralized, then the power-law distribution, for instance, must emerge from
an aggregation of local behaviors. But what is the right model for this aggregation?
In this section, we review some of the models that have been proposed and provide
additional pointers at the end of the chapter. In the case of random Web-like graph
models, these new models strongly differ from the classic random graph models
(Bollobs 1985; Erdos and Rnyi 1959, 1960), where the number n of vertices is
xed and each possible edge occurs independently of all other edges with a xed
probability p. In fact, it is fair to say that the study of the graphical properties of the
Internet and the Web have created new problems and new areas of active investigation
in the mathematical theory of random graphs.

3.2.1 Web page growth

Several empirical estimates of the size of the Web, the number of IP addresses, the
number of servers, the average size of a page, and so forth, have been published, for
instance in Lawrence and Giles (1999). Huberman and Adamic (1999) have observed
a power-law distribution of site sizes and proposed a generative model to account
for this distribution. The rst component of this model is that sites have short-term
(for instance, daily) size uctuations up or down that are proportional to the size of
the site. This is not an unreasonable assumption: a site with 100 000 pages may gain
or loose a few hundred pages in one day, whereas the same effect would be rather
surprising for a site with only 100 pages. A second ingredient is to assume that there
is an overall growth rate so that the size S(t) satises

S(t + 1) = (1 + t )S(t), (3.3)

where t is the realization of a 1 Bernouilli random variable at time t with proba-

bility 0.5, and is the absolute rate of the daily uctuations. Thus, after T steps,
1
T
S(T ) = T S(0) (1 + t ), (3.4)
i=0

so that
1
T
log S(T ) = T log + log S(0) + log(1 + t ). (3.5)
i=0
WEB GRAPHS 61

The last term can also be written as

l log(1 + ) + (T l) log(1 ),

where l is the number of positive uctuations. Assuming that the variables t are
independent, by the central limit theorem it is clear that for large values of T the
variable log S(T ) is normally distributed. Alternatively, log S(T ) can be associated
with a binomial distribution counting the number of times t is equal to +1 and the
binomial distribution can be approximated by a Gaussian. In any case, it follows
that S(T ) has a log-normal distribution which, as seen in Chapter 1, is related to
but different from the observed power-law distribution. However, Huberman and
Adamic (1999) report that, if this model is modied to include a wide distribution of
growth rates across different sites and/or the fact that sites have different ages (with
many more young sites than old ones), then simulations show that the distribution of
sizes across sites obeys a power-law distribution (see Mitzenmacher (2002) for more
general background about the relationship between the log-normal and power-law
distributions).
It must be pointed out, however, that the simple Fermi model of Chapter 1 pro-
vides a somewhat cleaner and analytically tractable model that seems to capture the
power-law distribution of website sizes, at least to a rst degree of approximation.
To see this, instead on focusing on daily uctuations, which are more of a second-
order phenomenon, it is sufcient to consider that websites are being continuously
created, that they grow at a constant rate during a growth period, after which their
size remains approximately constant, and that the periods of growth follow an expo-
nential distribution. Assuming, consistently with observations, a power-law exponent
= 1.08 for the size, this would give a relationship = 0.8 between the rate of
the exponential distribution and the growth rate .

3.2.2 Lattice perturbation models: between order and disorder

Watts and Strogatz (1998) have argued that the small networks found in biological,
technological, social, and man-made networks are somewhere between completely
organized (regular) networks, and completely disorganized networks. They are highly
clustered at the local level and have short characteristic path lengths (Table 3.2).
Hence, they have suggested a model based on the idea of progressively increasing the
level of disorder in a regular system.
More specically, they start from a regular arrangements of points, such as a lattice
with constant degree k and neighborhoods that are entirely local. Then they progres-
sively visit all the vertices and edges in the graph and randomly rewire each of the
local edges into a long-ranged edge with probability p, avoiding edge duplication.
Obviously, p = 0 corresponds to complete order and, as p is increased, the graph
becomes increasingly disordered until p = 1, at which stage the graph is completely
random. Note, however, that the procedure does not alter the total number of vertices
or edges in the graph, nor the average connectivity.
62 GENERATIVE MODELS FOR THE WEB GRAPH

Two quantities can be used to monitor the evolution of the undirected graph structure
during the rewiring process. First, the average diameter, d = d(p), corresponding
to the average distance (number of edges) between any two vertices, which is a
global property of the graph. Second, a more local property measuring the average
density of local connections, or cliquishness, dened as follows. If a vertex v has kv
neighbors, there are at most kv (kv 1)/2 edges between the corresponding nodes. If
cv is the corresponding fraction of allowable edges that
 actually occur, then the global
cliquishness is measured by the average c = c(p) = v cv /|V |, where |V | is the total
number of vertices. In a social network, d is the average number of friendships in the
shortest path between two people, and cv reects the degree to which friends of v are
friends of each other.
To ensure that the graph is both sparse but connected, the value k satises
n  k  log n  1,
where k  log n ensures that a corresponding random uniform graph remain con-
nected (Bollobs 1985). In this regime, Watts and Strogatz (1998) nd that d
n/2k  1 and c 3/4 as p 0, while d drandom log n/ log k and
c crandom k/n  1 as p 1. Thus, the regular lattice with p = 0 is a highly
clustered network where the average diameter d grows linearly with n, whereas the
random uniform network p = 1 is poorly clustered, with d growing logarithmically
with n. These extreme cases may lead one to conjecture that large c is associated with
large d, and small c with small d, but this is not the case.
Through simulations, in particular for the case of a ring lattice, Watts and Stro-
gatz nd that for small values of p the graph is a small-world network, with a high
cliquishness like a regular lattice and a small characteristic path length like a traditional
random graph. As p is increased away from zero, there is a rapid drop in d = d(p)
associated with the small-world phenomena, during which c = c(p) remains almost
constant and equal to the value assumed for the regular lattice over a wide range of
p. Thus, in this model, the transition to a small-world topology appears to be almost
undetectable at the local level. For small values of p, each new long-range connection
has a highly nonlinear effect on the average diameter d. In contrast, removing an edge
from a clustered neighborhood has at most a small linear effect on the cliquishness c.
In the case of a one-dimensional ring lattice, a mean eld solution for the average
path length and for the distribution of path lengths in the model is given in Newman
et al. (2000). The basic idea behind the mean eld approximation is to represent the
distribution of relevant variables over many realizations by their average values. The
authors apply the approximation to the continuous case rst and use the fact that when
the density of shortcuts is low, the discrete and continuous models are equivalent (see
also Barthlmy and Amaral (1999) for an analysis of the transition from regular
local lattice to small-world behavior and Amaral et al. (2000) for a classication of
small-world networks).
Even as a very vague model of the Web graph, however, the lattice perturbation
model has several limitations. First, there is no clear concept of an underlying lattice
and the notion of short and long links on the Web is not clearly dened. While short
WEB GRAPHS 63

Table 3.2 Small-world networks according to Watts and Strogatz (1998). n, number of ver-
tices; k, average degree; d, average distance or diameter; c, cliquishness. Random graphs have
the same number of vertices and same average connectivity. The difference in average connec-
tivity in the actor networks with respect to Table 3.1 results apparently from the inclusion of
TV series in addition to lms (A.-L. Barabasi 2002, personal communication).

n k d drandom c crandom

Actors 225 226 61 3.65 2.99 0.79 0.000 27

Power grid 4 941 2.67 18.7 12.4 0.080 0.005
C. elegans 282 14 2.65 2.25 0.28 0.05

links could correspond to links within organizations and some correlation between
link density and geographical distance may be expected, the whole point of the Web
is to make a link between pages that are geographically distant as easy to create
as a link between pages that are geographically close. More importantly perhaps, it
can be shown that the edge rewiring procedure does not yield a graph with power-
law distributed connectivity (see also Barthlmy and Amaral 1999). The degree
distribution remains bounded and strongly concentrated around its mean value. Thus,
although the lattice rewiring procedure yields small-world networks, it does not yield
a good model for the connectivity of the Web graph. Other models are needed to try
to account for the scale-free connectivity distribution.

3.2.3 Preferential attachment models, or the rich get richer

There are two aspects of many real networks that are not addressed by either the
traditional random graph models, or by the models in Watts and Strogatz (1998)
described above. First, in the traditional random graph approach or in the disordered
lattice approach, the number of vertices is xed in advance rather than growing in
time. Second, in many real networks, edges are not created independently at random,
but rather seem to follow some preferential attachment rule. The probability that a
new actor is cast with a well-established actor is higher than the probability of them
being cast with a less well-known actor. Likewise, a new Web page is more likely to
be linked to a well-known page, which is already well connected.
In Barabsi and Albert (1999) a rich-get-richer model is proposed that incorporates
these features. The model starts with a set M0 of vertices, at time zero. At each time
step, a new node is created together with m
M0 links. These new links connect
the new node to m randomly selected nodes from the set of already existing nodes.
Preferential attachment is modeled by assuming that the probability that a newly
created vertex v is connected to an existing vertex w is proportional to the  degree
kw (t) of w, so that the corresponding probability of attachment is given by kw / r kr .
This is the simplest formulation but other forms of preferential attachment are of
course possible. After t time steps, the network has M0 + t vertices and mt edges.
Simulations show that such a model evolves into a scale-free network, where the
64 GENERATIVE MODELS FOR THE WEB GRAPH

probability that a vertex has connectivity k follows a power-law distribution with

exponent = 2.9 with M0 = m = 5. Simulations also show that after a relatively
short transient phase, this distribution becomes stationary and does not depend on
the network size. The stationary power-law distribution is not observed if there is
preferential attachment and the number of vertices is kept xed or, vice versa, if the
number of nodes grows but attachment is random and uniform (see also Barabsi et al.
1999). In the case of uniform attachment for instance, the degrees are geometrically
distributed (see Barabsi et al. (1999) for an intuitive argument and Bollobs et al.
(2001) for a rigorous proof).
This behavior can be understood using a simple mean eld treatment assuming
continuous connectivity (Barabsi et al. 1999). The basic principle behind a mean
eld analysis is to replace variable quantities by their averages. Consider vertex v
after t time steps with a degree kv (t) when the total number of edges is mt. When a
new vertex is added, the probability that it is joined to v is mkv (t) divided by the sum
of all the degrees, which is twice the total number of edges. By taking expectations
and assuming that multiple edges between nodes are allowed, we have from the
preferential attachment rule that
kv (t) kv
kv (t + 1) kv (t) = m = . (3.6)
2mt 2t
Moving to a continuous formalism and replacing difference equations by differential
equations, the rate at which a vertex v acquires edges is kv /t = kv /2t. If tv is
the time at which vertex v is added to the network, we have the initial condition
kv (tv ) = m giving kv (t) = m(t/tv )0.5 . Thus, the connectivity of older vertices with
a smaller tv grows faster than the connectivity of younger vertices. This can also be
used to get an estimate of the exponent . Indeed,
   
mt 2 mt 2 m2 t
P (kv (t) < k) = P tv > 2 = 1 P tv
2 = 1 2 . (3.7)
k k k (t + M0 )
The last equality assumes that vertices are added at equal time intervals to the sys-
tem, yielding the density P (tv ) = 1/(M0 + t). The probability density P (k) satises
P (k) = [kv (t) < k]/k, which leads to the stationary solution P (k) = 2m2 /k 3 cor-
responding to = 3. The authors also report that nonlinear forms of preferential
attachment do not lead to stationary power-law distributions.
A mathematically rigorous treatment of this model can be found in Bollobs et al.
(2001), who note that the description of the model is rather vague in two respects.
First, when there are no edges in the network, the preferential attachment distribution
is not dened and if kv = 0 it remains equal to zero in the future. This can easily be
remedied by using a uniform distribution on the rst M0 vertices, until each one of
them has degree at least one. More importantly, the question of how to choose the m
earlier vertices of attachment at each time step is not precisely dened. Bollobs et
al. (2001) propose a more precise version, where vertices are chosen one at a time
and both multiple edges and loops are allowed. As it turns out, multiple edges and
loops remain rare and therefore do not really impact the overall behavior, or the value
WEB GRAPHS 65

of the exponent. Using this precise model, Bollobs et al. (2001) prove that = 3,
rst for the case of m = 1 and then for the general case for degrees up to O( n), by
deriving the general case from the case of m = 1.
It has been shown, via simulations, that this model of preferential attachment leads
to small-world networks, i.e. graphs with small diameter, with an asymptotic diameter
of roughly log n, where n is the number of growth steps (or, equivalently, the number
of vertices). Bollobs and Riordan (2003) have shown rigorously that log n is the
right asymptotic value when m = 1 and that for m  2 the correct asymptotic
bound is in fact log n/ log log n. It is worth observing that for m = 1 the graphs
are essentially trees so that the bound in this case, as well as other results, can be
derived from the theory of random recursive trees (see Mahmoud and Smythe 1995;
Pittel 1994). (A recursive tree is a tree on vertices numbered {1, 2, . . . , n}, where
each vertex other than the rst is attached to an earlier vertex. In a random recursive
tree this attachment occurs at random with a particular distribution such as uniform,
preferential attachment, and so forth.)
Of interest is the observation that a diameter growing like log n is in fact very com-
mon in a variety of random graphs, including random regular graphs (Bollobs and
de la Vega 1982). Thus, from a random graph standpoint, the small-world properties of
the Internet are perhaps not that surprising. In fact, rather than our naive six-degree-
of-separation surprise, the question becomes rather why the diameter would be so
large! The smaller diameter of log n/ log log n obtained when m > 1 is a step in that
direction.
The model above only yields a characteristic exponent value of = 3 and produces
only undirected edges. The graphs it produces are also too structured to be good
models of the Internet: for instance when m = 1 the graph consists of M0 trees, since
there are M0 initial components and m = 1 ensures that each component remains a
tree. It is clear, however, that the model can be modied to produce directed edges,
accommodate other exponents, and to produce more complex topologies.
To address the edge orientation problem, the orientation of each new edge could
be chosen at random with a probability of 0.5 and the preferential attachment rule
could take into consideration both indegrees and outdegrees. The xed exponent
problem can be obviated, for instance, with a richer mixture model where not only
new nodes, but also new links are added (without adding nodes), and links are also
rewired (Albert and Barabsi 2000) (see also Cooper and Frieze (2001) for a similar
model and an asymptotic mathematical analysis based on linear difference equations).
In one implementation, the model above is extended by incorporating at each step
the possibility of adding m new edges (without adding nodes) or rewiring m existing
edges guided, in both cases, by the preferential attachment distribution. The three basic
processes adding nodes, rewiring, adding links are assigned mixture probabilities
p, q and 1 p q. An extension of the mean eld treatment given above shows that
depending on the relative probability of each one of the three basic processes and m
(the number of new nodes or edges at each step), the system can lead to either an
exponential regime or a power-law regime, with different power-law exponents .
Clearly these variations also break the tree-like structure of the simple model.
66 GENERATIVE MODELS FOR THE WEB GRAPH

The preferential attachment model is quite different from the standard random graph
models where the size is xed. Even without the preferential attachment mechanism,
when graphs grow older, nodes tend to have higher connectivity simply due to their
age and this fact alone tends to remove some of the randomness (Callaway et al. 2001).
The simplicity of the preferential attachment model is its main virtue but also its main
weakness and it can only be viewed at best as a rst-order approximation. While
the model does reproduce the scale-free and small-world properties associated with
the Web graph, it is clear that more realistic models need to take into account other
higher order effects, including deletion of pages or links, differences in attachment
rates, weak correlations to a variety of variables such as Euclidean distance, and so
forth. Deviation from power-law scaling has been observed, especially for nodes with
low connectivity. In addition, the deviations seem to vary for different categories of
pages (Pennock et al. 2002). For example, the distribution of hyperlinks to university
home pages diverges strongly from a power-law distribution and seems to follow a far
more uniform distribution. Additional work is needed to better understand the details
of the distribution and its uctuations, and to create more precise generative models
that can capture such uctuations and other higher order effects.

3.2.4 Copy models

Kleinberg et al. (1999) have proposed a different copy model, where vertices are
born one at a time but, instead of preferential attachment, each new vertex chooses
an existing vertex and copies a random subset of its links (see also Gilbert 1997)
or, alternatively, chooses new neighbors uniformly on its own. This copy process is
in part suggested by the fact that often when a new page is created it is inspired by
another page and retains a subset of its links. Naturally, processes for deletion of
nodes and edges can also be added to the model.
To be more precise, in the simplest version of these models, at each time step a
node is added. With probability p a new edge is created between the new node and
one of the pre-existing nodes chosen uniformly. With probability 1 p, one of the
pre-existing nodes and one of its pre-existing edges is chosen uniformly and the edge
is copied (in the sense that a new link is created from the new page to the same page
pointed to by the pre-existing edge). This can again be viewed as a mixture of two
different elementary processes.
Simulations in Kumar et al. (1999a, 2000) report that this model also leads to power-
law degree distribution and also to a high number of dense bipartite subgraphs found
in the Web graph (see below). The copying process, however, introduces complex
dependencies between the variables and analytical treatments of these copy models
have not been developed so far. It is also clear that in its most simple form the copy
model yields a tree with n vertices and n 1 edges which is too crude to provide an
approximation of the Web graph. Finally, while copy models were originally intro-
duced as an alternative to the simple rich-get-richer model, they can in fact be viewed
as a generalization. In the simple version above, with probability p an edge is cre-
ated to an existing node uniformly, and with probability 1 p an edge is created
WEB GRAPHS 67

to an existing node with a probability that in the directed case strongly depends on
its indegree. If an existing page or node v has indegree |pa[v]| and each of its par-
ents i = 1, . . . , |pa[v]|has outdegree |ch[i]|, then the probability of connecting to v
through the copy mechanism is given by
|pa[v]|
 1
.
n|ch[i]|
i=1

This is because once the copy mechanism has been selected, there is a 1/n chance of
choosing a given parent i of v, and a 1/|ch[i]| chance of choosing the edge running
from i to v. Thus, everything else being equal, the copying process favors creating a
link to nodes that already have a large indegree and therefore is a form of preferential
attachment based on vertex degrees (to be exact, the most favored attachment nodes
are those with very high indegree and whose parent pages have low outdegree).

3.2.5 PageRank models

It should be clear by now that alternative or more general growth models can be
created by using a mixture of elementary mechanisms, and by using some form of
preferential attachment based on any kind of node or page metric. In particular, we
have seen that the relevance of a page can be assessed by its PageRank and that in some
ways this may be a good alternative to using the indegree of that page. As a result,
it is possible to develop models of preferential attachment based on PageRank (or
other measures) rather than connectivity. The model proposed in Pandurangan et al.
(2002) is an example of a mixture model where a new node is uniformly connected
with probability p to an existing node, and with probability 1 p to an existing
node in proportion to its PageRank. An obvious mixture generalization is to have a
probability p for selecting nodes uniformly, a probability q (p + q
1) of selecting
nodes proportionally to their Page Rank, and a residual probability 1 p q of
selecting nodes proportionally to their degree.
Experiments and mean eld analysis are reported in Pandurangan et al. (2002)
for the brown.edu domain (roughly 100 000 pages and 700 000 hyperlinks), the
smaller cs.brown.edu domain (roughly 25 000 pages and 175 000 hyperlinks), and
the larger WT10g corpus (for details see http://www.ted.cmis.csiro.au/
TRECWeb/wt10ginfo.ps.gz), containing roughly 1.69 million documents. The
authors conrm the power-law distribution of indegrees, outdegrees, degrees, and
PageRanks across all three scales. Mean eld analysis and simulations suggest that
preferential attachment guided by PageRank alone is not sufcient to reproduce all
power-law behaviors. The power-law behavior of the indegree, outdegree, and Page-
Rank distributions can be tted by a preferential attachment model based on degree
alone. In addition, the three distributions can also be tted by a uniform simulated
mixture model, where the probability of attachment based on degree is equal to the
probability of attachment based on PageRank (i.e. p q 1 p q 0.33).
68 APPLICATIONS

In time, additional components could be added to the mixture to reect attachment

biases that are driven by other considerations than the local (degree) or global (Page-
Rank) popularity of a node such as, for instance, the thematic aspects associated with
local communities and locally densely interconnected subgraph described below.

3.3 Applications
The graphical properties of the Web and other networks are worth investigating for
their own sake. In addition, these properties may also shed light on some of the
functional properties of these networks, such as robustness. In the case of the Web,
they may also lead to better algorithms for exploiting the resources offered by the
underlying network. Systems with small-world properties may have enhanced signal-
propagation speed and synchronizability. In particular, on the Web, an intelligent
agent that can interpret the links and follow the most relevant ones should be able
to nd information rapidly. An intelligent agent ought to be able also to exploit
the correlations that exist between content and connectivity. Furthermore, thematic
communities of various sorts ought to be detectable by patterns of denser connectivity
that stand above the general background of hyperlinks.

3.3.1 Distributed search algorithms

The existence of short paths between two vertices in the Web and many other small-
world networks arising in nature and technology is signicant but raises the questions
of whether such short paths, or an approximation to them, can be found efciently
using local knowledge. Because the average path length is typically bounded by a
polynomial in log n, mathematically it would be very satisfying if local algorithms
could be developed that can nd a short path between two vertices in the network in
average time also bounded by a polynomial of log n. This is in sharp contrast with the
polynomial bound for nding the shortest path using global knowledge of the graph
and breadth-rst/dynamic programming (Cormen et al. 2001).
The existence of navigation algorithms with average time bounded by a polynomial
in log n seems plausible for at least two reasons:
(a) we do not really need to nd the shortest path, just a reasonably short path; and
(b) unlike the case of breadth rst, which requires examining essentially all the
possible paths, the local knowledge or latent structural clues available in the
network may provide a means of restricting the search only to a small subset
of most promising paths.
To some extent, the feasibility of routing with local information in certain complex
networks is corroborated by experiments. In the original experiment described by
Milgram (1967), several people in a small town in Nebraska were given the task
of sending a letter to a target in Massachusetts, with the instruction that the letter
WEB GRAPHS 69

be forwarded at each step by the sender to someone they knew by their rst name.
The experiment revealed that the typical letter path had a length of six, hence the
term six degrees of separation. The success of this experiment suggests that it may
possible for intelligent agents to nd short paths in complex networks using only local
address information. Furthermore, the experiment also provides some clues about
possible algorithms. The geographic movement of the [message] from Nebraska to
Massachusetts is striking. There is a progressive closing in on the target areas as each
new person is added to the chain (Milgram 1967; see also Hunter and Shotland 1974;
Killworth and Bernard 1978; White 1970).
This observation, as well as casual observation of how we seek information over the
Web, suggests a trivial greedy algorithm, where at each step the agent locally chooses
whichever link is available to move as close as possible to the target. The problem
lies in the denition of closeness and the kind of information available regarding
closeness at each node. Clearly, in the extreme case, where distances are measured in
terms of path lengths in the underlying graph and where the distance to the target node
is available at each node (but not necessarily the path), an agent can nd the shortest
path requiring a number of steps equal to the distance (hence a polynomial in log n
steps on average in a small-world network) using the obvious greedy algorithm. But
agents navigating the Web using local information would not have precise access to
graph distances. Thus, more sophisticated or mathematically interesting models must
rely on a weaker form of local knowledge and weaker estimates or approximations
to the distance to the target. At one other simple extreme, consider that, when the
agent examines all the links emanating from one node, instead of having access to the
precise information on how much closer or further away the link would bring him to
the target node, the only local information he is given is noisy binary information on
whether each link brings him closer to or further away from the target.
In Kleinberg (2000a,b, 2001) local greedy navigation algorithms are discussed for
lattices that have been perturbed, along the lines of the model in Watts and Strogatz
(1998), with directed connections. These are square n = l l lattices in two dimen-
sions or, more generally, n = l m cubic lattices in m dimensions. Each vertex is fully
connected to all its neighbors within lattice distance l0 , with an additional q inde-
pendent long-ranged connections to nodes at lattice distance r chosen randomly with
probability proportional to r . When = 0, this corresponds to a uniform choice
of long-ranged connections.
The greedy algorithm to navigate from u to v always chooses the connection that
gets the closest to the target. In two dimensions (m = 2), the greedy algorithm has
a rapid delivery time, measured by the expected number of steps, which is bounded
by a function proportional to (log n)2 provided = m = 2. In particular, in the case
of uniform long-range connections ( = 0), no decentralized algorithm can achieve
rapid delivery time. This uniform case corresponds to a world in which short chains
exist but individuals, faced with a disorienting array of social contacts, are unable to
nd them (Kleinberg 2000b). For  = m, the delivery time is asymptotically much
larger and greater than a polynomial in l. Similar results hold in m dimensions, with
rapid (polynomial in log n) delivery if and only if = m. Likewise, in m dimensions,
70 APPLICATIONS

the greedy algorithm has a rapid delivery time bounded by a polynomial in log n
provided that = m. For  = m, the delivery time is asymptotically much larger.
This result is interesting but should be viewed only as a rst step. Its signicance
is undermined by the fact that the graphs being considered are regular instead of
power-law distributed. Furthermore, the behavior does not seem to be robust and
requires a very specic value of the exponent . Finally, the routing algorithm requires
knowledge of lattice distances between nodes and the global position of the target
that may not be trivial to replicate in a realistic Web setting. Further extensions and
analysis are the object of ongoing research in the context, for instance, of peer-to-
peer systems, such as Gnutella (Adamic et al. 2001), or Freenet (Zhang and Iyengar
2002). A model for searchability in social networks that leverages the list of attributes
identifying each individual is described in Watts et al. (2002).
From a different perspective, given the problem of navigating from u to v, an
intelligent agent starting at u can be modeled as an agent that somehow knows whether
any given edge brings it closer to its target v, with some error probability. This captures
the idea that a human trying to nd the date of birth of Albert Einstein is more likely
to follow a link to a page on Nobel prizes in physics than a link to a page on Nobel
prize in economics. If, for the sake of the argument, in a small-world network the
typical distance between u and v is log n, then an agent that can detect good hyperlinks
with an error smaller than 1/ log n will typically make no mistakes along a path of
length log n and therefore will typically reach the target. However, even if the agent
has an error rate of 0.5, then, by a simple diffusion (or random walk) argument, it will
take on the order of log2 n steps to reach the target, which is still polynomial in log n.
Thus, in this sense, the problem of building an intelligent agent, can be reduced to
the problem of building an agent that can discern good links with a 0.5 error rate or
better.

3.3.2 Subgraph patterns and communities

Both the real life networks and their models studied in this chapter show that the prop-
erty of a power-law degree distribution and small diameter do not entirely characterize
all the statistical topological properties of a network, and that there are additional prop-
erties and structures that can be considered which may be more characteristic of cer-
tain specic networks. In particular, there are subgraphs of the Web graph that appear
to occur more frequently than one would expect by chancel and which have been
used both to discover content as well as communities and to follow their dynamics
(Chakrabarti et al. 1999a, 2002; Flake et al. 2000, 2002; Gibson et al. 1998; Kleinberg
1999; Kleinberg and Lawrence 2001; Kumar et al. 1999b). For example, clique-like
structures (Kumar et al. 1999b) are overabundant in the Web graph: another departure
from standard random graphs. In fact, particular subgraph motifs appear to form the
building blocks and characteristic signatures of complex networks (Milo et al. 2002;
Oltvai and Barabsi 2002).
A statistically signicant high density of hyperlinks among a set of Web pages is an
indication that they may be topically related. We have seen that the basic background
WEB GRAPHS 71

density of the Web graph corresponds to a sparse graph where the number of edges
scales linearly with the number of vertices. Deviations from this background can be
dened in many ways. In one possible denition, a community can be dened as any
cluster of nodes such that the density of links between members of the community
(in either direction) is higher than the density of links between members of the com-
munity and the rest of the network. Variations on this denition are possible to detect
communities with varying levels of cohesiveness.
Another characteristic pattern in many communities is the hubsauthority bipartite
graph pattern (Kleinberg and Lawrence 2001), where a hub is an index or resource
page that points to many authority or reference pages in a correlated fashion. This is
studied further in Chapter 5. Likewise, an authority page is a page that is pointed to
by many hub pages. Other patterns discussed in the literature include the case where
authorities are linked directly to other authorities (Brin and Page 1998).
Once a particular subgraph pattern is selected as a signature of a community, the
Web graph can be searched to identify the occurrence of such signatures. In many
cases, the general graphical search problem is NP-complete. However, polynomial
time algorithms can be derived with some additional assumptions. In Flake et al.
(2002), for instance, communities are associated with maximum ow algorithms.
Large-scale studies reported in the literature have identied hundreds of thousands
of communities, including some unexpected ones (e.g. people concerned with oil
spills off the coast of Japan). Although analyses that are purely structural can discover
interesting results, it is clear that it is possible to integrate structure and content
analysis (Chakrabarti et al. 2002). For instance, once a community is found based
on hyperlink structure, words that are common to the pages in the community can
be identied and used to further dene and reliably extract the community from the
rest of the network. Likewise, textual analysis can be used to seed a community
search.
Discovering communities and their evolution may be useful for a number of appli-
cations including:

(a) catching trends at early stages;

(b) facilitating the transmission of information (or misinformation) and bringing

together individuals with common interests; and

(c) targeted marketing.

The importance of link analysis for search engines and information retrieval will be
explored further in Chapter 5. It must be pointed out, however, that the utilization of
link analysis in these applications opens the door for second-order applications. For
instance, search engines based on link analysis may tend to increase the popularity
of pages that are already popular and it may be important to develop methods to
attenuate this effect.
72 APPLICATIONS

3.3.3 Robustness and vulnerability

There is a clear sense that many of the networks studied in this chapter are in some
sense robust. In spite of frequent router problems, website problems, and temporary
unavailability of many Web pages, we have not yet suffered a global Web outage.
Likewise, mutation or even removal of a few enzymes in E. coli is unlikely to sig-
nicantly disrupt the function of the entire system. Mathematical support for such
intuition is provided by studying how the diameter or connectivity of the graph is
affected by deleting nodes randomly. When a few nodes are deleted at random, the
graph remains by and large connected and the small-world property remains intact.
In other words, the systems are robust with respect to random errors or mutations.
However, as pointed out in Albert et al. (2000), the systems considered in this chapter
are vulnerable to targeted attacks that focus on the nodes with the highest connectivity.
These targeted attacks can signicantly disrupt the properties of the systems.
To be more specic, with the same number of nodes and connections, a scale-free
graph has a smaller diameter in terms of average shortest path than a uniform random
graph. In a random uniform graph (sometimes also called random exponential graph,
when referring to the decay of the connectivity) all nodes are more or less equivalent.
In simulations reported in Albert et al. (2000) using random exponential and scale-
free graphs, when a small fraction f (up to a few percentage points) of nodes are
removed at random, the diameter increases monotonically (linearly) in an exponential
graph, whereas the diameter remains essentially constant in a scale-free graph with
the same number of nodes and edges. Thus, in scale-free graphs, the ability of nodes
to communicate is unaffected even by high rates of random failures. Intuitively, this
is not surprising and is rooted in the properties of the degree distribution of each kind
of graph. In an exponential network, all the nodes are essentially equivalent and have
similar degrees. As a result, each node contributes equally to the connectivity and to
the value of the diameter. In a scale-free graph with a skewed degree distribution,
randomly selected nodes are likely to have small degrees and their removal does not
disrupt the overall connectivity.
This situation, however, is somewhat reversed during an attack where specic nodes
are targeted. An attack is simulated in Albert and Barabsi (2000) by removing nodes
in decreasing order of connectivity. In an exponential graph, there is essentially no
difference between random deletions and attack. In contrast, the diameter of the scale-
free graph increases rapidly in the scale-free network, roughly doubling when 5% of
the nodes are removed. A higher degree of robustness of the undirected Web graph to
attacks, however, has been reported in other studies (Broder et al. 2000).
Similar effects are also observed when the connected components of both kinds of
graphs are studied, both under random error and targeted attack scenarios. In partic-
ular, the scale-free network suffers a substantial change in its topology under attack,
in the sense that it undergoes fragmentation and breaks down into many different
connected components when under attack, whereas it basically remains connected
during random removal of a few nodes.
WEB GRAPHS 73

To be more precise, two statistics that can be used to monitor the connected compo-
nents during node deletion are the size S of the largest connected component relative
to the size of the system and the average size s of the connected components other
than the largest one. In the case of an exponential graph, when f is small both S 1
and s 1 and only isolated nodes fall off the network. As f increases, s increases,
until a phase transition occurs around f = 0.28, where the connectivity breaks down,
the network fragments into many components, and s reaches a maximum value of 2.
Beyond the transition point, s decreases progressively toward unity and S rapidly
decreases toward zero. This behavior is observed both under random node deletion
and under attack.
In contrast, in a scale-free network under random deletion of nodes as f is increased
S decreases linearly and s remains close to unity. In case of attack, however, the
network breaks down in a manner similar to but faster than the breakdown of the
exponential network under random deletion: S drops rapidly to zero with a critical
threshold of 0.18 instead of 0.28. s has a peak of two, reached at the critical point.
When tested on a directed subgraph of the entire Web, containing 325 729 nodes and
1 469 680 links, this critical point is found to occur at f = 0.067 (Albert and Barabsi
2000).
There are many other important aspects of Internet/Web networksrobustness, secu-
rity, and vulnerability that cannot be addressed here for lack of space. We encourage
the interested reader to explore the literature further and limit ourselves to providing
two additional pointers. On the vulnerability side, it is worth mentioning that the
architecture and protocols of the Internet can be exploited to carry parasitic computa-
tions in which servers unwittingly perform computations on behalf of a remote node
(Barabsi et al. 2001). On the robustness side, peer-to-peer networks such as Gnutella
(Ripeanu et al. 2002) or Freenet (Clarke et al. 2000, 2002) are robust Internet network
architectures that do not rely on centralized servers.

3.4 Notes and Additional Technical References

Besides growth models, there are different possible mathematical approaches to den-
ing random power-law graphs. Some of these denitions generalize the original def-
initions of uniform random graphs given in Erdos and Rnyi (1959, 1960). One
possibility similar to G(n, e) (where e = |E|) is to introduce a uniform distribution
over all graphs that satisfy a particular power-law degree sequence. This is a special
case of random graphs with a given degree sequence, for which there is an abundant
literature. More precisely, the class G(, ) is dened by considering a uniform distri-
bution over all graphs where the number y of nodes with degree x satises y = e /x .
Alternatively one can use a denition similar to the denition of the class G(n, p)
for uniform random graphs, yielding an expected degree sequence rather than a xed
degree sequence (Chung and Lu 2002). As in the uniform case, the two models are
essentially equivalent, at least in an asymptotic sense. Theorems on the connectivity
of random power-law graphs are given in Aiello et al. (2001). In particular:
74 EXERCISES

when < 1, the graph is almost surely connected;

when 1 < < 2, there is a giant component of size O(n) and all smaller
components have size O(1);
when 2 < < 0 = 3.4785, there is a giant component of size O(n) and all
smaller components have size O(log n) (the size is O(log n/ log log n));
when 0 < , almost surely there are no giant components.
Thus, in short, when is small there is a giant component and when > 0 there
is no giant component, but a large number of small components. Additional results
include the fact that for > 4, the number of components of a given size follows
approximately a power-law distribution (see also Chung and Lu 2002). For compar-
ison purposes, a different, simpler, and relevant class of random graphs with given
degree sequence are random regular graph with xed degree. For instance, the diam-
eter of random directed graphs with constant outdegree is studied in Philips et al.
(1990).
Because the networks studied in this chapter are sparse, another relevant class of
random graphs is random sparse graphs, i.e. random graphs with |E| = o(n2 ) edges.
This class can again be dened by, for instance, introducing a uniform distribution
over all sparse graphs. The diameter of random sparse graphs is studied in Chung
and Lu (2001). Finally, quasi-random graphs are studied in Chung et al. (1989) and
sparse quasi-random graphs are studied in Chung and Graham (2002). Chung et al.
(2001) studies the problem of constructing a weighted graph to optimally realize a set
of pairwise distances between a subset of the vertices, with an application to what has
been called Internet tomography (reconstruction of the whole from low-dimensional
slices).

3.5 Exercises
Exercise 3.1. Provide at least two different back-of-the-envelope estimates for the
size (in terms of the number of vertices and the number of edges) for the graph
associated with the documents and hyperlinks of the Web. Rene such estimates by
sampling IP addresses (using the ping command) or Web pages by some kind of
crawling. Discuss the limitations of these estimates and how they could be improved.
For instance, the ping command samples only active machines that are authorized
to reply to the ping query. Consider also the possibility of probing typical ports
(e.g. 80, 8080).
Exercise 3.2. Crawl the entire website of your university or some other large
organization and compute the statistics of the corresponding graph including the
distribution of the connectivity and the diameter both in the directed and undi-
rected cases. [Easier alternatives to the crawl consist in downloading a publicly
available data set or using a program such as wget (http://www.gnu.org/
software/wget/wget.html).]
WEB GRAPHS 75

Exercise 3.3. Show that, if the connectivity of a large graph with n vertices has a
power law with exponent  2, then the graph is sparse in the sense that it has O(n)
vertices.
Exercise 3.4. In this chapter, we have listed the power-law exponent observed for
several phenomena. In general, the values of the exponent are relatively small, say
less than 10. How general is this and why?
Exercise 3.5. Fill in some of the entries in Table 3.1.
Exercise 3.6. Explore possible relationships between the Fermi model of Chapter 1
and the models introduced in this chapter to account for the power-law distribution
of the size of the sites, or the power-law distribution of the connectivity.
Exercise 3.7. Simulate directly random directed graphs where the indegrees and
outdegrees of the nodes follow a power-law distribution with the exponents given in
Table 3.1. Estimate the average and maximal shortest distance between vertices, and
how it scales with the number n of vertices. Study the distribution of this shortest
distance across all possible pairs of vertices.
Exercise 3.8. Run simulations of lattice perturbation models, along the lines of those
found in Watts and Strogatz (1998), on ring, square, and cubic lattices. Progressively
increase the amount of random long-ranged connections in these graphs and monitor
relevant quantities such as degree distribution, cliquishness, and average/maximal
distance between vertices.
Exercise 3.9. Simulate the rich get richer model and monitor relevant quantities
such as degree distribution, cliquishness, and average/maximal distance between ver-
tices. Examine modications of the basic model incorporating any of the following:
edge orientation, edge rewiring, edge creation and monitor relevant quantities. Try to
analyze these modications using the mean eld approach associated with Equa-
tions (3.6) and (3.7). Examine a growth model with uniform, rather than preferential,
attachment. Show through simulations or a mean eld type of analysis that the dis-
tribution of degrees is geometric rather than power law.
Exercise 3.10. Simulate a version of the copy model and model relevant quanti-
ties such as degree distribution, cliquishness, and average/maximal distance between
vertices.
Exercise 3.11. First simulate a random network with power-law connectivity and
small network properties using, for instance, the preferential attachment model of
Barabsi and Albert (1999). Then consider an agent that can navigate the network
and is faced with the typical task of nding a short path between two vertices u and v
using only local information. Suppose each time the agent considers an edge, he has
binary information on whether this edge brings it closer or not to the target, and that
this information comes with some error rate p. Design a navigation strategy for the
agent and estimate its time complexity by simulations and by a back-of-the-envelope
calculation. Simulate this strategy on several of the networks discussed in this chapter.
76 EXERCISES

Study how the typical travel time varies with p. Study the robustness of this strategy
with respect to uctuations in p. Modify the strategy to include a bias toward nodes
with high degree, or high PageRank and study its time complexity. How would you
model the effect of a search engine in this framework?
Exercise 3.12. Run simulations to study the effect of random and selective node
deletion on the diameter and fragmentation of scale-free and exponential graphs.
4

Text Analysis

Having focused in earlier chapters on the general structure of the Web, in this chapter
we will discuss in some detail techniques for analyzing the textual content of individ-
ual Web pages. The techniques presented here have been developed within the elds
of information retrieval (IR) and machine learning and include indexing, scoring,
and categorization of textual documents.
The focus of IR is that of accessing as efciently as possible and as accurately as
possible a small subset of documents that is maximally related to some user inter-
est. User interest can be expressed for example by a query specied by the user.
Retrieval includes two separate subproblems: indexing the collection of documents
in order to improve the computational efciency of access, and ranking documents
according to some importance criterion in order to improve accuracy. Categoriza-
tion or classication of documents is another useful technique, somewhat related to
information retrieval, that consists of assigning a document to one or more predened
categories. A classier can be used, for example, to distinguish between relevant
and irrelevant documents (where the relevance can be personalized for a particular
user or group of users), or to help in the semiautomatic construction of large Web-
based knowledge bases or hierarchical directories of topics like the Open Directory
(http://dmoz.org/).
A vast portion of the Web consists of text documents thus, methods for auto-
matically analyzing text have great importance in the context of the Web. Of course,
retrieval and classication methods for text, such as those reviewed in this chapter can
be specialized or modied for other types of Web documents such as images, audio
or video (see, for example, Del Bimbo 1999), but our focus in this chapter will be on
text.

4.1 Indexing
4.1.1 Basic concepts
In order to retrieve text documents efciently it is necessary to enrich the collection
with specialized data structures that facilitate access to documents in response to
Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
78 INDEXING
1
A manufacturer, importer, or seller of
digital media devices may not (1)
2, 83 sell, or offer for sale, in interstate
commerce, or (2) cause to be
3, 79 transported in, or in a manner
affecting, interstate commerce, a
computer digital media device unless the device
includes and utilizes standard security
2, 57 technologies that adhere to the
2, 245 security system standards....
books 2
Of course, Lissa did not necessarily
intend to read his books. She might
want the computer only to write her
security midterm. But Dan knew she came
from a middle-class family and could
1, 278 hardly afford the tuition, let alone her
1, 319 reading fees. Reading his books
protected might be the only way she could
3, 142 graduate....
3
Research in analysis (i.e., the
evaluation of the strengths and
3, 167 weaknesses of computer systems) is
essential to the development of
effective security, both for works
protected by copyright law and for
information in general. Such research
can progress only through the open
publication and exchange of complete
scientific results....
Inverted index Buckets Documents
Figure 4.1 Structure of a typical inverted index for information retrieval. Each entry
in the occurrence lists (buckets) is a pair of indices that identify the document and
the offset within the document where the term associated to the bucket appears. Docu-
ment (1) is an excerpt from Sen. Fritz Hollingss Consumer Broadband and Digital Televi-
sion Promotion Act, published at http://www.politechbot.com/docs/cbdtpa/
hollings.s2048.032102.html. Document (2) is an excerpt from Stallman (1997)
and (3) is an excerpt from the declaration submitted by ACM in the District Court of New
Jersey on the case of Edward Felten versus the Recording Industry Association of America
(http://www.acm.org/usacm/copyright/felten declaration.html).

user queries. A substring search, even when implemented using sophisticated algo-
rithms like sufx trees or sufx arrays (Manber and Myers 1990), is not adequate for
searching very large text collections. Many different methods of text retrieval have
been proposed in the literature, including early attempts such as clustering (Salton
1971) and the use of signature les (Faloutsos and Christodoulakis 1984). In practice,
inversion (Berry and Browne 1999; Witten et al. 1999) is the only effective technique
for dealing with very large sets of documents. The method relies on the construc-
tion of a data structure, called an inverted index, which associates lexical items to
their occurrences in the collection of documents. Lexical items in text retrieval are
called terms and may consist of words as well as expressions. The set of terms of
interest is called the vocabulary, denoted V . In its simplest form, an inverted index
is a dictionary where each key is a term V and the associated value b() is a
TEXT ANALYSIS 79

pointer to an additional intermediate data structure, called a bucket or posting list.

The bucket associated with a certain term is essentially a list of pointers marking
all the occurrences of in the text collection.
The general structure is illustrated in Figure 4.1. In the simplest case, the entries in
each bucket can simply consist of the document identier (DID), the ordinal number
of the document within the collection. However, in many cases it will be more useful to
have a separate entry for each occurrence of the term, where each entry consists of the
DID and the offset (in characters) of the terms occurrence within the document. The
latter approach has two advantages. First, recording all the occurrences of the term
allows us to present a user with a short context (i.e. a fragment of surrounding text) in
which the term occurs in the retrieved document. Second, the occurrence information
enables vicinity queries such as retrieving all documents where two assigned terms
are positionally close in the text.
The size of the inverted index is (|V |) and this data structure can typically be
stored in main memory.
It can be implemented using a hash table so that the expected access time is indepen-
dent of the size of the vocabulary. Buckets and the document repository must typically
be stored on disk. This poses a challenge during the construction phase of the inverted
index. If the buckets can be stored in the main memory, the construction algorithm is
trivial. It simply consists of parsing documents, extracting terms, inserting the term
in the inverted index if not present, and nally inserting the occurrence in the bucket.
However, if the buckets are on disk, this approach will generate a tremendous amount
of disk accesses making the algorithm impractical, since accessing a random byte of
information on disk can take up to 105 times longer than accessing a random byte
in main memory. Thus, for very large collections of text it is necessary to resort to
specialized secondary memory algorithms (Witten et al. 1999).
Searching for a single term in an indexed collection of documents is straight-
forward. First, we access the inverted index to obtain b(). Then we simply scan the
bucket pointed to by b() to obtain the list of occurrences. The case of Boolean queries
with multiple terms is only slightly more complicated. Suppose the query contains
literals associated with the presence or absence of k terms. The previous procedure
can be repeated separately for each term, yielding k lists of occurrences. These lists
are then combined by elementary set operations that correspond to the Boolean oper-
ators in the query: intersection for AND, union for OR, and complement for NOT
(Harman et al. 1992).

4.1.2 Compression techniques

The memory required for each pointer in the buckets can be reduced by sorting the
occurrences of each term by its DID. In this way, rather than storing each DID, it
sufces to store the sequence of differences between successive DIDs as a list of
gaps. For example, the sequence of DIDs
(14, 22, 38, 42, 66, 122, 131, 226, 312, 331)
80 LEXICAL PROCESSING

may be stored as a sequence of gaps

(14, 8, 16, 4, 24, 56, 9, 95, 86, 19).
The advantage is that frequent terms will produce many small gaps, and small inte-
gers can be encoded by short variable-length codewords. In particular, encod-
ing (Elias 1975) represents a gap g as the pair of bit strings (ug , bg ), where ug
is the unary code for the number log2 g + 1 (i.e. log2 g ones followed by a
zero) and bg is the standard binary code for the number g 2log2 g . For exam-
ple, the integer 7 would be represented as (110, 11) and 9 would be represented
as (1110, 001). In contrast, for a collection of n documents, storing an integer word
for a DID would require log2 n bits. Because terms are distributed according to
the Zipf distribution (see Chapter 1), on average this represents a signicant memory
saving (see Exercise 4.1 for details). Moffat and Zobel (1996) discuss this and other
index compression techniques, together with fast algorithms for query evaluation and
ranking.
In the case of large collections of Web pages, efcient indexing may be challenging
and traditional approaches may quickly became inadequate. The design of an efcient
distributed index for Web documents is discussed in Melnik et al. (2001).

4.2 Lexical Processing

4.2.1 Tokenization
Tokenization is the extraction of plain words and terms from a document, stripping out
administrative metadata and structural or formatting elements, e.g. removing HTML
tags from the HTML source le for a Web page. This operation needs to be performed
prior to indexing or before converting documents to vector representations that are
used for retrieval or categorization (see Section 4.3.1). Tokenization appears to be a
straightforward problem but in many practical situations the task can actually be quite
challenging.
The case of HTML documents is relatively easy. The simplest approach consists
of reducing the document to an unstructured representation, i.e. a plain sequence of
words with no particular relationship among them other than serial order. This can be
achieved by only retaining the text enclosed between <html> and </html> (see
Section 2.1), removing tags, and perhaps converting strings that encode international
characters to a standard representation (e.g. Unicode). The resulting unstructured rep-
resentation allows simple queries related to the presence of terms and to their posi-
tional vicinity. More generally, the additional information embedded in HTML can be
exploited to map documents to semi-structured representations. In this case, the rep-
resentation is sensitive to the presence of terms in different elements of the document
and allows more sophisticated queries like nd documents containing population
in a table header and famine in the title. Extracting semi-structured representations
from HTML documents is in principle very easy since tags provide all the necessary
TEXT ANALYSIS 81

structural information. However, the reader should be aware that, although HTML
has a clearly dened formal grammar, real world browsers do not strictly enforce
syntax correctness and, as a result, most Web pages fail to rigorously comply to the
HTML syntax. 1 Hence, an HTML parser must tolerate errors and include recovery
mechanisms to be of any practical usefulness. A public domain parser is distributed
with Libwww, the W3C Protocol Library http://www.w3.org/Library/.
An HTML parser written in Perl (with methods for text extraction) is available at
http://search.cpan.org/dist/HTML-Parser/.
Obviously, after plain text is extracted, punctuation and other special characters
need to be stripped off. In addition, the character case may be folded (e.g. to all
lowercase characters) to reduce the number of index terms.
Besides HTML, textual documents in the Web come in a large variety of formats.
Some formats are proprietary and undisclosed and extracting text from such le types
is severely limited by the fact that the associated formats have not been disclosed.
Other formats are publicly known, such as PostScript 2 or Portable Document Format
(PDF).
Tokenization of PostScript or PDF les can be difcult to handle because these are
not data formats but algorithmic descriptions of how the document should be rendered.
In particular, PostScript is an interpreted programming language and text rendering
is controlled by a set of show commands. Arguments to show commands are text
strings and two-dimensional coordinates. However, these strings are not necessarily
entire words. For example, in order to perform typographical operations such as kern-
ing, ligatures, or hyphenation, words are typically split into fragments and separate
show commands are issued for each of the fragments. Show commands do not need to
appear in reading order, so it is necessary to track the two-dimensional position of each
shownstring and use information about the font in order to correctly reconstruct word
boundaries. PostScript is almost always generated automatically by other programs,
such as typesetting systems and printer drivers, which further complicates matters
because different generators follow different conventions and approaches. In fact per-
fect conversion is not always possible. As an example of efforts in this area, the reader
can consult Neville-Manning and Reed (1996) for details on PreScript, a PostScript-to-
plain-text converter developed within the New Zealand Digital Library project (Witten
et al. 1996). Another converter is Pstotext, developed within the Virtual Paper project
(http://research.compaq.com/SRC/virtualpaper/home.html).
PDF is a binary format that is based on the same core imaging model as PostScript
but can contain additional pieces of information, including descriptive and adminis-
trative metadata, as well as structural elements, hyperlinks, and even sound or video.
In terms of delivered contents, PDF les are therefore much closer in structure to Web
pages than PostScript les are. PDF les can (and frequently do, in the case of digital
libraries) embed raster images of scanned textual documents. In order to extract text

1 See http://validator.w3.org/ for a public HTML validation service.

2 PostScript is a registered trademark of Adobe Systems Incorporated.
82 CONTENT-BASED RANKING

from raster images, it is necessary to resort to optical character recognition (OCR)

systems (Mori et al. 1992).

4.2.2 Text conation and vocabulary reduction

Often it is desirable to reduce all the morphological variants of a given word to a single
index term. For example, a document of interest might contain several occurrences of
words like sh, shes, sher, and shers but would not be retrieved by a query with
the keyword shing if the term shing never occurs in the text. Stemming consists
of reducing words to their root form (such as sh in our example), which becomes
the actual index term. Besides its effect on retrieval, stemming can be also useful to
reduce the size of the index.
One well known stemming algorithm for English was developed by Porter (1980)
as a simplication and systematization of previous work by Lovins (1968). It relies on
a preconstructed sufx list with associated rules. An example of a rewriting rule is if
the sufx of the word is IZATION and the prex contains at least one vowel followed
by a consonant, then replace the sufx with IZE. This would transform, for example,
the word BINARIZATION into the word BINARIZE. Porters stemmer applies rules
in ve consecutive steps. Source code in several languages and a detailed descrip-
tion of the algorithm are available at http://www.tartarus.org/martin/
PorterStemmer/.
Another technique that is commonly used for controlling the vocabulary is the
removal of stop words, i.e. common words such as articles, prepositions, and adverbs
that are not informative about the semantic content of a document (Fox 1992). Since
stop words are very common, removing them from the vocabulary can also signif-
icantly help in reducing the size of the index. In practice, stopwords may account
for a large percentage of text, up to 2030%. Naturally, removal of stop words also
improves computational efciency during retrieval.

4.3 Content-Based Ranking

A Boolean query to a Web search engine may return several thousands of matching
documents, but a typical user will only be able to examine a small fraction of these.
Ranking matching documents according to their relevance to the user is therefore
a fundamental problem. In this section we review some classic approaches that do
not exploit the hyperlinked structure of the Web. Ranking algorithms that take Web
topology into account will be discussed in Chapter 5.

4.3.1 The vector-space model

Text documents can be conveniently represented in a high-dimensional vector space
where terms are associated with vector components. More precisely, a text document
TEXT ANALYSIS 83

document d '

document d
x'
x

Figure 4.2 Cosine measure of document similarity.

d can be represented as a sequence of terms, d = ((1), (2), . . . , (|d|)), where

|d| is the length of the document and (t) V . A vector-space representation of
d is then dened as a real vector x R|V | , where each component xj is a statistic
related to the occurrence of the j th vocabulary entry in the document. The simplest
vector-based representation is Boolean, i.e. xj {0, 1} indicates the presence or the
absence of term j in the document being represented.
Vector-based representations are sometimes referred to as a bag of words, empha-
sizing that document vectors are invariant with respect to term permutations, since the
original word order (1), . . . , (|v|) is clearly lost. Representations of this kind are
appealing for their simplicity. Moreover, although they are necessarily lossy from an
information theoretic point of view, many text retrieval and categorization tasks can be
performed quite well in practice using the vector-space model. Note that typically the
total number of terms in a set of documents is much larger than the number of distinct
terms in any single document, |V |  |d|, so that vector-space representations tend
to be very sparse. This property can be advantageously exploited for both memory
storage and algorithm design.

4.3.2 Document similarity

We can dene similarity between two documents d and d  as a function s(d, d  ) R.
This function, among other things, allows us to rank documents with respect to a
query (by measuring the similarity between each document and the query). A classic
84 CONTENT-BASED RANKING

approach is based on the vector-space representation and the metric dened by the
cosine coefcient (Salton and McGill 1983). This measure is simply the cosine of the
angle formed by the vector-space representations of the two documents, x and x  (see
Figure 4.2),
xTx xTx
cos(x, x  ) = = , (4.1)
x x   xTx xTx
where the superscript T denotes the transpose operator and x T y indicates the dot
product or inner product between two vectors x, y Rm , dened as

. 
m
xTy = xi yi . (4.2)
i=1

Note that in the case of two sparse vectors x and y associated with two documents
d and d  , the above sum can be computed efciently in time (|d| + |d  |) (see
Exercise 4.2).
Several renements can be obtained by extending the Boolean vector model and
introducing real-valued weights associated with terms in a document. A more infor-
mative weighting scheme consists of counting the actual number of occurrences of
each term in the document. In this case xj N counts term occurrences in the cor-
responding document (see Figure 4.3). x may be multiplied by the constant 1/|d| to
obtain a vector of term frequencies (TF) within the document.
An important family of weighting schemes combines term frequencies (which are
relative to each document) with an absolute measure of term importance called
inverse document frequency (IDF). IDF decreases as the number of documents in
which the term occurs increases in a given collection. So terms that are globally rare
receive a higher weight.
Formally, let D = {d1 , . . . , dn } be a collection of documents and for each term j
let nij denote the number of occurrences of j in di and nj the number of documents
that contain j at least once. Then we dene
. nij
TFij = , (4.3)
|di |
. nj
IDFj = log . (4.4)
n
Here the logarithmic function is employed as a damping factor.
The TFIDF weight (Salton et al. 1983) of j in di can be computed as
xij = TFij IDFj (4.5)
or, alternatively, as
TFij IDFj
xij = . (4.6)
maxk di TFik maxk di IDFk
The IDF weighing is commonly used as an effective heuristic. A theoretical justi-
cation has been recently proposed by Papineni (2001), who proved that IDF is the
TEXT ANALYSIS 85
According to news.com, Apple has
warned one of its own dealers to stop
handing out a patch to allow DVD
burning with iDVD on non-Apple
hardware
According 1 .69 0 0
Act 0 0 1 .69
Apple 1 0 1 0
Computer 0 0 1 .69
Digital 0 0 1 .69
DVD 1 0 1 0
DVDs 0 0 0 0
Millennium 1 .69 0 .69
a 1 .69 0 0
allow 1 .69 0 0
burning 1 0 1 0
com 1 .69 0 0
customers 0 0 1 .69
dealers 1 .69 0 0
drivers 0 0 1 .69
external 0 0 1 .69
handing 1 .69 0 0
hardware 1 .69 0 0
has 1 0 1 0
iDVD 1 .69 0 0
invoked 0 0 1 .69
its 0 0 0 0
news 1 .69 0 0
non 1 .69 0 0
of 1 .69 0 0
on 1 0 1 0
one 1 .69 0 0
out 1 .69 0 0
own 1 .69 0 0
patch 1 .69 0 0
prevent 0 0 1 .69
stop 1 .69 0 0
to 3 0 1 0
the 0 0 1 .69
warned 1 .69 0 0
with 1 .69 0 0

Apple Computer has invoked the

Digital Millennium Copyright Act to
prevent its customers from burning
DVDs on external drives

Figure 4.3 Vector-space representations. For each of the two documents, the left vector counts
the number of occurrences of each term, while the right vector is based on TFIDF weights
(Equation 4.5).

optimal weight of a term with respect to the minimization of a distance function that
generalizes KullbackLeibler divergence or relative entropy (see Appendix A).

4.3.3 Retrieval and evaluation measures

Let us consider a collection of n documents D. Each document is represented by an
m-dimensional vector, where m = |V | and V is the set of terms that occurred in
86 PROBABILISTIC RETRIEVAL

the collection. Let q Rm denote the vector associated with a user query (terms that
are present in the query but not in V will be stripped off). Each document is then
assigned a score, relative to the query, by computing s(xi , q), i = 1, . . . , n. The set
R of retrieved documents that are presented to the user can be formed by collecting
the top-ranking documents according to the similarity measure. The quality of the
returned collection can be dened by comparing R to the set of documents R  that is
actually relevant to the query. 3
Two common metrics for comparing R and R  are precision and recall. Precision
is dened as the fraction of retrieved documents that are actually relevant. Recall
that is dened as the fraction of relevant documents that are retrieved by the system.
More precisely,
. |R R  | . |R R  |
= , = .
|R| |R  |
Note that in this context the ratio between relevant and irrelevant documents is
typically very small. For this reason, other common evaluation measures like accuracy
or error rate (see Section 4.6.5), where the denominator consists of |D|, would be
inadequate (it would sufce to retrieve nothing to get very high accuracy). Sometimes
precision and recall are combined into a single number called F measure dened as

. ( 2 + 1)
F = . (4.7)
2 +
Note that the F1 measure is the harmonic mean of precision and recall. If tends to
zero () the F measure tends to precision (recall).

4.4 Probabilistic Retrieval

Researchers in IR have long posed the question of how to dene an optimality criterion
for ranking retrieved documents in response to a specied user interest. Robertson
(1977) formulated an optimality principle called the probabilistic ranking principle
(PRP), stating that
if a reference retrieval systems response to each request is a ranking
of the documents in order of decreasing probability of relevance to the
user who submitted the request where the probabilities are estimated as
accurately as possible on the basis of whatever data have been made
available to the system for this purpose, the overall effectiveness of the
system to its user will be the best that is obtainable on the basis of those
data.

3 Of course this is possible only on controlled collections, such as those prepared for the Text REtrieval
Conference (TREC) (data, papers describing methodologies and description of evaluation measures and
assessment criteria are available at the TREC website http://trec.nist.gov/).
TEXT ANALYSIS 87

The PRP resembles the optimal Bayes decision rule for classication, a concept
that is well known for example in the pattern recognition community (Duda and Hart
1973). The Bayes optimal separation is obtained by ensuring that the posterior prob-
ability of the correct class (given the observed pattern) should be greater than the
posterior probability of any other class. PRP can be mathematically stated by intro-
ducing a Boolean variable R (relevance) and by dening a model for the conditional
probability P (R | d, q) of relevance of a document d for a given user need (for exam-
ple, expressed through a query q). Its justication follows from a decision-theoretic
argument as follows. Let c denote the cost of retrieving a relevant document and c
the cost of retrieving an irrelevant document, with c < c. Then in order to minimize
the overall cost, a document d should be retrieved next if
cP (R | d, q) + c(1 P (R | d, q))
cP (R | d  , q) + c(1 P (R | d  , q)) (4.8)
for every other document d  that has not yet been retrieved. But, since c < c, the
above condition is equivalent to
P (R | d, q)  P (R | d  , q),
that is, documents should be retrieved in order of decreasing probability of relevance.
In order to design a model for the probability of relevance some simplications are
needed. The simplest possible approach is called binary independence retrieval (BIR)
(Robertson and Sprck Jones 1976) as also used in the Bernoulli model for the Naive
Bayes classier, which we will discuss later, in Section 4.6.2. This model postulates
a form of conditional independence amongst terms. Following Fuhr (1992), let us
introduce the odds of relevance and apply Bayes theorem:
P (R = 1 | d, q) P (d | R = 1, q) P (R = 1 | q)
O(R | d, q) = = . (4.9)
P (R = 0 | d, q) P (d | R = 0, q) P (R = 0 | q)
Note that the last fraction is the odds of R given q, a constant quantity across
the collection of documents (it only depends on the query). The BIR assumption
concerns the rst fraction and was originally misidentied as a marginal form of term
independence (Cooper 1991). We can state it as
|V |

P (d | R = 1, q) xj P (j | R = 1, q) + (1 xj )(1 P (j | R = 1, q))
= .
P (d | R = 0, q) xj P (j | R = 0, q) + (1 xj )(1 P (j | R = 0, q))
j =1
(4.10)
The parameters to be estimated are therefore
.
j = P (j | R = 1, q)
.
and j = P (j | R = 0, q). If we further assume that j = j if j does not appear
in q we nally have
 j 1 j
O(R | d, q) = O(R | q) xj (1 xj ) , (4.11)
j 1 j
j :j q
88 LATENT SEMANTIC ANALYSIS

where the product only extends to indices j whose associated terms j appear in the
query. This can also be rewritten as
 1 j  j (1 j )
O(R | d, q) = O(R | q) , (4.12)
1 j j (1 j )
j q j :j q,
xj =1

where the last factor is the only part that depends on the document. The retrieval
status value (RSV) of a document is thus computed by taking the logarithm of the
last factor:
 j (1 j )
RSV(d) = log . (4.13)
j (1 j )
j :j dq

Documents are eventually ranked in reverse order of RSV.

4.5 Latent Semantic Analysis

The retrieval approach based on vector-space similarities can reach satisfactory recall
rates only if the terms in the query are actually present in the relevant documents.
Users, on the other hand, often formulate queries choosing terms that express concepts
related to the contents they want to retrieve. Natural language has a very rich expressive
power and even at the lexical level, the large variability due to synonymy and polysemy
causes serious problems to retrieval methods based on term matching. Synonymy
means that the same concept can be expressed using different sets of terms (e.g. bandit,
brigand, and thief). Synonymy negatively affects recall. Polysemy means that identical
terms can be used in very different semantic contexts. For example, in English, the
word bank refers to both a repository where important material is saved (as in blood
bank) and also to the slope beside a body of water (as in the bank of the Thames).
Polysemy negatively affects precision. Overall, synonymy and polysemy lead to a
complex relation between terms and concepts that cannot be captured through simple
matching.
Thus, although a query may conceptually be very close to a given set of documents,
its associated vector could be orthogonal or nearly orthogonal to those document
vectors, simply because the authors of the document and the user have a different
usage of language. Another more statistical way of thinking about this is to observe
that the number of terms that are present in a document is a rather small fraction of
the entire dictionary, reducing the likelihood that two documents use the same set of
words to express the same concept.
Latent semantic indexing (LSI) is a statistical technique that attempts to estimate
the hidden structure that generates terms given concepts. It uses a linear algebra tech-
nique known as singular value decomposition (SVD) to discover the most important
associative patterns between words and concepts. LSI is a data-driven method, where
a large collection of sentences or documents is employed to discover the statistically
most signicant co-occurrences of terms.
TEXT ANALYSIS 89

4.5.1 LSI and text documents

Let X denote a termdocument matrix, dened as

X = [x1 xn ]T . (4.14)

Each row in the matrix is simply the vector-space representation of a document. In

LSI we use occurrence counts as components. Each column contains the occurrences
of a term in each document in the data set. The LSI technique consists of computing
the SVD of X and setting to zero all the singular values except the largest K ones. In
practical applications K is often set to values around between 100 and 1000. This can
be seen as analogous to the PCA dimensionality reduction discussed in Section A.3:
documents are mapped to a lower-dimensional latent semantic space induced by
selecting directions of maximum covariance.
We will illustrate more in detail the LSI procedure through an example. Consider
the collection of 10 documents in Table 4.1. The rst ve documents are related to the
Linux operating system, and the remaining ve to news in the area of genomes.
For simplicity we focus only on frequent terms, in particular those occurring at
least twice. We remove common words as the and goes. The resulting term
document matrix XT is shown in the middle of Table 4.1 (we use the transpose
of X for the sake of space). Performing SVD in this example yields a singular values
matrix = diag(2.57, 2.49, 1.99, 1.9, 1.68, 1.53, 0.94, 0.66, 0.36, 0.10). By set-
ting = diag(2.57, 2.49, 0, 0, 0, 0, 0, 0, 0, 0), we obtain the reconstruction of X,
X = U SV T , shown at the bottom of table 4.1.
As we can see, the reconstructed matrix X is not as sparse as the original matrix X.
The new term weights account for word co-occurrences and appear to infer relations
among words that pertain to synonymy, at least in a loose sense. For example, a
query on the term Linux would now assign a relatively high score to documents 2
and 3, although they do not contain the search keyword at all. We can explain this
result arguing that SVD has inferred a link between Debian or Gentoo and Linux
through the co-occurrences of other words (they never occur together in our sample).
Similarly, we can note in Table 4.1 that both DNAand Dollyhave document vectors
closer to the document vector of genome.
When performing the numerical computation of the SVD of a very large document
matrix, it is important to take advantage of sparsity. A variety of well-known numerical
methods for computing the SVD of sparse matrices can be brought to bear (Berry
1992; Berry and Browne 1999; Letsche and Berry 1997).

4.5.2 Probabilistic LSA

The latent semantic indexing method described so far does not have a completely
sound probabilistic interpretation. As we have discussed, the approximation of X by
X obtained by setting to zero all but their rst K singular values is optimal in the sense
that X X2 is minimized (amongst all projections in latent spaces of dimension
90 LATENT SEMANTIC ANALYSIS

Table 4.1 Example application of LSI. Top: a collection of documents. Terms used in the
analysis are underlined. Center: the termdocument matrix XT . Bottom: the reconstructed
termdocument matrix X T after projecting on a subspace of dimension K = 2.

d1 : Indian government goes for open-source software

d2 : Debian 3.0 Woody released
d3 : Wine 2.0 released with xes for Gentoo 1.4 and Debian 3.0
d4 : gnuPOD released: iPod on Linux... with GPLed software
d5 : Gentoo servers running an open-source mySQL database
d6 : Dolly the sheep not totally identical clone
d7 : DNA news: introduced low-cost human genome DNA chip
d8 : Malaria-parasite genome database on the Web
d9 : UK sets up genome bank to protect rare sheep breeds
d10 : Dollys DNA Damaged

d1 d2 d3 d4 d5 d6 d7 d8 d9 d10

open-source 1 0 0 0 1 0 0 0 0 0
software 1 0 0 1 0 0 0 0 0 0
Linux 1 0 0 1 0 0 0 0 0 0
released 0 1 1 1 0 0 0 0 0 0
Debian 0 1 1 0 0 0 0 0 0 0
Gentoo 0 0 1 0 1 0 0 0 0 0
database 0 0 0 0 1 0 0 1 0 0
Dolly 0 0 0 0 0 1 0 0 0 1
sheep 0 0 0 0 0 1 0 0 1 0
genome 0 0 0 0 0 0 1 1 1 0
DNA 0 0 0 0 0 0 2 0 0 1

d1 d2 d3 d4 d5 d6 d7 d8 d9 d10

open-source 0.34 0.28 0.38 0.42 0.24 0.00 0.04 0.07 0.02 0.01
software 0.44 0.37 0.50 0.55 0.31 0.01 0.03 0.06 0.00 0.02
Linux 0.44 0.37 0.50 0.55 0.31 0.01 0.03 0.06 0.00 0.02
released 0.63 0.53 0.72 0.79 0.45 0.01 0.05 0.09 0.00 0.04
Debian 0.39 0.33 0.44 0.48 0.28 0.01 0.03 0.06 0.00 0.02
Gentoo 0.36 0.30 0.41 0.45 0.26 0.00 0.03 0.07 0.02 0.01
database 0.17 0.14 0.19 0.21 0.14 0.04 0.25 0.11 0.09 0.12
Dolly 0.01 0.01 0.01 0.02 0.03 0.08 0.45 0.13 0.14 0.21
sheep 0.00 0.00 0.00 0.01 0.03 0.06 0.34 0.10 0.11 0.16
genome 0.02 0.01 0.02 0.01 0.10 0.19 1.11 0.34 0.36 0.53
DNA 0.03 0.04 0.04 0.06 0.11 0.30 1.70 0.51 0.55 0.81
TEXT ANALYSIS 91
P (di ) P (zk | di ) P (j | zk) P (di | zk) P (zk) P (j | zk)

D Z D Z

(a) (b)

Figure 4.4 Bayesian networks describing the aspect model. Concepts (Z) are never observed
in the data but their knowledge would render words () and documents (D) independent. The
two networks are independence equivalent but use different parameterizations.

K). However, the L2 matrix norm is algebraically well suited for Gaussian variables
and is hard to justify its use in the case of the discrete space dened by vectors of
counts of words.
We now discuss a probabilistic approach for describing a latent semantic space
called the aspect model (Hofmann et al. 1999) or aggregate Markov model (Saul and
Pereira 1997). Let an event in this model be the occurrence of a term in a document
d. Let z {z1 , . . . , zK } denote a latent (hidden) variable associated with each event
according to the following generative scheme. First, select a document from a density
P (d). Second, select a latent concept z with probability P (z | d). Third, choose a
term to describe the concept linguistically, sampling from P ( | z), i.e. assume that
once the latent concept has been specied the document and the term are conditionally
independent, as shown by the Bayesian network representation of Figure 4.4. The
probability of each event (, d) is therefore

P (, d) = P (d) P ( | z)P (z | d). (4.15)
z

The aspect model has a direct interpretation in terms of dimensionality reduction.

Each document is uniquely determined by the mixing coordinates P (zk | d), k =
1, . . . , K, that form a nonnegative vector belonging to a probabilistic latent semantic
space. In other words, rather than being represented through terms, a document is
represented through latent variables that in turn are responsible for generating terms.
So if K  |V | dimensionality reduction occurs.
The analogy with respect to LSI can be established by introducing the matrices U ,
V , and whose elements are uik = P (di | zk ), vj,k = P (j | zk ) and kk = P (zk )
for i = 1, . . . , n, j = 1, . . . , m, and k = 1, . . . , K. In this way, all the n m
documentterm joint probabilities can be collected in the matrix

P = U V T . (4.16)

This matrix can be directly compared to the SVD representation used in traditional
LSI. Unlike X in LSI, P is a properly normalized probability distribution and its
entries cannot be negative. Moreover, the coordinates of terms in the latent space can
be interpreted as the distribution of terms, conditional on the hidden zk values.
Hofmann (2001) shows how to t the parameters of this model from data. In the
simplest case, they are estimated by maximum likelihood using the EM algorithm
(see Chapter 1) because of the latent variable Z. We assume the parameterization of
92 LATENT SEMANTIC ANALYSIS

Figure 4.4b. During the E step, the probability that the occurrence of j in di is due
to latent concept zk is computed as

P (zk | di , j ) P (wj | zk )P (di | zk )P (zk ). (4.17)

This can be seen as a special case of belief propagation (see Chapter 1) in the network
of Figure 4.4b. The M step simply consists of updating parameters as follows:


n
P (j | zk ) nij P (zk | di , j ), (4.18)
i=1
|V |

P (zk | di ) nij P (zk | di , j ), (4.19)
j =1

 |V |
n 
P (zk ) nij P (zk | di , j ). (4.20)
i=1 j =1

A possible renement of the above EM algorithm consists of replacing Equa-

tion (4.17) by

P (zk | di , j ) (P (j | zk )P (di | zk )) P (zk ). (4.21)

where is a positive parameter that is initialized to unity and exponentially decreased

in an outer loop that wraps around a full EM optimization (performed with constant ).
This outer loop iteration continues as long as predictive performance on a validation
set keeps improving. This renement can help to prevent overtting and has analogies
with the annealing process in physics. The resulting algorithm is called tempered EM
(TEM). Intuitively, smaller values of in Equation (4.21) increase the entropy of the
posterior distribution of Z performing a special form of regularization.
Once the model is trained, to dene a similarity measure based on PLSA, we need
to estimate which concepts are being expressed in the query q. As in LSI, documents
and queries that are not in the original matrix need to be folded in. In the case of PLSA
a point in the latent semantic space is a (posterior) distribution on Z, so folding in
consists of calculating P (zk | q) for k = 1, . . . , K. This is done by using the above
TEM procedure while keeping P (j | zk ) xed.
The original aspect model described above is not without its limitations. The fact
that there is a separate set of parameters for each document in the collection means
both that there are a very large number of parameters to be estimated and that the
model does not directly contain the concept of a new as yet unseen document
this can lead to overtting of the data. More recent work by Blei et al. (2002a) has
begun to address some of these issues by placing the aspect model in a more Bayesian
context.
TEXT ANALYSIS 93

4.6 Text Categorization

Text categorization consists of grouping textual documents into different xed classes
or categories. This may be useful, for example, to predict the topic of a Web page, or to
decide whether it is relevant with respect to the interests of a given user. This problem
is particularly well suited for a machine-learning solution. In the following we briey
review three of the most important and effective machine-learning algorithms that
are often applied to text categorization: k nearest neighbors (k-NN), Naive Bayes,
and support vector machines. Readers not familiar with general concepts in machine
learning are urged to read Section 1.5 before continuing this section.
Several additional machine-learning methods not covered here have been applied to
text categorization including decision trees (Lewis and Catlett 1994), neural networks
(Wiener et al. 1995), and symbolic approaches such as inductive logic programming
(Cohen 1995) and the induction of classication rules (Apt et al. 1994; Cohen 1996).
A review of these and other methods can be found in Sebastiani (2002).

4.6.1 k nearest neighbors

k-NN is a memory based classier that learns by simply storing all the training
instances. During prediction, k-NN rst measures the distances between a new point
x and all the training instances, returning the set N (x, D, k) of the k points that are
closest to x. For example, if training instances are represented by real-valued vectors
x, we could use Euclidean distance to measure the distance between x and all other
points in the training data, i.e. x xi 2 , i = 1, . . . , n. After calculating the distances,
the algorithm predicts a class label for x by a simple majority voting rule using the
labels in the elements of N (x, D, k), breaking ties arbitrarily. In spite of its apparent
simplicity, k-NN is known to perform well in many domains. An early result by Cover
and Hart (1967) shows that the asymptotic error rate of the 1-NN classier (as the
size of the training data set gets innitely large) is always less than twice the optimal
Bayes error (which is the lowest possible error rate achievable by any classier in a
particular feature space x).
In the case of text, majority voting can be replaced by a smoother metric where,
for each class c, a scoring function

s(c | x) = cos(x, x  ) (4.22)
x  Nc (x,D,k)

is computed through vector-space similarities between the new documents and the
subset of the k neighbors that belong to class c (Yang 1999), where Nc (x, D, k) is the
subset of N (x, D, k) containing only points of class c. Despite the simplicity of the
method, the performance of k-NN in text categorization is quite often satisfactory in
practice (Joachims 1998; Lam and Ho 1998; Yang 1999; Yang and Liu 1999). Han et
al. (2001) have proposed a variant of k-NN where the weights associated with features
are learned iteratively other statistically motivated techniques that extend the basic
k-NN classier are discussed in Hastie et al. (2001).
94 TEXT CATEGORIZATION

X1 X2 Xj Xk X |V |

aback abandon burn DVDs zoology

0 0 1 1 0

Figure 4.5 A Bayesian network for the Naive Bayes classier under the Bernoulli docu-
ment-based event model. The example document is the sentence burn DVDs and we will burn
you.

4.6.2 The Naive Bayes classier

This classier attempts to estimate the conditional probability of the class given the
document, namely P (c | d), for c = 1, . . . , S. Using Bayes theorem, we can write
this probability as
P (d | c, )P (c | )
P (c | d, ) = P (d | c, )P (c | ), (4.23)
P (d | )
where are the parameters of the model. In what follows we only include the depen-
dence on when necessary otherwise it can be implicitly assumed to be present.
Note that, since the classes are assumed to be mutually exclusive, we do not really
need to know  P (d), a term that can be thought of as a normalization factor guar-
anteeing that c P (c | d) = 1. The key idea behind the Naive Bayes classier is
the assumption that the terms in a document are conditionally independent given the
class. This assumption is clearly false in many if not most practical situations, but it
is often adequate to rst-order for the bag-of-words representation, where word order
in the document is not taken into account. Moreover, we should keep in mind that
we are interested in a good approximation of P (c | d), which does not necessarily
mean that we need a perfect model of P (d | c). In other words, we are interested in
discrimination among classes, not in a high quality generative model of the document
given the class (Ng and Jordan 2002). In practice, the classier is known to work
satisfactorily even when the conditional independence assumption is known not to
hold (Domingos and Pazzani 1997).
There is a subtle issue concerning the interpretation of the document in terms of
a probabilistic event (Lewis 1998; McCallum and Nigam 1998). If the document as
a whole is considered to be an event, then it should be naturally described by a bag
of words, and the words are the attributes of this event. In this case, each vocabulary
term is associated with a Bernoulli attribute whose realization is unity if the term
appears in the document, and zero otherwise. In addition to reducing documents to
bags of words, the Naive Bayes model postulates that binary attributes are mutually
independent given the class. This approach is substantially the same as the binary
TEXT ANALYSIS 95

independence retrieval model that was popular for probabilistic retrieval in the 1970s
(Robertson and Sprck Jones 1976) and has even more ancient origins (Maron 1961).
The conditional independence assumption in this model can be depicted graphically
using a Bayesian network such as that in Figure 4.5, suggesting that the class is the only
cause of the appearance of each word in a document. Under this model, generating a
document is like tossing |V | independent coins and the occurrence of each word in
the document is a Bernoulli event. Therefore we can rewrite the generative portion of
Equation (4.23) as
|V |

P (d | c, ) = xj P (j | c) + (1 xj )[1 P (j | c)], (4.24)
j =1

where xj = 1 [0] means that word j does [does not] occur in d and P (j | c) is
the probability of observing word j in documents of class c. Here represents the
set of probabilities (or parameters) P (j | c), which is the probability of the binary
event that word j is turned on within class c.
Alternatively, we may think of a document as a sequence of events W1 , . . . , W|d| .
Each observed Wt has a vocabulary entry (from 1 to |V |) as an admissible realization.
Note that the number of occurrences of each word, as well as the length of the
document, are taken into account under this interpretation. In addition, since the
document is a sequence, serial order among words should also be taken into account
when modeling P (W1 , . . . , W|d| | c). This could be done, for example, by using
a Markov chain. A simplifying assumption, however, is that word occurrences are
independent of their (relative) positions, given the class (Lewis 1992; Lewis and Gale
1994; Mitchell 1997). Equivalently, we assume that the bag-of-words representation
retains all the relevant information for assessing the probability of a document whose
class is known.
Under the word-based event model, generating a document is like throwing a die
with |V | faces |d| times, and the occurrence of each word in the document is a multi-
nomial event. Hence, the generative portion of the model is a multinomial distribution
|V |

P (d | ) = GP (|d|) P (j | c)nj , (4.25)
j =1

where nj is the number of occurrences of word j in d, and P (j | c) is the probabil-

ity that word j occurs at any position t [1, . . . , |d|]; because of the bag-of-words
assumption this does not depend  on t. Here the parameters are the set of probabil-
ities P (j | c), where now j =1 |V |P (j | c) = 1. McCallum and Nigam (1998)
found empirically that the multinomial model outperforms the Bernoulli model in
several evaluation benchmarks. Note that as in the case of the document-based event
model of Equation (4.24), the bag-of-words assumption results in a factorization
of P (W1 , . . . , W|d| | c) explaining, perhaps, why Naive Bayes is often used in the
literature to refer to both event models.
96 TEXT CATEGORIZATION

The normalization factor is the multinomial coefcient

|d|!
G=  .
j nj !

Neither P (|d|) nor G are needed for classication, since |d|, the number of words or
terms in a document, is assumed to be independent of the class. This last assumption
can be removed and P (|d| | c) explicitly modeled. Models of document length
(e.g. based on Poisson distributions) have been used for example in the context of
probabilistic retrieval (Robertson and Walker 1994). Note that in the case of the
Bernoulli model there are 2|V | possible different documents, while in the case of the
multinomial model there is an innite (but countable) number of different documents.
An additional model that may be developed and that lies somewhat in between
the Bernoulli and the multinomial models consists of keeping the document-based
event model but extending Bernoulli distributions to integer distributions, such as
the Poisson (Lewis 1998). Finally, extensions of the basic Naive Bayes approach
that allow limited dependencies among features have been proposed (Friedman and
Goldszmidt 1996; Pazzani 1996). However, these models are characterized by a larger
set of parameters and may overt the data (Koller and Sahami 1997).
Learning a Naive Bayes classier consists of estimating the parameters from the
available data. We will assume that the training data set consists of a collection of
labeled documents {(di , ci ), i = 1, . . . , n}. In the Bernoulli model, the parameters
include c,j = P (j | c), j = 1, . . . , |V |, c = 1, . . . , K. These are estimated as
normalized sufcient statistics
1 
n
c,j = xij , (4.26)
Nc
i:ci =c

where Nc = |{i : ci = c}| and xij = 1 if j occurs in di . Additional parameters are

the class prior probabilities c = P (c), which are estimated as
Nc
c = . (4.27)
n
Note that the estimates above correspond to ML estimates of the parameters. In
Chapter 1 we discussed how we can also derive Bayesian parameter estimates that
combine the observed data with prior knowledge. This can be achieved by using (for
example) Dirichlet priors for the Bernoulli likelihood model. Bayesian estimates can
be quite useful when estimating parameters from sparse data. For words in documents,
we might have a weak noninformative prior that any word can in theory appear in
documents from any class. This would avoid problems in making predictions on future
documents when such a word shows up in some class c where it did not appear in the
training data. A model based on ML parameter estimates would assign a probability
of zero to that document, irrespective of how well the other words in the document
matched class c. On the other hand, a model based on Bayesian estimates would assign
that word-class combination a low nonzero probability and still allow the other words
TEXT ANALYSIS 97

to play a role in determining the nal class prediction for the document. We might also
have more informative priors available, e.g. a Dirichlet prior on the distribution of
words in each class that reect word distributions as estimated from previous studies,
or that reect an experts belief in what words are likely to appear in each class.
In the case of the multinomial model of Equation (4.25), the generative
 parameters
are c,j = P (j | c). Note that these parameters must satisfy j c,j = 1 for each
class c. To estimate these parameters it is common practice to introduce Dirichlet
priors (see Chapter 1 for details). The resulting estimation equations are derived as
follows. In the case of the distributions of terms given the class, we introduce a
Dirichlet prior with hyperparameters qj and , resulting in the estimation formula

qj + ni:ci =c nij
c,j = |V |  , (4.28)
+ l=1 i:ci =c nil

where nij is the number of occurrences of j in di . A simple noninformative prior

assigns qj = 1/|V | and = |V |. Intuitively, this prior corresponds to the assumption
that we have observed each word exactly once in one document of each class. This
method (also known as Laplace smoothing) prevents the problem of estimating a null
value for a parameter if a certain term j never occurs in documents of a given class
c in the training set (without the smoothing, the probability that a new document
containing j belong to c would be zero). Similarly, the estimation formula for the
(unconditional) class probabilities is
qc  + Nc
c = . (4.29)
 + n
Also in this case we could assume a noninformative Dirichlet hyperparameters qc =
1/K and  = K.

4.6.3 Support vector classiers

Support vector machines (SVMs) were introduced in Cortes and Vapnik (1995),
extending earlier seminal work by Vapnik on statistical learning theory (Vapnik 1982).
The basic underlying idea, often referred to as structural risk minimization is closely
related to the theory of regularization but also to Bayesian approaches to learning
(Evgeniou et al. 2000) and is essentially guided by the principle that the hypothesis
that explains a nite set of examples should be searched in an appropriately small
hypothesis space.
SVMs are particularly well suited to deal with high-dimensional data such as vector-
space representations of text documents. In their standard formulation, they deal with
binary classication problems where the number of classes is restricted to two (see
later for a discussion on multiclass SVM).
Consider a training set D = {(xi , yi ), i = 1, . . . , n} with xi Rm , and where yi
{1, 1} is an integer that species whether xi is a positive or a negative example. A
98 TEXT CATEGORIZATION

linear discriminant classier is then dened by introducing the separating hyperplane

{x : f (x) = wT x + w0 = 0}, (4.30)
where w Rm and w0 R are adjustable coefcients that play the role of model
parameters. A binary classication function h : Rm {0, 1} can be obtained by tak-
ing the sign of f (x), i.e.

1, iff (x) > 0,
h(x) = (4.31)
0, otherwise.

Learning in this class of models consists of determining w and w0 from data. The
training examples are said to be linearly separable if there exists a hyperplane whose
associated classication function is consistent with all the labels, i.e. if yi f (xi ) > 0
for each i = 1, . . . , n. Under this hypothesis, Rosenblatt (1958) proved that the
following simple iterative algorithm terminates and returns a separating hyperplane:
Perceptron(D)
1 w0
2 w0 0
3 repeat
4 e0
5 for i 1, . . . , n
6 do s sgn(yi (w T xi + w0 ))
7 if s < 0
8 then w w + yi xi
9 w0 w0 + yi
10 e e+1
11 until e = 0
12 return (w, w0 ).

It can be shown that a sufcient condition for D to be linearly separable is that the
number of training examples n = |D| is less than or equal to m + 1 (see Exercise 4.4).
This is particularly likely to be true in the case of text categorization, where the
vocabulary typically includes several thousands of terms, and is often larger than the
number of available training examples n (see Exercise 4.5 for details).
Unfortunately, learning with the Perceptron algorithm offers little defense against
overtting. A thorough explanation of this problem requires concepts from statistical
learning theory that are beyond the scope of this book. To gain an intuition of why
this is the case, consider the scenario in Figure 4.6. Here we suppose that positive
and negative examples are generated by two Gaussian distributions (see Appendix A)
with the same covariance matrix and that positive and negative points are generated
with the same probability. In such a setting, the optimal (Bayes) decision boundary
is the one that minimizes the posterior probability that a new point is misclassied
and, as it turns out, this boundary is the hyperplane that is orthogonal to the segment
TEXT ANALYSIS 99

+
+
-
- +
+ +
-
- - +
- +
-
+ +

-
+
-
-
-
- - +

Figure 4.6 Alternative linear decision boundaries for a binary

classication problem (see text for explanation).

connecting the centers of mass of the two distributions (dotted line in Figure 4.6).
Clearly, a random hyperplane that just happens to separate training points (dashed
line) can be substantially far away from the optimal separation boundary, leading
to poor generalization to new data. The difculty grows with the dimensionality
of the input space m since for a xed n the set of separating hyperplanes grows
exponentially with m (a problem known as the curse of dimensionality). Remember
that in the case of text categorization m may be signicantly large (several thou-
sands).
The statistical learning theory developed by Vapnik (1998) shows that we can dene
an optimal separating hyperplane (relative to the training set) having two important
properties: it is unique for each linearly separable data set, and its associated risk of
overtting is smaller than for any other separating hyperplane. We dene the margin
M of the classier to be the distance between the separating hyperplane and the
closest training examples. The optimal separating hyperplane is then the one having
maximum margin (see Figure 4.7). Going back to Figure 4.6, the theory suggests that
the risk of overtting for the maximum margin hyperplane (solid line) is smaller than
for the dashed hyperplane. Indeed, in our example the maximum margin hyperplane
is signicantly closer to the Bayes optimal decision boundary.
In order to compute the maximum margin hyperplane, we begin by observing that
the distance of a point x from the separating hyperplane is

1
(w T x + w0 )
w
100 TEXT CATEGORIZATION

xTw + 0=0

+
+

+
-
- +
+
-
- - +
- +
-
+

+
-
+
-

-
- xTw + 0=M
- - + +

2
xTw + 0 = M
=2M
|| w ||
Figure 4.7 Illustration of the optimal separating hyperplane and margin.
Circled points are support vectors.

(see Figure 4.7). Thus, the optimal hyperplane can be obtained by solving the con-
strained optimization problem
1
max M subject to yi (w T xi + w0 )  M, i = 1, . . . , n, (4.32)
w,w0 w
where the constraints require that each training point should lie in the correct semis-
pace and at a distance not less than M from the separating hyperplane. Note that
although (w, w0 ) comprise n + 1 real numbers, there are only n degrees of freedom
since multiplying w and w0 by a scalar constant does not change the hyperplane. Thus
we can arbitrarily set w = 1/M and rewrite the optimization problem (4.32) as

min w subject to yi (w T xi + w0 )  1, i = 1, . . . , n. (4.33)

w,w0

The above problem can be transformed to its dual by rst introducing the vector of
Lagrangian multipliers Rn and writing the Lagrangian function


n
L(D) = 21 w2 + i [yi (w T x + w0 ) 1] (4.34)
i=1
TEXT ANALYSIS 101

and subsequently setting to zero the derivatives of (4.34) with respect to w, w0 ,

obtaining

n
max 21 T  + i subject to i  0, i = 1, . . . n, (4.35)

i=1

where  is an n n matrix with ij = yi yj xiT xj . This is a quadratic programming

(QP) problem that can be solved, in principle, using standard optimization pack-
ages. For each training example i, the solution must satisfy the KarushKuhnTucker
condition,
i [yi (w T x + w0 ) 1] = 0, (4.36)
and therefore either i = 0 or yi (w T x + w0 ) = 1. In other words, if i > 0 then the
distance of point xi from the separating hyperplane must be exactly M (see Figure 4.7).
Points with associated i > 0 are called support vectors. The decision function
h(x) can be computed via Equation (4.30) or, equivalently, from the following dual
form:
n
f (x) = yi i x T xi . (4.37)
i=1
We remark that in the case of text categorization it may be very important to exploit
the sparseness of the data vectors while computing dot products in Equations (4.35)
and (4.37).
If the training data are not linearly separable, then this analysis can be generalized by
introducing m nonnegative slack variables i and replacing the optimization problem
in Equation (4.33) with

n
yi (w T xi + w0 )  1 i , i = 1, . . . , n,
min w + C i subject to
w,w0
i=1
i  0, i = 1, . . . , n,
(4.38)
where the constant C controls the cost associated with misclassications.
This problem can also be dualized in the form

n
max 21 T  + i subject to 0
i
C, i = 1, . . . n, (4.39)

i=1

yielding another QP problem. The classier obtained in this way is commonly referred
to as a support vector machine (SVM).
Note that solving the QP problem using standard optimization packages would take
time O(n3 ) (assuming the number of support vector grows linearly with the number of
training example). This time complexity is a practical drawback for SVMs. However,
several approaches that can reduce complexity substantially have been proposed in
the literature (Joachims 1999a; Platt 1999).
If the data are considerably nonlinearly separable then an SVM classier will have
a low accuracy, i.e. even the best linear hyperplane may be quite inferior in terms
102 TEXT CATEGORIZATION

of prediction accuracy relative to a good nonlinear decision boundary. The methods

developed in this section can be further extended to accommodate nonlinear separation
by using kernel functions that map points x Rm into a higher-dimensional (possibly
innite-dimensional) space called the feature space, where data are linearly separable.
Details on kernel methods can be found in Cristianini and Shawe-Taylor (2000);
Schoelkopf and Smola (2002) and further details on the application of SVM to text
categorization can be found in Joachims (2002).
The SVM presented here can only deal with binary classication problems. For
more than two classes, a standard approach consists of reducing a multiclass problem
into several binary sub-problems, coding each class with a binary string. The simplest
(redundant) coding strategy is often referred to as one shotor one vs alland consists
of dening as many dichotomies of the instance space as there are classes, where
each class is considered as positive in one and only one dichotomy. A more general
reduction scheme is the information theoretic method based on error-correcting output
codes, introduced by Dietterich and Bakiri (1995) and more recently extended in
Allwein et al. (2000). Typically, these binary classiers are trained independently
but recent work (Crammer and Singer 2000; Guermeur et al. 2000) also considers
the case where classiers are trained simultaneously. The extraction of conditional
probabilities from multiclass SVM is studied in Passerini et al. (2002). Note that the
use of error-correcting codes is not limited to inherently binary classiers like SVMs.
For example, Ghani (2000) apply error-correcting codes in conjunction with the Naive
Bayes classier, reporting good results in text categorization tasks with many classes.

4.6.4 Feature selection

Even methods like SVMs that are especially well suited for dealing with high dimen-
sional data (such as vectorial representations of text) can suffer if many terms are
irrelevant for class discrimination. Feature selection is a dimensionality reduction
technique that attempts to limit overtting by identifying irrelevant components of
data points and has been extensively studied in pattern recognition (Kittler 1986) and
in machine learning (Kira and Rendell 1992; Koller and Sahami 1996; Langley 1994;
Liu and Motoda 1998).
Methods essentially fall into one of two categories: lters and wrappers. Filters
attempt to determine which features are relevant before learning actually takes place.
Wrapper methods, on the other hand, are based on estimates of the generalization
error computed by running a specic learning algorithm and searching for relevant
features by minimizing the estimated error (Kohavi and John 1997).Although wrapper
methods are in principle more powerful, in practice their usage is often hindered by
the computational cost. Moreover, they can overt the data if used in conjunction with
classiers having high capacity.
Information theoretic approaches are known to perform well for lter methods. In
particular, information gain is a popular metric in several machine-learning contexts,
including the induction of decision trees (Quinlan 1986). It measures the information
TEXT ANALYSIS 103

about the class that is provided by the observation of each term. More precisely,
let us denote by Wj an indicator variable such that Wj = 1 means that j appears
in a certain document. The information gain of Wj is then the mutual information
I(C, Wj ) between the class C and Wj (see Appendix A for a review of the main
concepts in information theory). This is also the difference between the marginal
entropy of the class H (C) and the conditional entropy H (C | Wj ) of the class, given
Wj :

K 
1
P (c, j )
G(Wj ) = H (C) H(C | Wj ) = P (c, j ) log . (4.40)
P (c)P (j )
c=1 j =0

Note that if C and Wj are independent, the mutual information is zero. Filtering
index terms simply consists of sorting terms by information gain and keeping only the
k terms with highest gain. Information gain has been used extensively for text catego-
rization (Craven et al. 2000; Joachims 1997; Lewis and Ringuette 1994; McCallum
and Nigam 1998; Yang 1999) and has been generally found to improve classication
performance compared to using all words.
One limitation of information gain is that relevance assessment is done separately
for each attribute and the effect of co-occurrences is not taken into account. However,
terms that individually bring little information about the class might bring signicant
information when considered together. In the framework proposed by Koller and
Sahami (1996), whole sets of features are tested for relevance about the class. Let
x denote a point in the complete feature space (i.e. the entire vocabulary V in the
case of text documents) and xG be the projection of x into a subset G V . In order
to evaluate the quality of G to represent the class, we measure the distance between
P (c | x) and P (c | xG ) using the average relative entropy:
 P (c | x)
G = P (x)P (c | x) log . (4.41)
x
P (c | xG )

The optimal set of features should yield a small G . Clearly this setup is only
theoretical, since Equation (4.41) is computationally intractable and the distributions
involved are hard to estimate accurately. Koller and Sahami (1996) use the notion of
a Markov blanket to reduce complexity. A set of features M V (with Wj not in M)
is a Markov blanket for Wj if Wj is conditionally independent of all the features in
V \ (M {Wj }) given M (see also the theory of Bayesian networks in Chapter 1).
In other words, once the features in M are known, Wj and the remaining features
are independent. Thus the class C is also conditionally independent of Wj given
M and, as a result, if G contains a Markov blanket for Wj then Gj = G , where
Gj = G \ {Wj }. The feature selection algorithm can then proceed by removing those
features for which a Markov blanket can be found. Koller and Sahami (1996) prove
that a greedy approach where features are removed iteratively is correct in the sense
that a feature deemed irrelevant and removed during a particular iteration cannot
became relevant again after some other features are later removed. Moreover, since
104 TEXT CATEGORIZATION

nding a Markov blanket may be computationally very hard and an exact blanket
may not exist, they suggest the following approximate algorithm:
ApproxMarkovBlanket(D, V , k, n )
1 GV
2 repeat
3 for Wj G
4 do for Wi G \ {Wj }
cov[W ,Wj ]
5 do ij var[W ]i var[W
i i]
6 Mj k features having highest ij
7 pj P (c | XMj = xMj , Wj = xj )
8 pj P (c | XMj = xMj )
9 D(xMj , xj ) H (pj , pj )

10 (Wj |Mj ) xM ,xj P (xMj , xj )D(xMj , xj )
j
11 j  arg minj (Wj |Mj )
12 G G \ {Wj  }
13 until |G| = n
14 return G

At each step, the algorithm computes for each feature Wj the set Mj (G \ {Fj })
containing the k features that have the highest correlation with Wj , where k is a
parameter of the algorithm and where correlation is measured by the Pearson corre-
lation coefcient in line 5. Then, in line 8, the quantity (Wj | Mj ) is computed as
the average cross entropy between the conditional distributions of the class that result
from the inclusion and the exclusion of feature Wj . This quantity is clearly zero if
Mj is a Markov blanket. Thus, picking j  that minimizes it (line 9) selects a feature
for which an approximate Markov blanket exists. Such a feature is removed and the
process iterated until n features remain in G.
Several other lter approaches have been proposed in the context of text catego-
rization, including the use of minimum description length (Lang 1995), and symbolic
rule learning (Raskinis and Ganascia 1996). A comparison of alternative techniques
is reported in Yang (1999).

4.6.5 Measures of performance

The performance of a hypothesis function h() with respect to the true classication
function f () can be measured by comparing h() and f () on a set of documents Dt
whose class is known (test set). In the case of two categories, the hypothesis can be
completely characterized by the confusion matrix
Predicted Actual category
Category +
TN FN
+ FP TP
TEXT ANALYSIS 105

where TP, TN, FP, and FN mean true positives, true negatives, false positives, and
false negatives, respectively. In the case of balanced domains (i.e. where the uncon-
ditional probabilities of the classes are roughly the same) accuracy A is often used
to characterize performance. Under the 0-1 loss (see Section 1.5 for a discussion),
accuracy is dened as
TN + TP
A= . (4.42)
|Dt |
Classication error is simply E = 1 A. If the domain is unbalanced, measures such
as precision and recall are more appropriate. Assuming (without losing generality)
that the number of positive documents is much smaller than the number of negative
ones, precision is dened as
TP
= (4.43)
TP + FP
and recall is dened as
TP
= . (4.44)
TP + FN
A complementary measure that is sometimes used is specicity
TN
= . (4.45)
TN + FN
As in retrieval, there is clearly a trade-off between false positives and false nega-
tives. For example, when using a probabilistic classier like Naive Bayes we might
introduce a decision function that assigns a document the class + if and only if
P (c | d) > t. In so doing, small values of the threshold t yield higher recall and
larger values yield higher precision. Something similar can be constructed for an
SVM classier using a threshold on the distance between the points and the sep-
arating hyperplane. Often this trade-off is visualized on a parametric plot where
precision and recall values, (t) and (t) are evaluated for different values of
the threshold (see Figure 4.9 for some examples). Sometimes these plots are also
called ROC curves (from Receiver Operating Characteristic) neglecting the caveat
that the original name was coined in clinical research for diagrams plotting speci-
city versus sensitivity (an alias for precision). A very common measure of perfor-
mance that synthesizes a precision-recall curve is the breakeven point, dened as
the best 4 point where (t  ) = (t  ) and can be seen as an alternative to the F
measure (discussed earlier in Section 4.3.3) for reducing performance to a single
number.
In the case of multiple categories we may dene precision and recall separately
for each category c, treating the remaining classes as a single negative class (see
Table 4.2 for an example). Interestingly, this approach also makes sense in domains
where the same document may belong to more than one category. In the case of mul-
tiple categories, a single estimate for precision and a single estimate for recall can
4 Since in general there may be more than one value of the threshold at which (t) = (t), we take
the value t  that maximizes (t  ) = (t  ).
106 TEXT CATEGORIZATION

be obtained by averaging results over classes. Averages, however, can be obtained in

two ways. When microaveraging, correct classications are rst summed individu-
ally:
K
TPc
= c=1

K
, (4.46)
c=1 TPc
K
c=1 TPc + FPc
= K . (4.47)
c=1 TPc + FNc

When macroaveraging, precision and recall are averaged over categories:

1 
K
M = c , (4.48)
K
c=1

1 
K
M = c . (4.49)
K
c=1

Compared to microaverages, macroaverages tend to assign a higher weight to classes

having a smaller number of documents.

4.6.6 Applications
Up to this point we have largely discussed the classication of generic text documents
as represented by a bag-of-words vector representation. From this viewpoint it did
not really matter whether the bag of words represented a technical article, an email, or
a Web page. In practice, however, when applying text classication to Web documents
there are several domain-specic aspects of the Web that can be leveraged to improved
classication performance. In this section we review some typical examples that
demonstrate how ideas from text classication can be applied and extended in a Web
context.

Classication of Web pages

Craven et al. (2000) have tackled the difcult task of extracting information from
Web documents in order to automatically generate knowledge bases (KB). In their
Web KB system, information extraction is carried out by rst training machine-
learning subsystems to make predictions about classes and relations, and then using
the trained subsystems to populate an actual KB starting from data collected from
the Web (see Chapter 6 for details on gathering collections of Web pages). Users
are expected to provide as input an ontology (i.e. a formal description of classes
and relations of interest) and training examples of actual instances of classes and
relations. A sample ontology for modeling academic websites is shown in Figure 4.8.
TEXT ANALYSIS 107
HomePage
Title

Department Other Activity Person

Location Name
Members Department
Email
Phone

Course Project Faculty Staff Grad Student

Instructor Members Position Position Advisor
TA P. Investigator Courses Courses
Projects Publications
Students
Publications

Figure 4.8 A sample ontology for representing knowledge about academic websites. The
diagram indicates the class hierarchy and the main attributes of each class. Figure adapted
from Craven et al. (2000).

Knowledge extraction consists essentially of

(1) assigning a new Web page to one node of the class hierarchy, and
(2) lling in the class attributes by extracting the relevant information from the
document.
For example, once the system has gathered a home page of a professor at your uni-
versity, it should rst assign the page to the correct class (faculty) and then it should
determine the value of attributes such as the email address, the courses taught, and
the published papers. The rst problem can be conveniently formulated as a text cat-
egorization task. In particular, Craven et al. (2000) study a Naive Bayes classier to
discriminate between the seven categories that appear as the leaves of the hierarchy
shown in Figure 4.8. Their results are shown in Table 4.2 in the form of a confusion
matrix. Each entry kl for k, l = 1, . . . , K is the number of documents in class l that
are assigned to class k, i.e. nondiagonal entries represent misclassications. Preci-
sion for each class k is obtained by dividing kk (the number of documents in class
kthat are correctly classied) by the total number of documents assigned to class k,
K
l=1 kl . Similarly the recall for class l is obtained
 dividing kl by the number of
documents that actually belong to class l, i.e. K k=1 kl .

Classication of news stories

The Reuters-21578 collection is perhaps the most widely use benchmark for text
categorization systems (Lewis 1997). It consists of 21 578 stories that appeared in
1987 in the Reuters newswire, assembled and manually labeled by personnel from
Reuters Ltd and Carnegie Group, Inc. There are 672 categories and each story may
belong to more than one category. 148 of the categories contain at least 20 docu-
108 TEXT CATEGORIZATION

Table 4.2 Experimental results obtained in Craven et al. (2000) using the
Naive Bayes classier on the Web KB domain.

Actual category
Predicted

category Cou Stu Fac Sta Pro Dep Oth Precision

Cou 202 17 0 0 1 0 552 26.2

Stu 0 421 14 17 2 0 519 43.3
Fac 5 56 118 16 3 0 264 17.9
Sta 0 15 1 4 0 0 45 6.2
Pro 8 9 10 5 62 0 384 13.0
Dep 10 8 3 1 5 4 209 1.7
Oth 19 32 7 3 12 0 1064 93.6

Recall 82.8 75.4 77.1 8.7 72.9 100.0 35.0

Table 4.3 Results reported by Joachims (1998) and Weiss et al. (1999) (last row) on 90 classes
of the Reuters-21578 collection. Performance is measured by microaveraged breakeven point.

Performance
Prediction method breakeven (%)

Naive Bayes (linear) 73.4

Rocchio (linear) 78.7
Decision tree C4.5 78.9
k-NN 82.0
Rule induction 82.0
Support vector (RBF) 86.3

Voting multiple decision trees 87.8

ments. The data set has been split into training and test data according to several
alternative conventions (see Lewis (1997) and Sebastiani (2002) for discussion). In
the so-called ModApte split, 9603 documents are reserved for training and 3299 for
testing (the rest being discarded). Ninety categories possess at least one training and
one test example. In this setting, Joachims (1998) experimented with several alter-
native classiers, including those described in this chapter (see Table 4.3). Note that
the support vector classier used in Joachims (1998) applies a radial basis func-
tion (RBF) kernel to learn nonlinear separation surfaces (see Schoelkopf and Smola
(2002) for a thorough discussion of kernel methods). It is worth noting that the sim-
ple k-NN classier achieves relatively good performance and outperforms Naive
Bayes in this problem. On the same data set, Weiss et al. (1999) have reported better
results using multiple decision trees trained with AdaBoost (Freund and Schapire
1996; Schapire and Freund 2000) specic results are provided in the last row of
Table 4.3.
TEXT ANALYSIS 109

Precision Acquisitions Precision Corn

100 100
90
90
80
80 70
60
70
50
60 40
Recall Recall
60 70 80 90 100 20 40 60 80 100

Precision Crude Precision Earnings

100 100
90 95
90
80
85
70
80
60 75
Recall Recall
40 50 60 70 80 90 100 75 80 85 90 95 100

Precision Grain Precision Interest

100 100
90 90
80 80
70 70
60 60
50 50
Recall Recall
50 60 70 80 90 100 40 60 80 100

Precision Moneyfx Precision Ship

100 100
90 90
80 80
70
70
60
60
50
50
40
Recall Recall
40 60 80 100 40 60 80 100

Precision Trade Precision Wheat

100 100
90 90
80 80
70 70
60
60
50
50
40
Recall Recall
50 60 70 80 90 100 40 60 80 100

Figure 4.9 Precision-recall curves obtained by training an SVM classier on the Reu-
ters-21578 data set (ModApte split, 10 most frequent classes), where each plot is for one
of the 10 classes.
110 TEXT CATEGORIZATION

Table 4.4 Experimental results obtained by Sahami et al. (1998)

in a junk email ltering problem.

Junk Not junk



Features Precision (%) Recall (%) Precision (%) Recall (%)

Words 97.1 94.3 87.7 93.4

Terms 97.6 94.3 87.8 93.4
Terms and features 100.0 98.3 96.2 100.0

A simplied problem in the Reuters data set is obtained by removing all but the 10
most populated categories. In this setting, Dumais et al. (1998) report a comparison
of several alternative learners obtaining their best performance (92% microaveraged
breakeven point) with support vector machines. Figure 4.9 shows precision-recall
curves we have obtained on this data set using an SVM classier.

Email and news ltering

Sahami et al. (1998) have applied the Naive Bayes classier to the problem of junk
email ltering. This application is an example of a problem where hand-programming
for data preprocessing (before learning) is particularly important. In this domain, the
bag-of-words representation removes important information related to word prox-
imity. Better features can be obtained by assessing the occurrence of terms such as
condential message, urgent and personal, or hot to make $, or perhaps by having
features associated with regular expressions that detect unusual usage of punctuation.
Sahami et al. (1998) manually identied a set of such terms and, in addition, enriched
the textual representation with other features such as the TLD of the senders address,
attachments, timestamps, etc.
Feature selection in Sahami et al. (1998) was performed by eliminating rare words
and then retaining the 500 most informative features as determined by mutual infor-
mation. Results obtained on a data set of 1538 training messages and 251 test mes-
sages are reported in Table 4.4. The data set contained 1578 junk messages and
211 legitimate messages. Since the loss of false positives is clearly higher than
the loss associated with false negatives, Sahami et al. (1998) proposed to classify
a message as junk only if the probability predicted by Naive Bayes was greater than
99.9% (an empirically determined threshold). Good results with Naive Bayes have
also been obtained by Androutsopoulos et al. (2000) on a publicly available data set
(http://www.aueb.gr/users/ion/publications.html). A compari-
son of alternative learning methods, including SVM, is reported in Drucker et al.
(1999).
NewsWeeder is a personalized ltering system for Usenet articles which uses rel-
evance feedback to score news according to user interests (Lang 1995). The data set
collected for training the NewsWeeder system contains 19 997 articles evenly sampled
from 20 newsgroups and has been widely used as a benchmark for text categorization
TEXT ANALYSIS 111









+
+

+

+
+


+


+

Figure 4.10 How unlabeled data (circles) may help categorization in

the case of large margin classiers.

systems (the task being to predict which newsgroup a message was posted to). The
data set of 20 newsgroups is available at http://www.ai.mit.edu/people/
jrennie/20Newsgroups/.

4.6.7 Supervised learning with unlabeled data

The classiers studied so far learn from labeled examples of text documents, i.e. the
category of the examples must be known. Labels must be assigned by human experts
and the process is clearly costly and time consuming. On the other hand, unlabeled
documents are abundantly available in many domains, particularly in the case of Web
documents. The available data can be written as D D  , where D = (di , f (di )),
i = 1, . . . n, is the labeled data set and D  = {dj , j = 1, . . . , n } are unlabeled doc-
uments. Is there any useful information about classes in D  ? Although it may seem
counterintuitive, the answer is yes, at least in principle, since unlabeled points allow
us to better characterize the shape of the data distribution.
There are several intuitive explanations of why this is the case. First, suppose pos-
itive and negative examples are generated by two separate distributions, for example,
two Gaussians, and that a very large number of samples are available. In this case
we may expect that the parameters of the two distributions can be estimated well, for
112 TEXT CATEGORIZATION

example, by a training a mixture of two Gaussians. Then, just a few labeled points may
be sufcient to decide which Gaussian is associated with the positive and negative
class (see Castelli and Cover (1995) for a theoretical discussion). A different intuition,
in terms of discriminant classiers, can be gained by considering Figure 4.10, where
unlabeled data points are shown as dots. The thicker hyperplane is clearly a better
solution than the thin maximum margin hyperplane computed from labeled points
only. Finally, as noted by Joachims (1999b), learning from unlabeled data is reason-
able in text domains because categories can be guessed using term co-occurrences.
For example, consider the 10 documents of Table 4.1 and suppose category labels
are only known for documents d1 and d10 (Linux and DNA, respectively). Then it
should be clear that term co-occurrences allow us to infer categories for the remain-
ing eight unlabeled documents. This observation links text categorization to LSI, a
connection that has been exploited for example in Zelikovitz and Hirsh (2001).
None of the classication algorithms we have studied so far can deal directly with
unlabeled data. We now present two approaches that develop the intuition above. The
use of unlabeled data is further discussed in the context of co-training in Section 4.7.

EM and Naive Bayes

Nigam et al. (2000) propose a solution based on the Naive Bayes classier and the
EM algorithm for handling unlabeled data. In their setting, the class variable C for
unlabeled documents is interpreted as a missing variable that can be lled in using
EM. More precisely, the E step consists of computing, for each document di Du , the
conditional probability P (c | di ) (in the case of labeled documents this probability is
clearly unity if c = ci and zero otherwise). This probability is then used to compute
the expected sufcient statistics for the parameters, which are in turn used as soft
counts for parameters re-estimation. Under the word-based multinomial event model
of Equation (4.25), the re-estimation equations in the M step are

qj + ni=1 nij P (c | di )
c,j = |V |  (4.50)
+ k=1 ni=1 nik P (c | di )
and
n
qc  + i=1 P (c | di )
c = . (4.51)
 + n
Nigam et al. (2000) validated the above method on three data sets: the 20 news-
groups, Web KB, and Reuters-21578 (ModApte split). For brevity we only report
their results on the rst data set in Figure 4.11. Note the signicant difference of
accuracy between Naive Bayes with no unlabeled documents and EM with 10 000
unlabeled documents when the number of labeled documents in the training set is
small. As the number of labeled documents increases, the parameters of the Naive
Bayes classier are estimated more reliably and the two approaches tend to the same
accuracy.
TEXT ANALYSIS 113

Transductive SVM
Support vector machines can be also be extended to handle nonlabeled data in the
transductive learning framework of Vapnik (1998). In this setting, the optimization
problem of Equation (4.33) (that leads to computing the optimal separating hyperplane
for linearly separable data) becomes:

yi (w T xi + w0 )  1, i = 1, . . . , n,
min w subject to (4.52)
y1 ,...,yn  ,w,w0 yj (w T xj + w0 )  1, j = 1, . . . , n .

The solution is found by determining a label assignment (y1 , . . . , yn  ) of the training
examples so that the separating hyperplane maximizes the margin of both the data in
D and D  . In other words missing values (y1 , . . . , yn  ) are lled in using maximum
margin separation as a guiding criterion. A similar extension of Equation (4.38) for
nonlinearly separable data is


n 
n
min w + C i + C  j
y1 ,...,yn ,w,w0

i=1 j =1


yi (w T xi + w0 )  1 i , i = 1, . . . , n,

y  (w T x  + w )  1  , j = 1, . . . , n ,
j j 0 j
subject to (4.53)

 0, i = 1, . . . , n,


i
j  0, j = 1, . . . , n .

Clearly an exact solution of the above transductive inference problem is intractable,

as it involves searching in the space of all possible label assignments on n points.
Joachims (1999b) proposes a heuristic algorithm that starts assigning points in D  ,
using the labeling induced by the classier obtained by training a standard SVM on D.
The algorithm then iteratively improves the solution by switching the labels of a posi-
tive and negative example in D  if the switch leads to a reduction of the objective func-
tion. Joachims (1999b) reports empirical results on Reuters-21578 (ModApte split
with 10 categories) and on a subset of Web KB data. The results show that using
a very small number of labeled documents still yields acceptable performance. For
example, Joachims found that the average breakeven point of the transductive SVM
was 60.8% on the Reuters-21578 data set (ModApte split, 10 most frequent classes)
when using a data set D of 17 labeled documents and a test set D  of 3299 documents.
By comparison, training a standard SVM on the same D and testing on the same D 
yields an average breakeven score of 48.4%. Similarly, in the Web KB domain the
average breakeven score was found to improve from 48.6% to 53.5% using a training
set |D| = 9 and a test set |D  | = 3957. Although positive, these results are obtained
in a quite specic experimental setup: the labeled data set D is a subset of the training
data specied in the ModApte split. However, the unlabeled data D  are the ModApte
test documents, i.e. the same documents where precision and recall are later estimated.
Zhang and Oles (2000) repeated the same experiment using a different data split and
114 EXPLOITING HYPERLINKS
Accuracy
100%

80% Using 10,000 unlabeled documents

60%

40%
No unlabeled documents

20%

50 100 500 1000 5000

Number of labeled documents

Figure 4.11 Classication accuracy as a function of the number of labeled documents using
EM and 10 000 unlabeled documents (solid line) and using no unlabeled documents and the
Naive Bayes classier (dotted line). Experimental results reported in (Nigam et al. 2000) for
the 20 newsgroups data set.

found that the transductive SVM actually maximizes the wrong margin i.e. pushes
toward a large separation of unlabeled data but achieve this by actually mislabeling
the unlabeled data. In the setting of Zhang and Oles (2000) the performance of the
transductive SVM was found to be worse than that of the inductive SVM.

4.7 Exploiting Hyperlinks

Text analysis in the case of Web documents can take advantage of the hyperlinked
structure to gain additional information about a document that is not contained in
the document itself. For example, the text in the source anchor text is very often a
synthetic description of the contents of the page being linked to. Even if we did not
observe a target document d(v), we can gain information about this document from
source anchors text in all the documents d(w) such that w pa[v]. In the case of
retrieval, for example, anchor text has been very usefully exploited to score target
pages (Brin and Page 1998).

4.7.1 Co-training
Links in hypertexts offer additional information that can also be exploited for improv-
ing categorization and information extraction. The co-training framework introduced
TEXT ANALYSIS 115

by Blum and Mitchell (1998) especially addresses this specic property of the Web,
although it has more general applicability. In co-training, each instance is observed
through two alternative sets of attributes or views and it is assumed that each view
is sufcient to determine the class of the instance. More precisely, the instance space
X is factorized into two subspaces X1 X2 and each instance x is given as a pair
(x1 , x2 ). For example, x1 could be the bag of words of a document and x2 the bag
of words obtained by collecting all the text in the anchors pointing to that document.
Blum and Mitchell (1998) assume that
(1) the labeling function that classies examples is the same as applied to x1 or to
x2 , and
(2) x1 and x2 are conditionally independent given the class.
Under these assumptions, unlabeled documents can be exploited in a special way for
learning. Specically, suppose two sets of labeled and unlabeled documents D and
D  are given. The iterative algorithm described in Blum and Mitchell (1998) then
proceeds as follows. Labeled data are used to infer two Naive Bayes classiers, one
that only considers the x1 portion of x and one that only considers the portion x2 . Then
these two classiers are used to guess class labels of documents in a subset of D  . A
xed amount of such instances that are classied as positive with highest condence
is then added as positive examples to D. A similar procedure is used to add self-
labeled negative examples to D and the process iterated, retraining the two Naive
Bayes classiers on both labeled and self-labeled data. Blum and Mitchell (1998)
report a signicant error reduction in an empirical test on Web KB documents.
Nigam and Ghani (2000) have shown that co-training offers signicant advantages
over EM if there is a true independence between the two feature sets.

4.7.2 Relational learning

Relational learning is another interesting approach for exploiting hypertextual infor-
mation. In this setting, data are assumed to be available in a relational format and
a learning algorithm can exploit relations among data items. In the case of Web
data, there are obvious relations that are encoded in the hyperlinked structure of Web
pages, but also local relations associated with the semi-structured organization of text
as determined by the descriptive markup in HTML. As in the co-training framework
described above, the general intuition is that, as well as the text on the pages being rel-
evant to the classication of the page, the hyperlinks also contain useful information
about the category a page belongs to.
FOIL (Quinlan 1990) is a classic algorithm for dealing with relational data and
learning rst-order clauses such as those used in the Prolog programming language.
Although a description of FOIL is out of the scope of this book, it is interesting to
remark that these approaches have been exploited with success to take hypertextual
information into account. Craven et al. (2000) applied FOIL to learn classication
rules in the Web KB domain (Figure 4.8 shows an ontology for that domain).
116 DOCUMENT CLUSTERING

Table 4.5 Examples of rst-order logic classication rules learned from Web KB data
(Craven et al. 2000). Note that terms such as jame and faculti result from text conation.

student(A) :- not(has data(A)), not(has comment(A)), link to(B,A),

has jame(B), has paul(B), not(has mail(B)).

faculty(A) :- has professor(A), has ph(A), link to(B,A), has faculti(B).

Two sample rules are reported in Table 4.5. For example, a Web page is classied as
a student page if it does not contain the terms data and comment and it is linked by
another page that contains the terms jame and paul but not the term mail. The method
was later rened by characterizing documents and links in a statistical way using the
Naive Bayes model in combination with FOIL Craven and Slattery (2001).
Graphical models such as Bayesian networks that are traditionally conceived for
propositional data (i.e. each instance of case is a tuple of attributes) have been more
recently generalized to deal with relational data as well. Relational Bayesian net-
works have been introduced in Jaeger (1997) and learning algorithms for probabilistic
relational models are presented in Friedman et al. (1999). Taskar et al. (2002) pro-
pose a probabilistic relational model for classifying entities, such as Web documents,
that have known relationships. This leads to the notion of collective classication,
where classication of a set of documents is viewed as a global classication prob-
lem rather than classifying each page individually. Taskar et al. (2002) use a general
probabilistic Markov network model for this class of problems and demonstrate that
taking relational information (such as hyperlinks) into account clearly improves clas-
sication accuracy. Cohn and Hofmann (2001) describe a probabilistic model that
jointly accounts for contents and connectivity integrating ideas from PLSA (see Sec-
tion 4.5.2) and PHITS (see Section 5.6.2).
Classication and probabilistic modeling of relational data in general is still a rela-
tively less explored research area at this time but is clearly a promising methodology
for classication and modeling of Web documents.

4.8 Document Clustering

4.8.1 Background and examples
As discussed in Chapter 1, clustering is the process of nding natural groups in data
where there are no class labels present, i.e. all of the training data are unsupervised. In
document clustering the objects to be clustered are typically represented using a bag-
of-words representation. Document clustering is useful across a variety of applications
related to the Web. We may for example want to automatically group Web pages
on a website into clusters based on their content for the purposes of automatically
generating links in real time to different clusters. If we have a relatively small number
of pages, it may be possible to manually assign pages to clusters. However, many
TEXT ANALYSIS 117

websites have tens of thousands of pages. In these cases it is clearly impractical to

manually label all of the Web pages and automated clustering is an attractive approach.
Another application of clustering of Web documents is the automated grouping of
the results of search engine queries. Search engines typically return a ranked list of
URLs relative to a specied query. For example, issuing the query World Cup to the
Google search engine on 2 December 2002 yielded the following results.
The top ranked URL was http://www.fifaworldcup.com/ and indeed
the next 10 ranked URLs, as well as many of the top 100, were also about soccer.
Clearly soccer is the largest cluster of URLs for this query.
However, the 12th ranked URL was http://www.dubaiworldcup.com,
a Web page about the richest horse race in the world and obviously a different
topic from soccer. As such it could be assigned to a different cluster.
Further down the list we nd items such as http://www.wcsk8.com/
for skateboarding, http://www.cricketworldcup.com/ for cricket,
http://www.pwca.org/ for paragliding, http://www.rugby2003.
com.au/ for rugby, and so on. Each of these could reasonably be expected to
be in different small clusters.
Robot soccer also appears on the list, http://www.robocup.org/, and
could perhaps be a subcluster of the main soccer cluster.
Skiing appears multiple times, both downhill skiing with http://www.fis-
ski.com/ and http://www.skiworldcup.org/ and cross-country
skiing with http://www.uwasa.fi/hkn/WC-Skiing/. Again per-
haps these could be viewed as two subclusters of a cluster about skiing.
Naturally it would be quite useful to automatically group these URLs in real time
into meaningful clusters. In fact some search engines already support this feature,
one example being http://www.vivisimo.com/. For the same query of World
Cupon 2 December 2002, the Vivisimo engine returned 175 URLs with the following
cluster titles and number of URLs in each cluster in parentheses: FIFA World Cup
(44); Soccer (42); Sports (24); History (19); Rugby World Cup (13); Womens World
Cup (10); Betting (8); Cricket (6); Skiing World Cup (3); and so on. The details of this
particular algorithm are not published but it can cluster about 200 URLs in response
to a query in human real time and automatically assign labels to the clusters (such
as those assigned to the World Cup clusters above).

4.8.2 Clustering algorithms for documents

There are a variety of classic clustering algorithms that can be applied to document
clustering including hierarchical clustering, K-means type clustering, or probabilistic
clustering (see the earlier discussion in Chapter 1 on clustering). In what follows we
provide a brief discussion of applying these algorithms to document clustering.
118 DOCUMENT CLUSTERING

Hierarchical clustering
Hierarchical clustering algorithms do not presume a xed number of clusters K in
advance but instead produce a binary tree where each node is a subcluster of the
parent node. Assume that we are clustering n objects. The root node consists of a
cluster containing all objects and the n leaf nodes correspond to clusters where
each contains one of the n original objects. This tree structure is known as a dendro-
gram.
Agglomerative hierarchical clustering algorithms start with a pairwise matrix of
distances or similarities between all pairs of objects. At the start of the algorithm all
objects are considered to be in their own cluster. The algorithm operates by iteratively
and greedily merging the two closest clusters into a new cluster. The resulting merging
process results in the gradual bottom-up construction of a dendrogram. The denition
of closest depends on how we dene distance between two sets (clusters) of objects
in terms of their individual distances. For example, if we dene closest as being the
smallest pairwise distance between any pair of objects (where the rst is in the rst
cluster, and the second is in the second cluster) we will tend to get clusters where
some objects can be relatively far away from each other. This leads to the well-known
chaining effect if, for example, the distances correspond to Euclidean distance in
some d-dimensional space, where the cluster shapes in d-space can become quite
elongated and chain like.This minimum-distance algorithm is also known as single-
link clustering.
In contrast we could dene closest as being the maximum distance between all
pairs of objects. This leads naturally to very compact clusters if viewed in a Euclidean
space, since we are ensuring at each step that the maximum distance between any
pair of objects in the cluster is minimized. This maximum-distance method is known
as complete-link clustering. Other denitions for closest are possible, such as com-
puting averages of pair-wise distances, providing algorithms that can be thought of
as between the single-line and complete-link methods.
A useful feature of hierarchical clustering for documents is that we can build
domain-specic knowledge into the denition of the pairwise distance measure bet-
ween objects. For example, the cosine distance in Equation (4.1) or weighting schemes
such as TFIDF (Equation (4.5)) could be used to dene distance measures. Alter-
natively, for HTML documents where we believe that some documents are struc-
turally similar to others, we could dene an edit distance to reect structural differ-
ences.
However, a signicant disadvantage of hierarchical clustering is that the agglom-
erative algorithms have a time complexity between O(n2 ) and O(n3 ) depending on
the particular algorithm being implemented. All agglomerative algorithms are at least
O(n2 ) due the requirement of starting with a pairwise distance matrix between all
n objects. For small values of n, such as the clustering of a few hundred Web pages
that are returned by a search engine, an O(n2 ) algorithm may be feasible. How-
ever, for large values of n, such as 1000 or more, hierarchical clustering is somewhat
impractical from a computational viewpoint.
TEXT ANALYSIS 119

Probabilistic model-based clustering

An alternative approach to hierarchical clustering is that of probabilistic model-based
clustering. Earlier in Chapter 1 we briey reviewed clustering using the Naive Bayes
model, and earlier in this chapter we discussed using the same model for classica-
tion. Recall from Chapter 1 that model-based clustering assumes the existence of a
generative probabilistic model for the data, typically in the form of a mixture model
with K components, where each component corresponds to a probability distribution
model for one of the clusters. The problem of clustering in this context amounts to
one of learning the parameters of this mixture model, specically, the parameters of
each component model and the mixture weights for each component.
As we saw with classication earlier in this chapter, the Naive Bayes model is
attractive as a component model for document clustering, since it contains only one
parameter per dimension and the dimensionality of our document vectors is likely be
high, e.g. as high as 500050 000 if we are using the bag-of-words representation.
Even with a model such as the Naive Bayes that is linear in the number of parameters,
these dimensionalities are still rather high. In practice a number of heuristics can be
used to reduce dimensionality, such as removing very common words as well as very
rare words. We cannot use standard class-based feature selection criteria directly (as
we did for classication earlier in this chapter), since in clustering we do not know a
priori what the classes (clusters) are.

4.8.3 Related approaches

Tantrum et al. (2002) discuss various strategies for combining ideas from hierarchical
clustering and model-based clustering with applications to sets of documents. In this
work the original bag-of-words representation is reduced to a 50-dimensional space by
applying latent semantic indexing before clustering, and then performing clustering
in the reduced-dimensional space. Dhillon and Modha (2001) investigate a spherical
K-means algorithm for clustering of sparse high-dimensional document vectors and
again use ideas from linear algebra to reduce dimensionality improve the quality
of clustering. The motivation in both of these approaches is to combine dimension
reduction with clustering, with the promise of being able to achieve more reliable
clustering in the lower-dimensional space. A caveat here is that dimension-reduction
techniques could in fact destroy the very cluster structure that we are seeking, unless
the objective function for dimension reduction somehow incorporates the notion of
what constitutes a good clustering of the data.
From a computational viewpoint, Dhillon et al. (2001) describe a number of prac-
tical methods for efcient implementation of the spherical K-means algorithm for
document clustering. McCallum et al. (2000b) propose the use of approximate dis-
tance measures that are cheap to compute and that can divide a data set into subsets
called canopies.The general idea is that only objects within the same canopy need to
be considered when updating clusters associated with that canopy and this is particu-
larly efcient when both the dimensionality of the problem and the number of objects
120 INFORMATION EXTRACTION

are both very large. The authors describe speed-ups of up to an order of magnitude for
greedy agglomerative clustering, K-means clustering, and probabilistic model-based
clustering, using their approach.
A number of other representations (besides hierarchical clustering, K-means, or
mixtures models) have also been proposed for clustering documents. Taskar et al.
(2002) describe how their Markov network model for relational data can be used
to incorporate both text on the page as well as hyperlink structure for clustering of
Web pages. Slonim and Tishby (2000) propose a very general information-theoretic
technique for clustering called the information bottleneck. The technique appears to
work particularly well for document clustering (Slonim et al. 2002).
Zamir and Etzioni (1998) describe a specic algorithm for the problem discussed
earlier of clustering the results of search engines that was . They use snippetsreturned
by Web search engines as the basis for clustering and propose a clustering algorithm
that uses sufx tree data structures based on phrases between algorithms. In the
data sets used in their experiments they show that the resulting algorithm is both
computationally efcient (linear in the number of documents) and nds clusters that
are approximately as good as those obtained from clustering the full text of the Web
pages.

4.9 Information Extraction

An interesting application of machine-learning and articial intelligence techniques
to text documents in general (and to text data on the Web in particular) is that of
information extraction (Cohen et al. 2000). The general idea is to automatically extract
unstructured text data from Web pages and to represent this extracted information in
some well-dened schema (e.g. in a form suitable to enter into a relational database
table). For example, we might want to extract information on authors and books from
various online bookstore and publisher pages. Furthermore we might want to search
for online reviews of these books, in newspaper and magazine articles, at online stores,
etc. In another type of application a company might want to continuously crawl the
Web searching for information about certain technologies or products of interest.
In general the problem of information extraction can often be characterized as that
of detecting information about entities (individuals, organizations, objects, etc.)
in Web pages and then performing some form of parsing to extract the relevant
information such as the name of the entity, attributes of the entity, and relationships
among multiple entities. This is a rich research topic here we only provide the
briefest of overviews but hopefully this will whet the readers interest in exploring
some of the references below and learning more about this topic. Our overview is
loosely based on Cohen and McCallum (2002).
Early efforts in information extraction relied on hand-built models, for example the
FASTUS system (Appelt et al. 1995) that uses cascades of nite-state machines to
parse text sequences into lexical units, entity names, groups of words associated with
verbs, and so forth. More recent work, including, for example, the Web-KB system
TEXT ANALYSIS 121
START

TITLE

EMAIL
AUTHOR

AFFILIATION

ADDRESS
ABSTRACT

END

Figure 4.12 A toy example of states and transitions in an HMM for extracting various elds
from the beginning of research papers. Not shown are various possible self-transition probabil-
ities (e.g. multiple occurrences of the state author), transition probabilities on the edges, and
the probability distributions of words associated with each state. See Figure 7 in McCallum et
al. (2000b) for a more detailed and realistic example of an HMM for this problem.

described earlier in this chapter, has focused largely on the use of machine-learning
and statistical techniques that leverage human-labelled data to automate the process
of constructing information extractors. In effect these systems use the human-labelled
text to learn how to extract the relevant information (Cardie 1997; Kushmerick et al.
1997).
One general approach is to dene information extraction as a classication problem.
For example, we might want to classify short segments of text in terms of whether
they correspond to the title, author names, author addresses and afliations, and so
forth, from research papers that are found by crawling the Web (for example this is one
component of the functionality in the CiteSeer digital library system (Lawrence et al.
1999)). A classication approach to this problem would be to represent each document
as a sequence of words and then use a sliding window of xed width k as input to a
classier each of the k inputs to the classier is a word in a specic position. The
system can be trained on positive and negative examples that are typically manually
labeled. A variety of different classiers such as Naive Bayes, decision trees, and
relational rule representations have been used for sliding-window based classication
(Baluja et al. 2000; Califf and Mooney 1998; Freitag 1998; Soderland 1999).
A limitation of the sliding window is that it does not take into account sequential
constraints that are naturally present in the data, e.g. the fact that the author eld
almost always precedes the address eld in the header for a research paper. To
take this type of structure into account, one approach is to train stochastic nite-
122 EXERCISES

state machines that can incorporate sequential dependence. One popular choice has
been hidden Markov models (HMMs), which were mentioned briey in Chapter 1
during our discussion of graphical models. An HMM contains a nite-state Markov
model, where each state in the model corresponds to one of the elds that we wish
to extract from the text. For example, in parsing headers of research papers we could
have states corresponding to the title of the paper, author name, etc. An example of
a simple Markov state diagram is shown in Figure 4.12. The key idea in HMMs is
that the true Markov state sequence is unknown at parse-time instead we see noisy
observations from each state, in this case the sequence of words from the document.
Each state has a characteristic probability distribution over the set of all possible
words, e.g. the distribution of words from the state title will be quite different from
the distribution of words for the state author names. Thus, given a sequence of
words, and an HMM, we can parse the observed sequence into a corresponding set
of inferred states the Viterbi algorithm provides an efcient method (linear in the
number of observed words) for producing the most likely state-sequence, given the
observations.
For information extraction, the HMM can be trained in a supervised manner with
manually labeled data, or bootstrapped using a combination of labeled and unlabeled
data (where the EM algorithm is used for training HMMs on unlabeled data). An early
application of this approach to the problem of named entity extraction (automatically
nding names of people, organizations, and places in free text) is described in Bikel
et al. (1997). A variety of related ideas and extensions are described in Leek (1997),
Freitag and McCallum (2000), McCallum et al. (2000a) and Lafferty et al. (2001).
For a detailed description of the use of machine-learning methods (and HMMs in
particular) to automatically create a large-scale digital library see McCallum et al.
(2000c).
Information extraction is a broad and growing area of research at the intersec-
tion of Web research, language modeling, machine learning, information retrieval,
and database theory. In this section we have barely scratched the surface of the many
research problems, techniques, and applications that have been proposed, but nonethe-
less, we hope that the reader has gained a general high-level understanding of some
of the basic concepts involved.

4.10 Exercises
Exercise 4.1. Given a collection of m documents to be indexed, compare the memory
required to store pointers as integer document identiers and as the difference between
consecutive document identiers using Eliass coding. (Hint: use Zipfs Law for
term frequency and note that the most frequent term occurs in n documents, the second
most frequent in n/2 documents, the third in n/4 documents and so on.)
Exercise 4.2. A vector x Rm is said to be sparse if

 = |{j = 1, . . . , m : xj  = 0}|  m.
TEXT ANALYSIS 123

The dot product of two vectors x and x  normally requires (m) time. In the case of
sparse vectors, this can be reduced to ( +  ). Describe an efcient algorithm for
computing the dot product of sparse vectors. (Hint: use linear lists to store the vectors
in memory.)
Exercise 4.3. Consider the vector-space representation of documents and compare
the cosine distance (Equation (4.1)) to the ordinary Euclidean distance. Show that for
vectors of unit length the ranking induced by the two distances is the same.
Exercise 4.4. Show that a data set of n vectors in Rm is linearly separable if n
m+1.
Exercise 4.5. Consider a data set of n text documents. Suppose L is the average
number of words in each document. How many documents must be collected, on
average, before the sufcient condition of Exercise 4.4 fails? (Hint: use Heaps law
to estimate vocabulary growth.)
Exercise 4.6. Collect subjects of your email mailbox and build a termdocument
matrix X using a relatively small set of frequent terms. Compute the SVD decompo-
sition of the matrix using a software package capable of linear algebra computations
(e.g. octave) and examine the resulting reconstructed matrix X. Try querying your
messages and compare the results obtained with a simple Boolean model where you
simply retrieve matching subjects and the LSI model. Experiment with different val-
ues of k (the dimension of the latent semantic space).
Exercise 4.7. Build your own version of the Naive Bayes classier and evaluate its
performance on the Reuters-21578 data set. In a rst setting, keep all those classes
having at least one document in the training set and one in the test set. Repeat the
experiment using only the top 10 most frequent categories. Beware that categories
in the collection are not mutually exclusive (some documents belong to several cate-
gories).
Exercise 4.8. Write a program that selects the most informative features using mutual
information. Test your program on the 20 Newsgroups corpus. Train your Naive Bayes
classier using the most k informative terms for different values of k and plot your
generalization accuracy. What is your best value of k?
Repeat the exercise by using Zipfs Law on text, removing words that are either too
frequent or too rare. Try different cutoff values for the k1 most frequent words and the
k2 most rare words. Compare your results to those obtained with mutual information.
Finally combine the two methods above and compare your results.
Exercise 4.9. Obtain a public domain implementation of Support Vector Machines
(e.g. Joachimss SVMlight , http://svmlight.joachims.org/) and set up
an experiment to classify documents in the Web KB corpus. Since SVMs are
binary classiers you need to devise some coding strategy to tackle the multiclass
problem. Compare results obtained using the two easiest strategies: one-against-all
and all-pairs.
This Page Intentionally Left Blank
5

Link Analysis

Computing the relevance of a document is a major issue in Web-based information

retrieval. As we have seen in Chapter 4, when considering unstructured collections
of documents, it is possible to compute relevance with respect to user queries such
as those involving keyword-based Boolean expressions, or those involving measures
of similarity between documents. In each of these cases the score is a function of
the document and the query. Large collections of hypertext such as the Web are
more interesting in this respect since they allow us to compute scoring functions that
include topological information about the hypertext graph. The basic assumption is
that hyperlinks contain information about the human judgment of a document. To
a rst approximation, the more incoming links that exist for a document, the more
likely it is that the document was judged to be important by the authors of other
documents linking to it.
Incoming links embody a common notion of popularity that also exists in other
domains or in other webs. Networks of interaction have been studied for a long time
in social sciences (Wasserman and Faust 1994), where nodes correspond to persons
or organizations, and edges represent some type of social interaction. Intuitively,
increasing the number of incoming links to a node should increase a common-sense
measure of standing or popularity or prestige for that node. However, it should be also
common sense that just counting the number of links does not necessarily provide an
accurate measure of standing. For example, measuring the prestige of an enterprise
by the mere number of its clients could be misleading, since different clients may
have very different weights.
An important example is the network obtained by considering the scientic litera-
ture. Nodes in this case are papers, books, or entire journals, and edges correspond to
citations. It makes sense to assume that the more citations a paper or a book receives,
the more it can be assumed to be important, since it has been judged as useful by other
scientists. The systematic construction of such networks through citation indexes was
introduced by Gareld (1955), who later proposed a measure of standing for journals
that is still in use. This measure, called impact factor (Gareld 1972), is dened as
the average number of citations per recently published item. More precisely, if C is
the total number of citations in a given time interval [t, t + t1 ] to articles published
Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
126 EARLY APPROACHES TO LINK ANALYSIS

by a given journal during [t t2 , t], and N is the total number of articles published
by that journal in [t t2 , t], the impact factor is dened as C/N (typically t1 = 1
year and t2 = 2 years). Thus, the impact factor is a very simple measure, since it
basically corresponds to the normalized indegree of a journal in a subgraph of the
citation network. Graph-based link analysis for the scientic literature goes back to
the 1960s (Garner 1967). However, these ideas were not exploited in the development
of rst-generation Web search tools.
The paper by Bray (1996) reports an early attempt to apply social networks concepts
to the Web. He suggested a Web visualization approach where the . . . appearance of
a site should reect its visibility, as measured by the number of other sites that have
pointers to it . . . and . . . its luminosity, as measured by the number of pointers with
which it casts navigational light off-site. . . . Visibility and luminosity dened in this
way are directly related to the indegree and the outdegree of websites, respectively.
More recently, toward the end of the 1990s, link analysis methods became more
widely known and used in a search engine context, leading to what is sometimes
called the second generation of Web searching tools.
This chapter reviews the most common approaches to link analysis and how these
techniques are be applied to compute the popularity of a document or a site. The algo-
rithms presented in this chapter extract emergent properties from a complex network
of interconnections, attempting to model (indirectly) subjective human judgments.
It remains debatable as to whether popularity (as implied by the mechanism of cita-
tions) captures well the notions of relevance and quality as they are subjectively
perceived by humans, and whether link analysis algorithms can successfully model
human judgments.

5.1 Early Approaches to Link Analysis

Our notation for hypertext will be straightforward. For each vertex v in the hypertext
graph G = (V , E), d(v) denotes the contents of the document at vertex v. If d(v)
is considered to be an isolated document, then its score with respect to a query q
is s(v | d(v), q). When considering d(v) in its hypertext context, the score should
also depend on G and will be denoted as S(v | d(v), q, G). In the following, we will
simplify the notation of these scores by just writing s(v) and S(v) if the dependencies
on d(v), q, and G are obvious from the context.
To quantify visibility (luminosity) S(v) could simply be designed to grow with
the indegree (outdegree) of v as hinted by Bray (1996). Clearly, however, such an
approach suffers from a fundamental limitation: it would fail to capture the relative
importance of different parents (children) in the graph. For example, a Web page
with a small number of links coming from important sites should be considered more
popular than a Web page with a larger number of links and whose sources are all from
unimportant or irrelevant sites. Hence, rather than a mere count, popularity should be
computed as a weighted sum of the citations a document receives through hypertext
links. Ideas having this avor are less recent than we might expect.
LINK ANALYSIS 127

The use of hypertext information in information retrieval is older than the Web.
Mark (1988), for example, was concerned with retrieving hypertext cards in a medical
domain and noted that . . . often cards do not even mention what they are about, but
assume that the reader understands the context because he or she has read earlier
cards. He then proposed a simple algorithm for scoring documents where relevance
information was transmitted from documents to their parents in the hypertext graph
G. More precisely, the global score of v given the query and the topology of G was
computed as:
1 
S(v) = s(v) + S(w). (5.1)
| ch[v]|
w| ch[v]|
This simple algorithm somewhat resembles message passing schemes that are very
common in connectionism (McClelland and Rumelhart 1986) or in graphical model-
ing (Pearl 1988). As such, it requires G to be a DAG so that a topological sort 1 can be
chosen for updating the global scores S. The DAG assumption is reasonable in small
hypertexts with a root document and a relatively strong hierarchical structure (in this
case, even if G is not acyclic, not much information would be lost by replacing it with
its spanning tree). The Web, however, is a large and complex graph. This may explain
why search engines largely ignored its topology for several years.
The paper by Marchiori (1997) was probably the rst one to discuss the quantitative
concept of hyper information to complement textual information in order to obtain
the overall information contained in a Web document. The idea somewhat resembles
Frisses approach. Indeed, if we rewrite Equation (5.1) as
S(v) = s(v) + h(v), (5.2)
then s(v) can be thought of as the textual information (that only depends on the
document and the query), h(v) corresponds to the hyper information that depends on
the link structure where v is embedded, and S(v) is the overall information. Marchiori
(1997) did not cite Mark (1988), but nonetheless he identied a fundamental problem
with Equation (5.1). If an irrelevant page v has a single link to a relevant page w,
Equation (5.1) implies that S(v)  S(w). The scenario would be even worse in a
chain of documents v0 , v1 , . . . , vk . Here if S(vk ) is very high but S(v0 ), . . . , S(vk1 )
are almost zero, then v0 would receive a global score higher than vk , even though a
user would need k clicks to reach the important document.
As a remedy, Marchiori suggested that in this case the hyper information of v0
should be computed as

h(v) = F r(v,w) S(w), (5.3)
wch[v]

where F (0, 1) is a fading constant and r(v, w) {1, . . . , |ch[v]|} is the rank
of w after sorting (in ascending order) the children of v according to the value of
1 A topological sort is an ordering < of the vertices such that v < v  if and only if there is a directed
path from v  to v (Cormen et al. 2001).
128 NONNEGATIVE MATRICES AND DOMINANT EIGENVECTORS

S(w). When applying Equation

 (5.3) to a linear chain v0 , v1 , . . . , vk of documents,
one would get S(v0 ) = ki=1 F i S(vi ), so the contribution of the score of vk fades
exponentially as one moves back in the graph. As it turns out, Equation (5.3) in general
implies a recursive form of computation that cannot be carried out in a cyclic graph.
Marchiori (1997) suggested a solution to this problem assuming a nite horizon of
propagation, i.e. S(v) was computed on the tree rooted at v and having a xed small
depth k.
Before we discuss link analysis of the Web graph, it will rst be useful to estab-
lish certain basic mathematical results relating to nonnegative matrices, graphs, and
Markov chains. The reader familiar with these topics can safely skip the next section.

5.2 Nonnegative Matrices and Dominant

Eigenvectors
A square matrix A is said to be nonnegative, written A  0, if all its elements are
nonnegative. Important examples of nonnegative matrices include graph incidence
(of adjacency) matrices and stochastic matrices. Given a directed graph G = (V , E),
the incidence matrix A of G is dened as the 0-1 matrix with aij = 1 if and only if
(i, j ) E. Note that for simplicity we have assumed that the vertices in V are iden-
tied by the integers 1, 2, . . . , n, where n = |V |. Stochastic matrices are commonly
used to describe rst-order Markov chains as discussed in Appendix A. In a system
with n discrete states, each entry of a stochastic matrix A contains the transition
probability aij = P (St = j | St1 = i). In this case, the probability axioms imply
that each aij  0 and that the elements of each row i should sum to unity.
A nonnegative n n matrix A is said to be irreducible if, for each pair of indices
(vertices) i and j , there exists a corresponding integer t such that (At )ij > 0. If A
is the adjacency matrix of a (directed) graph, this property tells us that the graph is
(strongly) connected. By contrast, a reducible matrix is associated with a graph with
more than one (strongly) connected component. In this case, if there exists a path of
length t from a node i to itself, (At )ii > 0. The greatest common divisor (gcd) of
the set {t : (At )ii > 0} is called the period of i. If A is irreducible, then the period
is the same for all indices (nodes) i. The common period is the gcd of the lengths of
all the cycles in the graph. Interestingly, these topological properties of a graph have
a correspondence in the spectral structure of its adjacency matrix, as shown by the
PerronFrobenius theorem (Seneta 1981).
For a nonnegative irreducible primitive matrix A, the PerronFrobenius theorem
allows us to conclude that there exists an eigenvalue of A such that
(1) is real and positive, and  | | for every other eigenvalue   = ;

(2) corresponds to a strictly positive eigenvector;

(3) is a simple root of the characteristic equation (A In ) = 0.

LINK ANALYSIS 129

3 4 3 4

1 2 1 2

(a) (b)

3 4 3 4

1 2 1 2

5 5

(c) (d)
Figure 5.1 Graphs with different types of incidence matrices. (a) is primitive,
(b) is irreducible (with period 4) but not primitive, (c) and (d) are reducible.

In this case, is called the dominant eigenvalue of A and the associated eigenvector
is called the dominant eigenvector. Denoting by (1 , . . . , n ) the eigenvalues of A,
in the following we will assume that the dominant eigenvalue is always 1 .
Note, however, that although there cannot be multiple roots there may be some
other eigenvalue j  = 1 such that |j | = |1 |. It can be shown that if there are k
eigenvalues having the same magnitude as the dominant eigenvalue, then they are
equally spaced in the complex circle of radius 1 . Moreover, if A is the adjacency
matrix of a graph, k is the gcd of the lengths of all the cycles in the graph. In order
to get a dominant eigenvalue that is strictly greater than all other eigenvalues further
conditions are necessary.
A matrix A is said to be primitive if there exists a positive integer t such that
At > 0 (note the strict inequality). A primitive matrix is also irreducible, but the
converse is not true in general. For a primitive matrix, condition 1 of the Perron
Frobenius theorem holds with strict inequality. This means that all the remaining
eigenvalues are smaller in modulus than the dominating eigenvalue. Moreover, if the
adjacency matrix of a graph is primitive, then the gcd of the lengths of all cycles is
unity. Figure 5.1 illustrates some examples.
130 NONNEGATIVE MATRICES AND DOMINANT EIGENVECTORS

Root set R 5
1

10
6

11 7
12
3

4
8
13

14

15

Figure 5.2 Example of a base subgraph obtainedby starting from the vertex set {1, 2, 3, 4}.

The fundamental property of primitive matrices also suggests a simple iterative

algorithm for computing the dominant eigenvalue and the associated eigenvector.
Let x Rn and let (a1 , . . . , an ) denote the coordinates of x in the basis formed
by the eigenvectors (v1 , . . . , vn ). If we expand the product At x, remembering that
Avi = i vi , we obtain 
At x = ai ti vi , (5.4)
i
but since 1 > |i |, i > 1, the rst term dominates the above sum as t gets large,
i.e. ai i vi At for large t. This gives us a vector proportional to the dominant
eigenvector, provided that x is not orthogonal to v1 . Since the PerronFrobenius
theorem tells us that v1 is strictly positive, any random positive vector will yield the
correct solution, for example, x = 1 = (1, 1, . . . , 1)T .
In the special case of primitive stochastic matrices, it is easy to see that 1 = 1 since,
by the denition of a stochastic matrix, A1 = 1. Moreover, since all the remaining
eigenvalues are strictly smaller than one in modulus, the sequence At converges at
an exponential rate and it can be shown that
lim At = 1T r. (5.5)
t
LINK ANALYSIS 131

In this case, r is known as the stationary distribution of the Markov chain.

5.3 Hubs and Authorities: HITS

Kleinbergs algorithm, called Hypertext Induced Topic Selection (HITS), simul-
taneously computes a pair of scoring values associated with hypertext documents
(Kleinberg 1998, 1999). The semantics attached to these quantities essentially match
Brays concepts of visibility and luminosity.
HITS works on a small graph associated with a selected collection of hypertext
documents, for example a focused portion of the Web that is expected to be related to
a given topic of interest. In the original formulation of HITS, the subgraph of interest
(also known as the base subgraph) is computed by selecting the neighbors of a root
set R of Web pages that are known to be relevant with respect to the topic. The root
set is expanded to include all the children and a xed number of parents of nodes in
R. Details of the procedure are given in the following algorithm (see Figure 5.2 for
an illustration).
BaseSubgraph(R, d)
1 SR
2 for each v in R
3 do S S ch[v]
4 P pa[v]
5 if |P | > d
6 then P arbitrary subset of P having size d
7 S SP
8 return S

Note that the children of a given node (line 3) are forward links and can be
obtained directly from each page v. Parents (line 4) correspond to backlinks and
can be obtained from a representation of the Web graph obtained, for example,
through a crawl. Several commercial search engines currently support the spe-
cial query link:url that returns the set of documents containing url as a link.
In the case of small scale applications, this approach can be used to obtain the
set of parents in line 4. Parameter d is the maximum number of parents of a
node in the root set that can be added. As we know (see Chapter 3), some pages
may have a very large indegree. Thus, bounding the number of parents is cru-
cial in practical applications. Algorithm BaseSubgraph returns a set of nodes S.
In what follows, HITS is assumed to work on the subgraph of the Web induced
by S.
Let G = (V , E) denote the subgraph of interest, where V = S. For each vertex
v V , let us introduce two positive real numbers a(v) and h(v). These quantities
are called the authority and the hubness weights of v, respectively. Intuitively, a
document should be very authoritative if it has received many citations. As discussed
132 HUBS AND AUTHORITIES: HITS

above, citations from important documents should be weighted more than citations
from less-important documents. In the case of HITS, the importance of a document as
a source of citations is measured by its hubness. Intuitively, a good hub is a document
that allows us to reach many authoritative documents through its links. The result is
that the hubness of a document depends on the authority of the cited documents, and
the authority of a document depends on the hubness of the citing documents. We are
apparently stuck in a loop, but let us observe that this recursive form of dependency
between hubs and authority weights naturally leads to the denition of the following
operations:

a(v) h(w), (5.6)
wpa[v]

h(v) a(w). (5.7)
wch[v]

The two operations above can be carried out to update authority and hubness weights
starting from initial values. This approach is meaningful because Kleinberg (1999)
showed that iterating Equations (5.6) and (5.7), intermixed with a proper normaliza-
tion step, yields a convergent algorithm. The output is a set of weights that can be
therefore considered to be globally consistent. Kleinbergs algorithm is listed below.
For convenience, weights are collected in two n-dimensional vectors a and h.

HubsAuthorities(G)
1 1 [1, . . . , 1] R|V |
2 a 0 h0 1
3 t 1
4 repeat
5 for each v inV
6 do at (v) wpa[v] ht1 (w)

7 ht (v) wch[v] at1 (w)
8 at at /at 
9 ht ht /ht 
10 t t +1
11 until at at1  + ht ht1  <
12 return (at , ht )

To show that HubsAuthorities terminates, we need to prove that for each > 0
the condition controlling the outer loop will be met for t large enough. Formally,
this means that the sequences {at }iN and {ht }tN converge to limits a  and h ,
respectively. The proof of this result is based on rewriting HITS using linear algebra.
In particular, if we denote by A the incidence matrix of G, it can be easily veried that
the updating operations can be written compactly in vector notation as at = AT ht1
LINK ANALYSIS 133

The authority weight of a page is the sum

of the hubness weights of its parents

The hubness weight of a page is the sum

of the authority weights of its children

Figure 5.3 Graphical explanation of the basic operations in HITS

(see Equations (5.6) and (5.7)).

and ht = Aht1 . As a result, after t iterations,

at = t (AT A)t1 AT 1, (5.8)
ht = t (AAT )t 1, (5.9)
where t and t are scalar normalization factors. Thus, a sufcient condition for
HubsAuthorities to terminate is that the sequences of vectors {t (AT A)t1 AT 1}
and t (AAT )t 1 converge for t . It is possible to prove that these sequences
converge under fairly unrestrictive hypotheses. For example, it can be shown that a
sufcient condition is that M be a nonnegative nonsingular symmetric matrix. In this
case, the dominant eigenvector 1 (M) (i.e. the eigenvector associated with the largest
eigenvalue 1 (M)) is nonnegative, and for every vector x such that 1 (M)T x  = 0
we have
Mt x
lim 1 (M). (5.10)
t M t x

Since 1 cannot be orthogonal to a nonnegative vector, the sequences {at } and {ht }
converge to 1 (AT A) and 1 (AAT ), respectively.
As an example, in Figure 5.4 we show the authority and hubness weights computed
by HubsAuthorities on the graph of Figure 5.2. We can note some unobvious
weight assignments. For example, vertex 3 has the largest indegree in the graph but
nonetheless its authority is rather small because of the low hubness weight of its
parents.
Bharat and Henzinger (1998) have suggested an improved version of HITS that
addresses some specic problems that are encountered in practice. For example, a
mutual reinforcement effect occurs when the same host (or document) contains many
identical links to the same document in another host. To solve this problem, Bharat
and Henzinger (1998) modied HITS by assigning weight to these multiple edges that
134 PAGERANK
Authority

Hubness

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Figure 5.4 Authority and hubness weights in the example graph of Figure 5.2.

are inversely proportional to their multiplicity. The method presented by Bharat and
Henzinger (1998) also addresses the problem of links that are generated automatically,
for example, by converting messages posted to Usenet news groups into Web pages.
Finally, they address the so-called topic drift problem, i.e. some nodes in the base
subgraph may be irrelevant with respect to the user query and documents with highest
authority or hubness weights could be about different topics. This problem can be
addressed either by pruning irrelevant nodes or by regulating the inuence of a node
with a relevance weight.

5.4 PageRank
The theory developed in this section was introduced by Page et al. (1998) and resem-
bles in many ways the recursive propagation idea we have seen in HITS. However,
unlike HITS, only one kind of weight is assigned to Web documents. Intuitively, the
rank of a document should be high if the sum of its parents ranks is high. To a rst
approximation, this intuition might be embodied in the equation
 r(w)
r(v) = , (5.11)
|ch[w]|
wpa[v]

where r(v) is the rank assigned to page v and is a normalization constant. Note
that each parent w contributes by a quantity that is proportional to its rank r(w) but
inversely proportional to its outdegree. This is a fundamental difference with respect
to authority in HITS. The endorsement signal that ows from a given page w to each
of its children decreases as the number of outgoing links (and, therefore, the potential
of being a good hub) increases. Equation (5.11) can also be written in matrix notation,
as
r = Br = Mr. (5.12)
LINK ANALYSIS 135

The matrix B is obtained from the adjacency matrix A of the graph by dividing each
element by the corresponding row sum, i.e.

auv
 , if ch[u]  = ,
buv = w auw (5.13)

a = 0, otherwise.
u,v

As implied by Equation (5.12), the vector r is a right eigenvector of B with an

associated eigenvalue .
An interesting property of this solution can be seen by interpreting the computation
expressed by Equation (5.12) as the description of a random walk through the Web
graph. More precisely, suppose each vertex in the graph is associated with a realization
of a discrete random variable St that models the position of a hypothetical surfer at
a given time t. The rank of a page v could be then thought of as the asymptotic
probability that the surfer is currently browsing that page, i.e. P (St = v). Under this
perspective, M is interpreted as the transition matrix for a rst-order Markov chain:

rt (v) = P (St = v) = P (St = v | St1 = w)P (St1 = w)
w

= mwv rt1 (w). (5.14)
w

This equation updates the probability that our random surfer will browse page v at
time t, given the vector of probabilities at time t 1 and the transition probabilities
mwv . In matrix notation this can be written as

rt = Mrt1 . (5.15)

To satisfy probability axioms, M must be a stochastic matrix, i.e. its rows should sum
to one. Since the rows of B are normalized, the probability axioms are satised if
= 1. This simply means that the random surfer picks one of the outlinks in the page
being visited according to the uniform distribution.
A fundamental question is whether iterating Equation (5.14) converges to some
sensible solution regardless of the initial ranks r0 . To answer this we need to inspect
different cases in the light of the theory of nonnegative matrices developed in Sec-
tion 5.2.
Four interesting cases are illustrated in Figure 5.5. The rst graph has a primitive
adjacency matrix. Therefore Equation (5.5) holds, and values of r(v) corresponding
to the steady state are indicated inside each node (below the node index). In the
same gure, arcs are labeled by the amount of rank that is passed from a node to its
children. As expected, Equation (5.11) holds everywhere. The second graph is more
problematic, since its adjacency matrix is irreducible but not primitive. In this case,
passing ranks from nodes to their children results in a cyclic updating. The random
walk recursion of Equation (5.14) converges in this case to a limit cycle rather than to
a steady state, and the periodicity of the limit cycle is the period of the matrix, or, as
136 PAGERANK

3 4 3 4
1/8
1/8 1/4 abcd cdab

1/8 1/8
1/4

1 1/8 2 1 2
1/4 1/4 bcda dabc

1/8 1/8

5
1/8

(a) (b)

3 1/9 4 3 0 4
1/3 2/9 0 0
1/9
0 0
1/9 2/9 0

1/9 1/3
1 2 1 2
1/9 1/3 0 0

0 0 0
0

5 5
0 0

(c) (d)
Figure 5.5 Rank propagation on graphs with different types of incidence matrices. Equa-
tion (5.14) converges to a nontrivial steady state in case (a) and (c), to a limit cycle in case (b),
and to zero in case (d).

we know, the gcd of the lengths of the cycles (4 in the example). This is indicated in
Figure 5.5b by four values a, b, c, d of rank that cyclically bounce along the nodes. The
third graph of Figure 5.5 has a reducible adjacency matrix. In this case M t converges
to a matrix whose last column is all zero, reecting the fact that the node should have
LINK ANALYSIS 137

zero rank as it has no parents. Finally, the fourth graph has also a reducible adjacency
matrix but this time the maximum eigenvalue is less than one, so M t converges to the
zero matrix. This is due to the existence of a node (1) with no children that effectively
acts as a rank sink.
The situation in Figure 5.5d is of course undesirable but is very common in the
actual Web. Many pages have no outlinks at all. Furthermore, pages that remain on the
crawling frontier and are never fetched will likely produce dangling edges in the graph
that is obtained from crawling. To solve this difculty, observe that the connectivity
of node 1 should be dened as illegal, since it violates the basic hypothesis underlying
the random walk model: the sum of the probabilities of the available actions should
be one in each node, but once in the sink nodes our random surfer would be left with
no choices. A sensible correction consists of giving the random surfer a method of
escape by adding allowable actions. One possibility is to assume that the surfer, who
cannot possibly follow any link, will restart browsing by picking a new Web page
at random. This is the same as adding a link from each sink to each other vertex,
i.e. introducing an escape matrix E dened as evw = 0 if |ch[v]| > 0 and evw = 1/n
otherwise, for each w. Then the transition matrix becomes
M = (B + E).
M is now a stochastic matrix and the Markov chain model for a Web surfer is sound. In
general, however, there is no guarantee that M is also primitive (if there are cycles with
zero outdegree as in Figure 5.1b, these bring irreducible but periodic components).
This difculty will be addressed shortly and for now let us assume that M is primitive.
The following iterative algorithm was suggested in the original paper on PageRank
(Page et al. 1998). It takes as input a nonnegative square matrix M, its size n, and a
tolerance parameter .
PageRank(M, n, )
1 1 [1, . . . , 1] Rn
2 z n1 1
3 x0 z
4 t 0
5 repeat
6 t t +1
7 xt M T xt1
8 dt xt1 1 xt 1
9 xt x1 + dt z
10 xt1 xt 1
11 until < 
12 return xt

The quantity dt is the total rank being lost in sinks. Adding dt z to M T xt1 is basically
a normalization step. As it turns out, if M is a stochastic primitive matrix, then dt = 0
in each iteration (no normalization is necessary) and PageRank converges to the
138 STABILITY

stationary distribution of M. Otherwise, the above algorithm implicitly repairs the

matrix M into a stochastic matrix and converges to the corresponding stationary
distribution (see Exercise 5.4).
Now we address the problem of irreducible but periodic components. These also
act as rank sinks because they never pass rank to other parts of the graph. Moreover,
periodicity may hurt convergence and the algorithm PageRank above is not anymore
guaranteed to terminate. The solution suggested in Page et al. (1998) consists of forc-
ing some source or rank by introducing a static stochastic process that models the
distribution of Web pages that a random surfer periodically jumps to. This distribu-
tion can be any nonnegative vector e such that e1 = 1. The probability distribution
that results from combining the Markovian random walk distribution x and the static
rank source distribution is a mixture model with parameter :

r = e + (1 )x.

The simplest choice for e is a uniform distribution, i.e. e = (1/n)1. Intuitively, this
approach can be motivated by the metaphor that browsing consists of following exist-
ing links with some probability 1 or selecting a nonlinked page with probability .
When the latter choice is made, each page in the entire Web is sampled according to
the probability distribution e. In the case of the uniform distribution, Equation (5.15)
will be rewritten as
rt = [H + (1 )M]T rt1 , (5.16)

where H is a square matrix with huv = 1/n for each u, v. In this way we have obtained
an ergodic Markov chain whose underlying transition graph is fully connected. The
associated transition matrix H + (1 )M is primitive and therefore the sequence rt
converges to the dominant eigenvector. The stationary distribution r associated with
the Markov chain described by Equation (5.16) is known as PageRank. In practice,
is typically chosen to be between 0.1 and 0.2 (Brin and Page 1998).

5.5 Stability
An important question is whether the link analysis algorithms based on eigenvectors
(such as HITS and PageRank) are stable in the sense that results do not change
signicantly as a function of modest variations in the structure of the Web graph. More
precisely, suppose the connectivity of a portion of the graph is changed arbitrarily,
i.e. let G = (V , E) be the graph of interest and let us replace it by a new graph
G = (V , E), where some edges have been added or deleted. How will this affect the
results of algorithms such as HITS and PageRank?
Ng et al. (2001) proved two interesting results about the stability of algorithms
based on the computation of dominant eigenvectors.
LINK ANALYSIS 139

5.5.1 Stability of HITS

First, Ng et al. (2001) derived a bound on the number of hyperlinks that can added
or deleted from one page without signicantly affecting the authority (or hubness)
weights computed by HITS. The bound essentially depends on the eigengap, namely
.
the difference = 1 2 between the two largest eigenvalues of M = AT A (since
this matrix is symmetric the eigengap is a real number).
The result can be formally stated as follows. For every > 0, suppose G is
perturbed by adding or deleting at most k hyperlinks from one page,
 2

k
d+ d , (5.17)
4 + 2
where d is the maximum outdegree of G. The principal eigenvector associated with
the perturbed graph G then satises
a a2
. (5.18)
Moreover, Ng et al. (2001) show that it is possible to perturb a symmetric matrix,
by a quantity that grows as , that produces a constant perturbation of the dominant
eigenvector. Thus, matrices with small eigengap can have low robustness with respect
to perturbation.
In practice, it is not difcult to construct graphs where even adding or deleting
even a single edge results in large variations in the authority and hubness weights. For
example, consider two isolated communities, one possessing a hub h but no emerging
authority (so the authority weight is dispersed among all the nodes), and the other
possessing a well recognized authority a but no important hubs. Adding an edge
from h to a would result in a large change in weight assignments, since the authority
weight of a would be increased at the expense of some amount of authority weights
contributed by all the nodes in the rst community.

5.5.2 Stability of PageRank

In this case suppose r is the stationary distribution associated with matrix
H + (1 )M
(see Equation (5.16)). If the adjacency matrix A is perturbed to a new matrix A, then
Ng et al. (2001) show that

2 j V r(j )
r r  , (5.19)

where V denotes the set of vertices touched by the perturbation. This demonstrates
two interesting facts: rst, the parameter of the mixture model in Equation (5.16)
has a stabilization role; second, if the set of pages affected by the perturbation have a
140 PROBABILISTIC LINK ANALYSIS

5 1h 1a
1
2h 2a
6
3a

2
4h
7
5a
3

4 6h 6a

7h 7a

G U
Figure 5.6 Forming a bipartite graph in SALSA.

small rank, the overall change will also be small. Bianchini et al. (2001) later proved
the tighter bound
1 
r r  2 (j )r(j ), (5.20)

j V

where (j )  2 depends on the edges incident on j affected by the perturbation.

5.6 Probabilistic Link Analysis

The probabilistic interpretation of PageRank based on random walks can be extended
to link analysis algorithms that, like HITS, distinguish the importance of a node as
an authority or as a hub.

5.6.1 SALSA
Lempel and Moran (2001) have proposed a probabilistic extension of HITS called the
Stochastic Approach for Link Structure Analysis (SALSA). Similar extensions have
been proposed independently by Raei and Mendelzon (2000) and Ng et al. (2001).
In all of these proposals, the random walk is carried out by following hyperlinks both
in the forward and in the backward direction.
SALSA starts from a graph G = (V , E) of topically related pages (like the base
subgraph of HITS) and constructs a bipartite undirected graph U = (V , E) as (see
Figure 5.6)
V = Vh Va ,
LINK ANALYSIS 141

where
.
Vh = {vh : v V , ch[v]  = },
.
Va = {va : v V , pa[v]  = },
.
E = {(uh , va ) : (u, v) E}.
The sets Vh and Va are called the hub side and the authority side of U , respectively.
Two separate random walks are then introduced. In the hub walk, each step con-
sists of
(1) following a Web link from a page uh to a page wa , and
(2) immediately afterward following a backlink going back from wa to vh , where
we have assumed that (u, w) E and (v, w) E.
For example, jumping from 1h to 5a and then back from 5a to 2h in Figure 5.6. In
the authority walk, a step consists of following a backlink rst and a forward link
next. In both cases, a step translates into following a path of length exactly two in
U . Note that, by construction, each walk starts on one side of U , either the hub side
or the authority side, and will remain conned to the same side. The Markov chains
associated with the two random walks have transition matrices H and T , respectively,
dened as follows:
 1 1
huv = ,
deg(uh ) deg(wa )
w:(u,w)E,
(v,w)E
 1 1
tuv = .
deg(va ) deg(wh )
w:(w,u)E,
(w,v)E

The hub and authority weights are then obtained as principal eigenvectors of the
matrices H and T . Note that these two matrices could also be dened in an alternative
way. Let A be the adjacency matrix of G, Ar the row-normalized adjacency matrix
(as in Equation (5.13)) and let Ac the column-normalized adjacency matrix of G
(i.e. dividing each nonzero entry by its column sum). Then H consists of the nonzero
rows and columns of Ar ATc , while T consists of the nonzero rows and columns of
ATc Ar .
Note that huv > 0 implies that there exists at least one page w that has links to both
u and v. This is known as co-citation in bibliometrics (Kessler 1963) (see Figure 5.9
and Exercise 5.2). Similarly, tuv > 0 implies there exists at least one page that is
linked to by both u and v, a bibliographic coupling (Small 1973).
Lempel and Moran (2001) showed theoretically that SALSA weights are more
robust that HITS weights in the presence of the Tightly Knit Community (TKC)
Effect. This effect occurs when a small collection of pages (related to a given topic)
is connected so that every hub links to every authority and includes as a special
case the mutual reinforcement effect identied by Bharat and Henzinger (1998) (see
142 PROBABILISTIC LINK ANALYSIS

Section 5.3). It can be shown that the pages in a community connected in this way
can be ranked highly by HITS, higher than pages in a much larger collection where
only some hubs link to some authorities. Clearly the TKC effect could be deliberately
created by spammers interested in pushing the rank of certain websites. Lempel and
Moran (2001) constructed examples of community pairs Cs connected in a TKC
fashion, and Cl sparsely connected, and proved that authorities of Cs are ranked
above the authorities of Cl by HITS but not by SALSA.
In a similar vein, Raei and Mendelzon (2000) and Ng et al. (2001) have proposed
variants of the HITS algorithm based on a random walk model with reset, similar to
the one used by PageRank. More precisely, a random surfer starts at time t = 0 at a
random page and subsequently follows links from the current page with probability
1 , or (s)he jumps to a new random page with probability . Unlike PageRank,
in this model the surfer will follow a forward link on odd steps but a backward link
on even steps. For large t, two stationary distributions result from this random walk,
one for odd values of t, that corresponds to an authority distribution, and one for
even values of t that correspond to a hubness distribution. In vector notation the two
distributions are proportional to

a2t+1 = 1 + (1 )Ar h2t , (5.21)

h2t = 1 + (1 )ATc a2t1 . (5.22)

The stability properties of these ranking distributions are similar to those of PageRank
(Ng et al. 2001).
Some further improvements of HITS and SALSA, as well as theoretical analyses
on the properties of these algorithms can be found in Borodin et al. (2001).

5.6.2 PHITS
Cohn and Chang (2000) point out a different problem with HITS. Since only the
principal eigenvector is extracted, the authority along the remaining eigenvectors is
completely neglected, despite the fact that it could be signicant. An obvious approach
to address this limitation consists of taking into account several eigenvectors of the
co-citation matrix, in the same spirit as PCA is used to extract several factors that
are responsible for variations in multivariate data. As we have discussed in Sec-
tion 4.5.2, however, the statistical assumptions underlying PCA are not sound for
multinomial data such as termdocument occurrences or bibliographical citations.
PHITS can be viewed as probabilistic LSA (see Section 4.5.2) applied to co-citation
and bibliographic coupling matrices. In this case citations replace terms. As in PLSA,
a document d is generated according to a probability distribution P (d) and a latent
variable z is then attached to d with probability P (z | d). Here z could represent
research areas (in the case of bibliographic data) or a (topical) community in the
case of Web documents. Citations (links) are then chosen according to a probability
distribution P (d  | z).
LINK ANALYSIS 143

Figure 5.7 A link farm. Shaded nodes are all copies of the same page.

5.7 Limitations of Link Analysis

Search engines can be spammed by websites faking high relevance with respect to
some topics for the sole purpose of attracting visitors. Marchiori (1997) called this
phenomenon Search Engine Persuasion (SEP). First-generation engines that based
their ranking on classical information retrieval measures were clearly very sensitive
to SEP. Common early techniques used to fool these engines included the use of
inappropriate site titles or descriptions, the inclusion of META keywords in the HTML
code, or even the use of extra text invisible to human surfers. Link analysis was
immediately recognized as a solid defense against SEP. In their seminal paper on
PageRank, Page et al. (1998) stated that
. . . for a page to get a high PageRank, it must convince an important page,
or a lot of non-important pages to link to it. At worst, you can have manip-
ulation in the form of buying advertisements (links) on important sites.
But this seems well under control since it costs money. This immunity
to manipulation is an extremely important property.
In the intervening time period, ranking pages using link analysis has become the stan-
dard approach followed by all the second generation search engines. Consequently,
website owners have realized the enormous importance of maximizing PageRank or
similar link-based scoring functions in order to increase visibility. For some websites,
having high search engine ranking (with respect to some keywords) is so important
that it justies considerable nancial investments. Buying links as a form of adver-
tisement has become reality and perhaps in some cases it is a sensible alternative to
paying search engine companies directly for advertised links.
A link farm is a densely connected Web subgraph articially built for the purpose
of accumulating PageRank (or similar measures of popularity). Link farms can be
built in many ways but link exchange is a common approach. A website owner who
144 LIMITATIONS OF LINK ANALYSIS

W
G in
sink

G G W
core out rest

G W \G

Figure 5.8 Canonical decomposition of the Web relative to a subgraph G.

decides to join the farm agrees to store a copy of a hub page on her server and to
link it from the root of her site. In return, the main URL of her site is added to the
hub page, which is in turn redistributed to the sites participating in the link exchange.
The result is a densely connected subgraph like the one shown in Figure 5.7.
It may appear that, since structures of this kind are highly regular, they should
be relatively easy to detect (see Exercise 5.8) and thus link farming should not be
a serious concern for search engines. However, it is possible to build farms that are
more tightly entangled in the Web and are therefore more difcult to detect by simple
topological analyses. This problem has been recently pointed out by Bianchini et al.
(2001), who have shown that every community, dened as an arbitrary subgraph G
of the Web, must satisfy a special form of energy balance. The overall PageRank
assigned to pages in G grows with the energy that ows in from pages linking to
the community and decreases with the energy dispersed in sinks and passed to pages
outside the community. With reference to Figure 5.8, let Gout denote the subgraph
of G induced by pages that contain hyperlinks pointing outside to G and let Gsink
denote the sink subgraph of G. Also, let Win be the subgraph induced by the pages
outside G that link to pages in G. Then the equation
rG  = |G| + EG
in
EG
sink
EG
out
. (5.23)
holds, where |G| is the default energy that is assigned to the community,
1 
EGin
= fG (w) r(w)

wWin

is the energy owing in from outside, where fG (w) is the fraction of links in w that
point to pages in G,
1 
EGout
= (1 fG (w)) r(w)

wGout

is the energy owing out to pages outside the community, and

1 
sink
EG = r(w).

wGsink
LINK ANALYSIS 145

Suppose a set of sponsoring pages S is generated articially to boost the rank

of a target website. The above energy balance analysis shows that the PageRank of
the target can be increased linearly with |S| regardless of the topology in S, making
it difcult to detect the origin of spam using methods that are only based on graph
topology. The fact that spamming activities of this kind are indeed possible is borne
out, for example, by some recent anecdotes such as the popularity battle of the Church
of Scientology against its main opponent Xenu.net. 2
Websites that allow their users to post HTML code, for example, Weblogs (Walker
2002), potentially offer a cheap way for constructing articial rank-boosting commu-
nities. Specialized algorithms that exploit more information than just topology are
likely to be needed in the near future in order to prevent these spamming activities.
This may also help prevent the future topology of the Web (and associated notions of
popularity) from becoming signicantly controlled by economic interests.
One example in this direction is the paper by Davison (2000a) that describes a
machine-learning method for the automatic discrimination of links that are unrelated
to the intrinsic merit of the pages. In this case, hyperlinks are described by several
binary features that include, for example, tests about the structure of the URL, identity
of source and target host or domain, or the total number of outlinks found on the source
page.
Finally, we note that assessing the quality of pages returned by a search engine
is difcult because quality is ultimately dened by human judgement. Amento et al.
(2000) have reported an empirical study in which 16 human experts in ve popular
topics related to TV entertainment and music were asked to rank the quality of a
set of Web pages. The experiment was aimed at testing whether expert judgements
were correlated with scores based on link analysis. The study revealed that ranking
documents according to the authority score (as computed by HITS) or according to
PageRank yields high precision for the top 5 or 10 ranked documents. However, it
was also found that alternative metrics such as page indegree or total number of pages
in the website perform equally well.

Exercises
Exercise 5.1. Draw a graph of reasonable size, connecting vertices at random, and
compute the principal eigenvectors of the matrices AT A and AAT to get authority and
hubness weights. A very rapid way of doing this is by using linear algebra software
such as Octave. Now select a vertex having nonzero indegree but small authority and
try to modify the graph to increase its authority without increasing its indegree nor
the indegree of its parents.
Exercise 5.2. The two matrices involved in Equations (5.8) and (5.9) were introduced
several years before in the eld of bibliometrics. In particular, C = AT A is known as
the co-citation matrix (Kessler 1963) and B = AAT is known as the bibliographic
2 See http://www.operatingthetan.com/google/ for details.
146 LIMITATIONS OF LINK ANALYSIS

Co-citations (Small, 1973)

Bibliographic coupling (Kessler, 1963)

Figure 5.9 Co-citations and bibliographic coupling.

0.25
0.6
0.1
1
0.1 2

0.4 0.65
0.3

0.5

0.2 3

Figure 5.10 Markov chain for Exercise 5.3.

coupling matrix (Small 1973). Show that cuv is the number of documents that cite
both documents u and v, while buv is the number of pages that are cited by both u
and v (see Figure 5.9).
Exercise 5.3. Consider the Markov chain in Figure 5.10 (where arcs are labeled by
transition probabilities). Is it ergodic? What is the steady-state distribution?
LINK ANALYSIS 147

Exercise 5.4. Let N be the n n row normalized adjacency matrix of a graph as in

Section 5.4. Suppose N is not stochastic and let R be any matrix such that N + R
is a stochastic matrix. Show that PageRank(N , n, ) and PageRank(N + R, n, )
converge to the same solution.
Exercise 5.5. In Equation (5.21), the stationary distributions of authority and hubness
for randomized HITS are dened within a proportionality constant. Determine what
this constant is.
Exercise 5.6. Use the stability results in Section 5.5.2 to estimate how often you
should crawl the Web (and recompute PageRank) in order to guarantee that

r r < ,
 being an assigned tolerance. Assume for simplicity that a constant number of pages
are changed in a given unit of time.
Exercise 5.7. Implement the PageRank computation and simulate the results on a
relatively large articial graph (build the graph using ideas from Chapter 3). Then
introduce link farms in your graph and study the effect they have on the PageRank
vector as a function of the number and the size of the farms.
Exercise 5.8. Propose an efcient algorithm to detect link farms structured as in
Figure 5.7 in a large graph.
This Page Intentionally Left Blank
6

Advanced Crawling Techniques

In this chapter we rene the basic design of Web crawlers, building on the material
presented in Section 2.5.

6.1 Selective Crawling

The search engine coverage experiments reported in Section 2.5.2 suggest that no
crawler can realistically harvest the entire Web. An important fact is that the time
required to complete a large crawl can be signicant, independent of whatever tech-
nology is available at the site where the search engines operate. Moreover, fetching
and indexing a larger set of documents has signicant implications on the scalability
of the overall system and, consequently, on the cost of the required hardware and
maintenance services. Thus, to optimize available resources, a crawler should ideally
be capable of recognizing the relevance or the importance of sites or pages, and limit-
ing fetching to the most important subset of pages that can be downloaded in a given
amount of time.
Relevance can be estimated with respect to several different criteria. To formalize
the intuitive idea of selective crawling, we need to introduce, for each URL u, a scoring
( )
function s (u), with respect to some underlying relevance criterion and parameters
. In the simplest case we can assume a Boolean relevance function, i.e. s(u) = 1
if the document pointed to by u is relevant and s(u) = 0 if the document is not
interesting. More generally, we can think of s(d) as a real-valued function, such as
the conditional probability that the document belongs to a certain category given its
contents. Note that in all of these cases the scoring function depends only on the URL
and the criterion, but does not depend on the state of the crawler. Moreover, we will
assume that the scoring function does not depend on time.
One general approach for building selective crawlers consists of changing the policy
of insertions and extractions in the queue of discovered URLs. For simplicity, let us
consider again the algorithm Simple-Crawler on page 47. Suppose now that URLs
in Q are sorted according to the value returned by s(u). In this way we obtain a
best-rst search strategy (see, for example, Russell and Norvig 1995) such that URLs
Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
150 SELECTIVE CRAWLING
rt /t qt (G ) / t
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

1N 2N 3N 4N N 1N 2N 3N 4N N
5 5 5 5 5 5 5 5

Figure 6.1 Efciency curves obtained by Cho et al. (1998) for a selective crawler. Both were
obtained on a set of 784 592 Stanford University URLs. Dotted lines correspond to a BFS
crawler, solid lines to a crawler guided by the estimated number of backlinks, and dashed lines
to a crawler guided by estimated PageRank. The diagonal lines are the reference performance
of a random crawler. The target measure of relevance is the actual number of backlinks. In the
right-hand plot, the importance target G is 100 backlinks.

having higher score are fetched rst. If s(u) provides a good model of the relevance
of the document pointed to by the URL, then we expect that the search process will
be guided toward the most relevant regions of the Web. From this perspective the
problem is to dene interesting scoring functions and to devise good algorithms to
compute them.
In the following we consider in detail some specic examples of scoring functions.

Depth. As we have seen in Chapter 3, the distribution of the number of pages on

different Web servers generally follows a power law. Thus, many sites have few
pages and, at the same time, there is a small but signicant fraction of sites con-
taining a huge number of pages. A very simple strategy for maximizing coverage
breadth consists of limiting the total number of documents downloaded from a
single site, for example, by setting a threshold, or by keeping track of the depth in
the site directory tree, or by limiting the acceptable length of the path from the site
homepage to the document,

(depth) 1, if |root(u)  u| < ,
s (u) = (6.1)
0, otherwise,

where root(u) is the root of the site containing u. The rationale behind this approach
is that by maximizing breadth it will be easy for the end-user to eventually nd the
desired information. For example a page that is not indexed by the search engine
may be easily reachable from other pages in the same site.
Popularity. It is often the case that we can dene criteria according to which certain
pages are more important than others. For example, search engine queries are
answered by proposing to the user a sorted list of documents. Typically, users tend
to inspect only the few rst documents in the list (in Chapter 8 we will discuss
ADVANCED CRAWLING TECHNIQUES 151

empirical data that support this statement). Thus, if a document rarely appears near
the top of any lists, it may be worthless to download it and its importance should be
small. A simple way of assigning importance to documents that are more popular
is to introduce a relevance function based on the number of backlinks

1, if indegree(u) > ,
s(backlinks)
(u) = (6.2)
0, otherwise,

where is an assigned threshold.

(backlinks)
Clearly s (u) can be computed exactly only if we have already crawled the
entire Web. In practice we can use an approximated value from the partial subgraph
obtained in a previous crawl, or we can incrementally update the score as soon as
new parents of u are encountered during the crawl.

PageRank. As discussed in Section 5.4, PageRank is a measure of popularity that

recursively assigns each link a weight that is proportional to the popularity of
the source document. It may be seen as a renement of the indegree measure
dened above in Equation (6.2). Also PageRank needs to be estimated using partial
knowledge of the graph that is rened during crawling.

The efciency of a selective crawler can be measured by comparing it to an ideal

crawler that fetches documents in order of relevance (whatever the relevance function).
Suppose we are given a website with N pages and for n = 1, . . . , N let sn denote
the score of the page that is ranked in position n by the relevance function. Suppose
at some point in time the crawler has fetched t pages and let rt denote the number
of fetched pages whose score is higher than st . In the ideal crawler we would have
rt = t and the efciency curve rt /t would be constant and equal to one. In a random
crawler, only a fraction t/N of the fetched pages will have a score higher than st ,
i.e. rt /2 = t/N . Selective crawlers should obtain intermediate efciency curves.
Alternatively, one may assign a xed importance target, G, and consider as relevant
all the pages whose score is higher than G. For example, we might be interested in
quickly fetching pages having at least 20 backlinks. The efciency in this case may be
measured as qt (G)/t, where qt is the number of fetched pages having a score higher
than G. Cho et al. (1998) report several experimental results comparing the above
strategies on a controlled data set of Stanford University pages. Figure 6.1 shows
some of the results when the target measure of relevance is the number of backlinks.
Interestingly, the estimated PageRank outperforms the estimated number of backlinks
as a guiding criterion.
In a more recent study by Najork and Wiener (2001) it was found that traversing the
Web in breadth-rst order is a good strategy and allows us to download hot pages
rst. The study was conducted on a large set of 328 million pages from 7 million
distinct servers.
152 FOCUSED CRAWLING
0.5
Random
Sibling
0.4 Linked, same domain
Linked
Linked, different domains
0.3

0.2

0.1

0.0
Mean Mean Term Mean
TFIDF Probability Overlap

Figure 6.2 Results of the study carried out by Davison in 2000. (a) Three indicators that
measure text similarity (see Section 4.3) are compared in ve different linkage contexts. In
this data set, about 56% of the linked documents belong to the same domain. It can be seen
that the similarity between linked documents is signicantly higher than the similarity between
random documents. (b) Similarities measured between anchor text and text of documents in
ve different contexts.

6.2 Focused Crawling

In many cases a search engine does not need to be completely general purpose and
cover every possible site on the Web, but instead can be focused on particular topics
that are of interest to a specic community of users. In these cases, relevance can be
dened with respect to the expected utility of the pages for users of the search engine.
For example, a crawler that provides information to a vertical portal specialized in,
say, music, could be quite safely programmed to ignore sites with content in different
topics such as health or sports. A focused crawler is a rened selective crawler that
searches for information related to certain topics rather than being driven by generic
quality measures. The fraction of sites that specialize in a particular topic may be
small enough to enable a focused crawler to download them almost exhaustively and
in a relatively short time.

6.2.1 Focused crawling by relevance prediction

The URL ordering technique suggested for selective crawlers can be also extended
to focused crawlers. In this case, it is necessary to determine whether a document is
relevant or not to the topic of interest in order to dene a scoring function for driving
the search process. One way to do this is to use the text categorization techniques
presented in Section 4.6. More precisely, if we denote by c the topic of interest, a
score can be computed as the conditional probability that the document is relevant,
given the text in the document:
(topic)
s (u) = P (c | d(u), ). (6.3)
ADVANCED CRAWLING TECHNIQUES 153
Cycling domain AIDS/HIV domain
Average Average
relevance relevance
(unfocused) (unfocused)

0.4 0.4

0.2 0.2

0 0
0 2000 4000 6000 8000 10000 Documents 0 1000 2000 3000 4000 5000 Documents
fetched fetched
Average Average
relevance relevance
(soft focused) (soft focused)

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 1000 2000 3000 4000 5000 Documents 0 1000 2000 3000 4000 Documents
fetched fetched

Figure 6.3 Efciency of a BFS and a focused crawler compared in the study by Chakrabarti
et al. (1999b). The plots show the average relevance of fetched documents versus the number
of fetched documents. Results for two topics are shown (cycling on the left, AIDS/HIV on the
right).

Here represents the adjustable parameters of the classier. An obvious consideration

is that the score cannot be computed exactly (we have not yet downloaded the docu-
ment pointed to by u). There are a number of different strategies for approximating
the topic score.

Parent based. In this case we compute the score for a fetched document and extend
the computed value to all the URLs in that document. More precisely, if v is a
parent of u, we approximate the score of u as
(topic)
s (u)  P (c | d(v), ). (6.4)

The rationale is a general principle of topic locality. If a page deals with, say,
music, it may be reasonable to believe that most of the outlinks of that page will deal
with music as well. In a systematic study based on the analysis of about 200 000 doc-
uments, Davison (2000b) found that topic locality is measurably present in the Web,
under different text similarity measures (see Figure 6.2a).
Chakrabarti et al. (1999b) use a hierarchical classier and suggest two implemen-
tations of the parent-based scoring approach. In hard focusing, they check if at least
one node in the category path of d(v) is associated with a topic of interest; if not
154 FOCUSED CRAWLING

Layer 2

Layer 1

Figure 6.4 Example of a two-layered context graph. The central white node
is a target document. Adapted from Diligenti et al. (2000).

the outlinks of v are simply discarded (not even inserted in the crawling queue Q).
In soft focusing, relevance is computed according to Equation (6.4), but if the same
URL is discovered from multiple parents, a revision strategy is needed in order to
(topic)
update s (u) when new evidence is collected. Chakrabarti et al. (1999b) found
no signicant difference between the two approaches.
Anchor based. Instead of the entire parent document, d(v), we can just use the text
d(v, u) in the anchor(s) where the link to u is referred to, as the anchor text is often
very informative about the contents of the document pointed to by the corresponding
URL. This semantic linkagewas also quantied by Davison (2000b), who showed
that the anchor text is most similar to the page it references (see Figure 6.2b).

To better illustrate the behavior of a focused crawler on real data, consider the
efciency diagrams in Figure 6.3 that summarize some results obtained by Chakrabarti
et al. (1999b). An unfocused crawler starting from a seed set of pages that are relevant
to a given topic will soon begin to explore irrelevant regions of the Web. As shown in
the top diagrams, the average relevance of the fetched pages dramatically decreases
as crawling goes on. Using a focused crawler (in this case, relevance is predicted
by a Naive Bayes classier trained on examples of relevant documents) allows us to
maintain an almost steady level of average relevance.
There are several alternatives to a focused crawler based on a single best-rst queue,
as detailed in the following.

6.2.2 Context graphs

Diligenti et al. (2000) suggested a strategy that takes advantage of knowledge of
the Internet topology to train a machine-learning system to predict how far some
ADVANCED CRAWLING TECHNIQUES 155

relevant information can be expected to be found starting from a given page. Intu-
itively, suppose the crawler is programmed to gather homepages of academic courses
in articial intelligence. The backlinks of these pages are likely to lead to professors
home pages or to the teaching sections of the department site. Going one step further,
backlinks of backlinks are likely to lead into higher level sections of department sites
(such as those containing lists of the faculty). More precisely, the context graph of a
node u (see Figure 6.4) is the layered graph formed inductively as follows. Layer 0
contains node u. Layer i contains all the parents of all the nodes in layer i 1. No
edges jump across layers. Starting from a given set of relevant pages, Diligenti et al.
(2000) used context graphs to construct a data set of documents whose distance from
the relevant target was known (backlinks were obtained by querying general purpose
search engines). After training, the machine-learning system predicts the layer a new
document belongs to, which indicates how many links need to be followed before rel-
evant information will be reached, or it returns other to indicate that the document
and its near descendants are all irrelevant. Denoting by n the depth of the considered
context graph, the crawler uses n best-rst queues, one for each layer, plus one extra
queue for documents of class other. This latter queue is initialized with the seeds.
In the main loop, the crawler extracts URLs from the rst nonempty queue and in this
manner favors those that are more likely to rapidly lead to relevant information.

6.2.3 Reinforcement learning

Reinforcement learning (see, for example, Sutton and Barto 1998) is a framework for
deciding in an optimal way what actions an agent operating in a discrete environment
should take in order to maximize its expected future reward. Compared to supervised
learning, reinforcement learning is characterized by the absence of external supervi-
sion. The agent learns from the received rewards (or punishments). More formally,
the discrete-time environment is characterized by a nite set of states S. At time t
the environment perceived by the agent is in state st S and, correspondingly, the
agent has access to a nite set of actions A(st ). After performing action at A(st )
the state transitions to st+1 and the agent receives an immediate reward rt+1 . A policy
describes the joint probability distribution on states and actions, i.e. (s, a) is the
probability of taking action a when in state s. An optimal policy should maximize the
expected rewards received by the agent over time. In order to specify what an optimal
policy is, we rst dene the value of state s under a policy as


V (s) = E

rt+k+1 | st = s .
k
(6.5)
k=0

Similarly, we can dene the action-value function as



Q (s, a) = E

rt+k+1 | st = s, at = a .
k
(6.6)
k=0
156 FOCUSED CRAWLING

An optimal policy maximizes the value function over all the states: V (s) 
V (s) for all s S. According to the Bellman optimality principle, underlying
the foundations of dynamic programming (Bellman 1957), a sequence of optimal
decisions has the property that, regardless of the action taken at the initial time, the
subsequent sequence of decisions must be optimal with respect of the outcome of the
rst action. This translates into

V (s) = max E[rt+1 + V (st+1 ) | st = s, at = a], (6.7)

aA(s)

which allows us to determine the optimal policy once V is known for all s. An
optimal policy also maximizes Q (s, a) for all s S, a A(s):

Q (s, a) = E[rt+1 + max Q (st+1 , b) | st = s, at = a]. (6.8)

b

The advantage of the state-action representation is that once Q is known for all s
and a, all the agent needs to do in order to maximize the expected future reward is to
choose the action that maximizes Q (s, a).
In the context of focused Web search, immediate rewards are obtained (downloading
relevant documents) and the policy learned by reinforcement can be used to guide
the agent (the crawler) toward high long-term cumulative rewards. As we have seen
in Section 2.5.3, the internal state of a crawler is basically described by the sets of
fetched and discovered URLs. Actions correspond to fetching a particular URL that
belongs to the queue of discovered URLs. Even if we simplify the representation by
removing from the Web all the off-topic documents, it is clear that the sets of states and
actions are overwhelming for a standard implementation of reinforcement learning.
LASER (Boyan et al. 1996) was one of the rst proposals to combine reinforcement
learning ideas with Web search engines. The aim in LASER is to answer queries
rather than to crawl the Web. The system begins by computing a relevance score
r0 (u) = TFIDF(d(u), q) and then propagates it in the Web graph using the recurrence
 rt (v)
rt+1 (u) = ro (u) + , (6.9)
|pa[u]|
vpa[u]

which is iterated until convergence for each document in the collection, where and
 are free parameters. After convergence, documents at distance k from u provide a
contribution proportional to k times their relevance to the relevance of u.
McCallum et al. (2000c) used reinforcement to search computer science papers
in academic websites. In order to simplify the problem of learning Q (s, a) for an
enormous number of states and actions they propose the following two assumptions:
the state is independent of the relevant documents that have been fetched so
far;
actions can only be distinguished by means of the words in the neighborhood
of the hyperlink that correspond to each action (e.g. the anchor text).
ADVANCED CRAWLING TECHNIQUES 157

In this way, learning Q reduces to learning a mapping from text (e.g. bag of words
representing anchors) to a real number (the expected future reward).

6.2.4 Related intelligent Web agents

The Fish algorithm by De Bra and Post (1994) uses a population of agents that
autonomously spider the Web. Like shes in an information sea, agents accumulate
energy as they collect relevant documents for a given query while they consume
energy for using network resources. Since agents need energy to survive, those which
end up exploring irrelevant portions of the Web perish, implementing a selection
mechanism. The Shark algorithm by Hersovici et al. (1998) improves the Fish search
by introducing a real-valued relevance score that also depends on anchor text and a
discounted fraction of the score that was given to ancestor pages. One of the main
limitations of these approaches is the lack of adaptation.
WebWatcher (Armstrong et al. 1995) is an interactive recommendation system that
assists a user during browsing. It introduced the use of machine learning for predicting
the best hyperlink to follow according to a given user goal. A similar approach based
on heuristic searching was proposed in Lieberman (1995). Arachnid (Menczer 1997)
is based on a distributed population of adaptive agents that search information related
to user-provided keywords. Similar in spirit to the Fish algorithm, Arachnid agents
receive energy by relevance feedback provided by the user. An important feature of
this approach is that hyperlinks are selected by a neural network associated with the
agent, whose weights are adjusted by reinforcement learning. An extension of this
system is described in Menczer and Belew (2000).
A more recent framework that generalizes focused crawling is intelligent crawl-
ing (Aggarwal et al. 2001). In this case, the crawler does not need a collection of
topical examples for training. Users describe their information needs by means of
predicates, i.e. general specications that generalize keyword-based queries to also
include document-to-document similarity queries, or characterizations of a topic as
obtained from a text classier. The intelligent crawler proposed in Aggarwal et al.
(2001) is capable of auto-focusing using documents that satisfy the query predicate.
CiteSeer (http://citeseer.nj.nec.com/cs) is a search engine focused
on computer science literature (Bollacker et al. 1998; Lawrence et al. 1999). It fetches
PostScript or PDF papers from paper repositories (for example, a researchers Web
pages) and autonomously builds a citation index. As a major difference with respect to
the system described by McCallum et al. (2000c), documents are located by querying
traditional Web search engines and by collecting submitted URLs of pages containing
links to research articles. The system performs several information extraction opera-
tion from the documents and, in particular, it extracts citations and references in the
body of a paper in order to build a scientic literature web. Having represented the
collection of papers as a web, CiteSeer can also apply link analysis algorithm and
compute for example authority and hubness weights of each paper. Link analysis in
this case can also be used to determine which documents are related to any given one.
158 DISTRIBUTED CRAWLING

Interestingly, since CiteSeer creates a Web page for each online article, it effectively
maps the online subset of the computer science literature web to a subset of the World
Wide Web. A recent study has shown that papers that are available online tend to
receive a signicantly higher number of citations (Lawrence 2001).
The DEADLINER system described in Kruger et al. (2000) is a search engine
specialized in conference and workshop announcements. One of the input components
of the system is a context-graph focused crawler (see Section 6.2.2) that gathers
potentially related Web documents. An SVM text classier is subsequently used to
rene the set of documents retrieved by the focused crawler.

6.3 Distributed Crawling

A single crawling process, even if multithreading is used, will typically be insufcient
for large-scale engines that need to fetch large amounts of data rapidly. When using
a single centralized crawler, all the fetched data must pass through a single physical
link. This is often problematic, regardless of the bandwidth available at the crawling
center, because it will be unlikely to have comparable connection speeds from different
geographical regions. Distributing the crawling activity via multiple processes can
be seen as a form of divide and conquer that can help build a scalable system.
Splitting the load decreases hardware requirements and at the same time increases the
overall download speed and reliability if separate processes access the Internet through
different physical links. The advantage is particularly evident if separate crawlers run
in separate data centers that are located in different countries or continents. However,
while the physical links reect geographical neighborhoods, we know that the edges
of the Web graph are instead associated with communitiesthat can and often do cross
geographical borders. Hence, running separate and independent crawling processes
can result in a signicant overlap among the collections of fetched documents. The
performance of a parallelization approach can be measured in terms of
communication overhead the fraction of bandwidth spent to coordinate the
activity of the separate processes, with respect to the bandwidth usefully spent
for document fetching;
overlap the fraction of duplicate documents downloaded by all the processes;
coverage the fraction of documents reachable from the seeds that are actually
downloaded; and
quality e.g. some of the scoring functions dened in Section 6.1 depend on the
link structure and this information can be partially lost using separate crawling
processes.
The literature on this important topic is not abundant. In a recent study, Cho and
Garcia-Molina (2002) have dened a framework based on several dimensions that
characterize the interaction among a set of crawlers.
ADVANCED CRAWLING TECHNIQUES 159

Si Sj

Aii Ajj

Aii Aij Ajj Aji

Aij Aji

Crawler i Crawler j
Figure 6.5 Two crawlers statically coordinated.

Coordination refers to the way different processes agree about the subset of pages
each of them should be responsible for. If two crawling processes i and j are
completely independent (not coordinated), then the degree of overlap can only
be controlled by having different seeds Si and Sj . If we assume the validity of
topological models such as those presented in Section 3, then we can expect that
the overlap will eventually become signicant unless a partition of the seed set is
properly chosen. On the other hand, making a good choice is a challenge, since
the partition that minimizes overlap may be difcult to compute, for example,
because current models of the Web are not accurate enough. In addition it may be
suboptimal with respect to other desiderata that motivated distributed crawling in
the rst place, such as distributing the load and scaling-up.
A pool of crawlers can be coordinated by partitioning the Web into a number of
subgraphs and letting each crawler be mainly responsible for fetching documents
from its own subgraph. If the partition is decided before crawling begins and not
changed thereafter, we refer to this as static coordination. This option has the great
advantage of simplicity and implies little communication overhead. Alternatively,
if the partition is modied during the crawling process, the coordination is said to
be dynamic. In the static approach the crawling processes, once started, can be seen
as agents that operate in a relatively autonomous way. In contrast, in the dynamic
case each process is subject to a reassignment policy that must be controlled by an
external supervisor.
Connement species, assuming statically coordinated crawlers, how strictly each
crawler should operate within its own partition. Consider two processes, i and
160 WEB DYNAMICS

j , and let Aij denote the set of documents belonging to partition i that can be
reached from the seeds Sj (see Figure 6.5). The question is what should happen
when crawler i pops from its queue foreign URLs pointing to nodes in a different
partition. Cho and Garcia-Molina (2002) suggest three possible modes: rewall,
crossover, and exchange. In rewall mode, each process remains strictly within its
partition and never follows interpartition links. In crossover mode, a process can
follow interpartition links when its queue does not contain any more URLs in its
own partition. In exchange mode, a process never follows interpartition links, but
it can periodically open a communication channel to dispatch the foreign URLs
it has encountered to the processes that operate in those partitions. To see how
these modes affect performance measures, consider Figure 6.5 again. The rewall
mode has, by denition, zero overlap but can be expected to have poor coverage,
since documents in Aij \ Aii are never fetched (for all i and j ). Cross-over mode
may achieve good coverage but can have potentially high overlap. For example,
documents in Aii Aij can be fetched by both process i and j . The exchange mode
has no overlap and can achieve perfect coverage. However, while the rst two modes
do not require extra bandwidth, in exchange mode there will be a communication
overhead.

Partitioning denes the strategy employed to split URLs into non-overlapping sub-
sets that are then assigned to each process.A straightforward approach is to compute
a hash function of the IP address in the URL, i.e. if n {0, . . . , 232 1} is the inte-
ger corresponding to the IP address and m the number of processes, documents such
that n mod m = i are assigned to process i. In practice, a more sophisticated solu-
tion would take into account the geographical dislocation of networks, which can
be inferred from the IP address by querying Whois databases such as the Rseaux
IP Europens (RIPE) or the American Registry for Internet Numbers (ARIN).

6.4 Web Dynamics

In the nal section of this chapter we address the question of how information on the
Web changes over time. Knowledge of this rate of change is crucial, as it allows us
to estimate how often a search engine should visit each portion of the Web in order
to maintain a fresh index.
In Chapter 3 we discussed general aging properties of Web graphs. A precise
notion of recency in this context has been proposed by Brewington and Cybenko
(2000). The index entry for a certain document, indexed at time t0 , is said to be -
current at time t if the document has not changed in the time interval between t0 and
t . Basically is a grace period: if we pretend that the user query was made
time units ago rather than now, then the information in the search engine would be up
to date. A search engine for a given collection of documents is said to be (, )-current
if the probability that a document is -current is at least . According to this denition,
we can ask interesting questions like how many documents per day should a search
ADVANCED CRAWLING TECHNIQUES 161

engine refresh in order to guarantee it will remain (0.9,1 week)-current? Answering

this question requires that we develop a probabilistic model of Web dynamics. The
model will be complicated because of two concomitant factors: pages change over
time, and the Web itself grows and evolves in time (see Chapter 3).

6.4.1 Lifetime and aging of documents

To begin with consider a single document. The model we are going to develop is based
on the same ideas underpinning reliability theory in industrial engineering (see, for
example, Barlow and Proshan 1975). Let T be a continuous random variable repre-
senting the lifetime of a component in a piece of machinery or equipment. Assuming
the component was initially installed or replaced at time zero, lifetime is the time
when the component breaks down (dies). Let F (t) be the cumulative distribution
function (cdf) of lifetime. The reliability (or survivorship function) is dened as
.
S(t) = 1 F (t) = P (T > t), (6.10)
i.e. the probability that the component will be functioning at time t. A variable closely
related to lifetime is the age of a component, i.e. the time elapsed since the last
replacement (see Figure 6.6 to better understand the relationship between lifetime
and age). Its cdf, dened as G(t) = P (age < t), is obtained by integrating the
survivorship function and normalizing:

t
0 S( ) d
G(t) =
. (6.11)
0 S( ) d

Note that G(t) is the expected fraction of components that are still operating at
time t. The age probability density function (pdf) g(t) is thus proportional to the
survivorship function. Returning from the reliability metaphor to Web documents,
S(t) is the probability that a document that was last changed at time zero will remain
unmodied at time t, while G(t) is the expected fraction of documents that are older
than t. The probability that the document will be modied before an additional time
h has passed is expressed by the conditional probability P (t < T
t + s | T > t).
The change rate (t) (also known as the hazard rate in reliability theory, or mortality
force in demography) is then obtained by dividing by h and taking the limit for small
h,
. 1
(t) = lim P (t < T
t + h | T > t)
h0 s

1 1 t+h f (t)
= lim f ( ) d = , (6.12)
h0 S(t) h t S(t)
where f (t) denotes the lifetime pdf. Combining (6.10) and (6.12) we have the ordinary
differential equation (see, for example, Apostol 1969)
F  (t) = (t)(1 F (t)), (6.13)
162 WEB DYNAMICS
Observed
lifetime
age(t )

Actual
lifetime

age(t )

Observation
timespan

Figure 6.6 Sampling lifetimes can be problematic as changes can be missed for two reasons.
Top, two consecutive changes are missed and the observed lifetime is overestimated. Bottom,
the observation time-span must be large enough to catch changes that occur in a long range.
Sampling instants are marked by double arrowheads.

with F (0) = 0. We will assume that changes happen randomly and independently.
According to a Poisson process, the probability of a change event at any given time
is independent of when the last change happened (see Appendix A). For a constant
change rate (t) = , the solution of Equation (6.13) is

F (t) = 1 et , f (t) = et .

Brewington and Cybenko (2000) observed that the model could be particularly
valuable for analyzing Web documents. In practice, however, the estimation of f (t)
is problematic for any method based on sampling. If a document is observed at two
different instants t1 and t2 , we can check for differences in the document but we can-
not know how many times the document was changed in the interval [t1 , t2 ], a phe-
ADVANCED CRAWLING TECHNIQUES 163
1
Observed pdf Observed pdf
of page age 0.9 of page age
2
10 0.8
0.7

3 0.6
10
0.5
0.4
4 0.3
10
0.2
0.1
5
10 2 3
0 200 400 600 800 1000 1200 1400 1 10 10 10
Age in days Age in days

2 Observed pdf Observed cdf

10 1
of page lifetime of page lifetime
0.9
0.8

3
0.7
10
0.6
0.5
0.4
4
10 0.3
0.2
0.1

2 3
0 100 200 300 400 500 600 1 10 10 10
Lifetime in days Lifetime in days

Figure 6.7 Empirical distributions of page age (top) and page lifetime (bottom) on a set of
7 million Web pages. Adapted from Brewington and Cybenko (2000).

nomenon known as aliasing (see Figure 6.6). On the other hand, the age of a document
may be readily available, if the Web server correctly returns the Last-Modified
timestamp in the HTTP header (see Section 2.3.4 for a sample script that obtains this
information from a Web server). Sampling document ages is not subject to aliasing
and lifetime can be obtained indirectly via Equation (6.11). In particular, if the change
rate is constant, it is easy to see that the denominator in Equation (6.11) is a/ and
thus from Equation (6.12) we obtain
g(t) = f (t) = et . (6.14)
In other words, assuming a constant change rate, it is possible to estimate lifetime
from observed age.
This simple model, however, does not capture the essential property that the Web
is growing with time. Brewington and Cybenko (2000) collected a large data set of
roughly 7 million Web pages, observed between 1999 and 2000 in a time period
of seven months, while operating a service called The Informant. 1 The resulting
distributions of age and lifetime are reported in Figure 6.7.
1 Originally http://informant.dartmouth.edu, now http://www.tracerlock.com.
164 WEB DYNAMICS

P ()
P0 e

P0 e(t)

P0

t
Figure 6.8 Assuming an exponentially growing Web whose pages are never changed,
it follows that the age distribution is also exponential at any time .

What is immediately evident is that most of the collected Web documents are
young, but what is interesting is why this is the case. Suppose that growth is modeled
by an exponential distribution, namely that the size at time is P ( ) = P0 e , where
P0 is the size of the initial population and the growth rate. If documents were
created and never edited, their age would be simply the time since their creation. The
probability that a random document at time has an age less than t is

new docs in ( t, ) e e( t)
Gg (t, ) = = (6.15)
all docs e 1
(see Figure 6.8) and thus the resulting age density at time is

e( t)
gg (t, ) = [H (t) H (t )], (6.16)
e 1
where H (t) is the Heaviside step function, i.e. H (t) = 0 for t < 0 and H (t) = 1
otherwise.
In other words, this trivial growth model yields an exponential age distribution,
like the model in Section 6.4.1 that assumed a static Web with documents refreshed
at a constant rate. Clearly, the effects of both Web growth and document refreshing
should be taken into account in order to obtain a realistic model. Brewington and
Cybenko (2000) experimented with a hybrid model (not reported here) that combines
an exponential growth model and exponential change rate. Fitting this model with
ages obtained from timestamps, they estimated = 0.001 76 (in units of days1 ).
This estimate corresponds to the Web size doubling in about 394 days. In contrast,
if we estimated from the lower bounds of 320 million pages in December 1997
(Lawrence and Giles 1998b) and 800 million pages in February 1999 (Lawrence and
Giles 1999), roughly 426 days later, we would obtain = 0.022, or a doubling time
ADVANCED CRAWLING TECHNIQUES 165

of 315 days. Despite the differences in these two estimates, they are of the same order
of magnitude and give us some idea of the growth rate of the Web.
Another important problem when using empirically measured ages is that servers
often do not return meaningful timestamps. This is particularly true in the case of
highly dynamic Web pages. For example, it is not always possible to assess from the
timestamp whether a document was just edited or was generated on-the-y by the
server. These considerations suggest that the estimation of change rates should be
carried out using lifetimes rather than ages. The main difculty is how to deal with
the problem of potentially poorly chosen timespans, as exemplied in Figure 6.6.
Brewington and Cybenko (2000) suggest a model that explicitly takes into account
the probability of observing a change, given the change rate and the timespan. Their
model is based on the following assumptions.
Document changes are events controlled by an underlying Poisson process,
where the probability of observing a change at any given time does not depend
on previous changes. Given the timespan and the change rate , the probability
that we observe one change (given that it actually was made) is therefore
P (c | , ) = 1 e . (6.17)
It should be observed that, in their study, Brewington and Cybenko (2000)
found that pages are changed with different probabilities at different hours of
the day or during different days of the week (most changes being concentrated
during ofce hours). Nonetheless, the validity of a memoryless Poisson model
is assumed.
Mean lifetimes are Weibull distributed (see Appendix A), i.e. denoting the mean
lifetime by t = 1/, the pdf of t is
 
t 1 (t/)
w(t) = e , (6.18)

where is a scale parameter and is a shape parameter.
Change rates and timespans are independent and thus
 
P (c | ) = P (c, | ) d = P ( )P (c | , ) d.
0 0

The resulting lifetime distribution is


f (t) = et w(1/) d(1/), (6.19)
0

where w(1/) is an estimate of the mean lifetime. Brewington and Cybenko (2000)
used the data shown in Figure 6.7 to estimate the Weibull distribution parameters,
obtaining = 1.4 and = 152.2. The estimated mean lifetime distribution is plotted
in Figure 6.9.
166 WEB DYNAMICS
Estimated Estimated
1
probability density function cumulative distribution function
of mean page lifetime of page lifetime
0.004
0.8

0.003 0.6

0.002 0.4

0.001 0.2

0 t
200 400 600 800 1 5 10 50 100 500 1000
Lifetime in days Lifetime in days

Figure 6.9 Estimated density and distribution of mean lifetime resulting from
the study of Brewington and Cybenko (2000).

fetch fetch
Grace period

0 t- t I
Figure 6.10 A document is -current at time t if no changes have occurred before
the grace period that extends backward in time until t .

These results allow us to estimate how often a crawler should refresh the index
of a search engine to guarantee that it will remain (, )-current. Let us consider
rst a single document and, for simplicity, let t = 0 be the time when the crawler
last fetched the document. Also, let I be the time interval between two consecutive
visits (see Figure 6.10). The probability that for a particular time t the document
is unmodied in [0, t ] is e(t) for t [, I ) and 1 for t (0, ). Thus, the
probability that a specic document is -current at time t is
  I
1 1 (t) 1 e(I )
dt + e dt = + , (6.20)
0 I I I I
but since each document has a change rate , whose reciprocal is Weibull distributed,
the probability that the collection of documents is -current is
  
1 e(I /t)
= w(t) + dt. (6.21)
0 I t/I
This allows us to determine the minimum refresh interval I to guarantee (, )-
currency once the parameters of the Weibull distribution for the mean change rate are
known. Assuming a Web size of 800 million pages, Brewington and Cybenko (2000)
ADVANCED CRAWLING TECHNIQUES 167

determined that a reindexing period of about 18 days was required to guarantee that
95% of the repository was current up to one week ago.

6.4.2 Other measures of recency

Freshness and index age are two alternative and somewhat simpler measures of
recency (Cho and Garcia-Molina 2000b). The freshness (t) at time t of a given
document is a binary function that indicates whether the document is up-to-date in
the index at time t. The expected freshness is therefore the probability that the doc-
ument did not change in (0, t], i.e. E[(t)] = et (see Equation (6.17)). Note that
freshness essentially corresponds to the concept of -currency for = 0. Hence, if
d is refreshed regularly each I time units, the average freshness is
1 eI
=
I
as follows from Equation (6.20) with = 0.
The index age of a document is the age of the document if the local copy is outdated,
or zero if the local copy is fresh. Thus if the document was modied at time s (0, t],
its index age is t s. From Equation (6.14) it follows that the expected age at time t
is  t
1 et
E[a(t)] = (t s)es ds = t ,
0
whose average in (0, I ] is
1 eI 1 I
a=2
+ .
I 2
When a collection of documents is considered, the above quantities can be averaged
over the collection. Because of the linearity of integral, we can average both time
punctual values and time averages.

6.4.3 Recency and synchronization policies

The bandwidth requirements that can be inferred from the analysis reported in Sec-
tion 6.4.1 could be somewhat pessimistic. Clearly, not all the sites change their pages
at the same rate. Cho and Garcia-Molina (2000b) conducted another study in 1999,
monitoring changes of about 720 000 popular 2 Web pages. Results are reported in
Figure 6.11. In terms of the overall popular Web, these diagrams are in good qual-
itative accordance with the ndings of Brewington and Cybenko (2000), indicating
that a vast fraction of this Web is dynamic. However, it is interesting to note that a
very simple Web partitioning, based on four top-level domains, yields dramatically
different results (Cho and Garcia-Molina 2000b). In particular, dot com websites
2 According to PageRank (see Section 5.4), so this forms a biased representative of the Web population.
168 WEB DYNAMICS
> 4 months
Average
14 months
.gov
1 week 1 month
.edu 1 day 1 week
1 day
.net

.com
0.0 0.2 0.4 0.6 0.8 1.0

Figure 6.11 Average change interval found in a study conducted by Cho and Garcia-Molina
(2000a). 270 popular sites were monitored for changes from 17 February to 24 June 1999. A
sample of 3000 pages was collected from each site by visiting in breadth rst order from the
homepage.

are much more dynamic compared to educational or governmental sites. While this
is not surprising, it suggests that a resource allocation policy that does not take into
account site (or even document) specic dynamics may waste bandwidth re-fetching
old information that, however, is recent in the index. From a different viewpoint, this
nonuniformity suggests that, for a given bandwidth, the recency of the index can be
improved if the refresh rate is differentiated for each document.
To understand how to design an optimal synchronization policy we will make
several simplifying assumptions. Suppose there are N documents of interest and
suppose we can estimate the change rate i , i = i, . . . , N , of each document. Suppose
also that it will be practical to program a crawler that regularly fetches each document
i with a refresh interval Ii . Suppose also that the time required to fetch each document
is constant. The fact that we have limited bandwidth should be reected in a constraint
involving Ii . If B is the available bandwidth, expressed as the number of documents
that can be fetched in a time unit, this constraint is


N
1

N. (6.22)
Ii
i=1

The problem of optimal resource allocation consists of selecting the refresh intervals
Ii so that a recency measure of the resulting index (e.g. freshness or index age) will be
maximized (Cho and Garcia-Molina 2000b). For example, we may want to maximize
freshness

N
(I1 , . . . , IN ) = arg max (i , i ) (6.23)
Ii ,...,IN
i=1

subject to (6.22). We might be tempted to arrange a policy that assigns to each doc-
ument a refresh interval Ii that is proportional to the change rate i . However, this
intuitive approach is suboptimal, and can be proven to be even worse than assign-
ing the same interval to each document. The optimal intervals can be easily derived
ADVANCED CRAWLING TECHNIQUES 169

as the solution of a constrained optimization problem such as the one described by

Equation (6.23).
If very large collections of documents need to be monitored, bandwidth limitations
may even prevent us from frequently monitoring document changes. Cho and Ntoulas
(2002) have recently proposed a sampling approach where a small fraction of pages
from a given site is checked for changes and the results are used to estimate the change
rate of the entire site.
WebFountain (Edwards et al. 2001) is a fully distributed and incremental crawler
with no central control or centralized queue of URLs. Incremental in this case means
that the repository entry of a given document is updated as soon as it is fetched from
the Web and the crawling process is never regarded as complete. The goal is to keep
the repository as fresh and as complete as possible. The model in this case is not
based on assumptions on the distribution of document change rates. Changes are
simply detected when a document is re-fetched and documents are grouped into a set
of buckets, each containing documents having similar rates of change. The trade-off
between re-fetching (to improve freshness) and exploring (to improve coverage) is
controlled in this case by maintaining separate queues for old and new URLs.
The optimal ratio of between the number of old and new URLs is determined as the
solution of a constrained optimization problem (see Edwards et al. (2001) for details).
Wolf et al. (2002) also formulate crawling as an optimization problem and suggest
a pragmatic metric, the embarrassment level, that focuses on the expected number of
times a search engine client is returned a stale URL.

Exercises
Exercise 6.1. Write a simplied crawling program that organizes the list of URLs
to be fetched as a priority queue. Order the priority queue according to the expected
indegree of the page pointed to by each URL and compare your results to a best-rst
search algorithm, using the actual indegree as a target measure of relevance.
Exercise 6.2. Extend the crawler developed in Exercise 6.1 to search for documents
related to a specic topic of interest. Collect a set of documents of interest to be used as
training examples and use one of text categorization tools studied in Chapter 4 to guide
the crawler. Compare results obtained using the parent-based and the anchor-based
strategies.
Exercise 6.3. Write a program that recursively scans your hard disk and estimates
the lifetime of your les.
Exercise 6.4. Suppose the page mean lifetime is Weibull distributed with parameters
and . What are the average mean lifetime, the most likely mean lifetime, and the
median mean lifetime? What would be reasonable values for these quantities starting
from an estimated distribution like the one shown in Figure 6.9?
Exercise 6.5. Explain what we mean when we say that a collection of stored docu-
ments is (, )-current.
170 WEB DYNAMICS

Exercise 6.6. Suppose you want to build a (0.8,1-week)-current search engine for
a collection of 2 billion documents whose average size is 10 Kb. Suppose the mean
change is Weibull distributed with = 1 and = 100 days. What is the required
bandwidth (suppose the time necessary to build the index is negligible)?
7

Modeling and Understanding

Human Behavior on the Web

7.1 Introduction
Up to this point in the book we have focused on how the Web works and its general
characteristics, such as the inner workings of search engines and the basic properties
of the connectivity of the Web. This provides us with a rich source of information
on how the Web is constructed and interconnected, illustrating its inner workings
from an engineering viewpoint. However, there is a critically important aspect of the
picture that we have not yet discussed, namely how humans interact with the Web. In
this chapter we will examine in some detail this aspect of human interaction across
a broad spectrum of activities, such as how we use the Web to browse, navigate, and
issue search queries.
The Web can be viewed as an enormous distributed laboratory for studying human
behavior in a virtual digital environment. From a data analysis viewpoint, the Web
provides rich opportunities to gather observational data on a large-scale and to use
such data to construct, test, refute, and adapt models of how humans behave in the
Web environment. For example, if we were to record all search queries issued over a
12-month period at a large search-engine website, we could then use this data for
(1) exploration, e.g. we can generate summary statistics on how site-visitors are
issuing queries, such as how many queries per session, or the distribution of
the number of terms across queries;
(2) modeling, e.g. we can investigate whether there is any dependence between the
content of a query and the time of day or day of week when the query is issued;
(3) prediction, e.g. we can construct a model to predict which Web pages are likely
to be most relevant for a query issued by a site-visitor for whom we have past
historical query and navigation data.
In this chapter we will explore how we can go from exploratory summary statistics
(such as raw counts, percentages, and histograms) to more sophisticated predictive
Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
172 WEB DATA AND MEASUREMENT ISSUES

models. These types of models are of broad interest across a wide variety of elds:
from the social scientist who wishes to better understand the social implications of
Web usage, to the human factors specialist who wishes to design better software
tools for information access on the Web, to the network engineer who seeks a better
understanding of the human mechanisms that contribute to aggregate network trafc
patterns, to the e-commerce sales manager who wants to better predict which factors
inuence the Web shopping behavior of a site-visitor.
We begin with an overview of Web data measurement issues. While the Web allows
us to collect vast amounts of data, there are nonetheless some important factors that
limit the quality of this data, depending, for instance, on how and where the data are
collected. In Section 7.2 we discuss in particular those factors that can have a direct
impact on any inferences we might make from the available data. We follow this in
Section 7.3 by looking at a variety of empirical studies that investigate how we navigate
and browse Web pages. Section 7.4 then discusses a number of different probabilistic
mechanisms (such as Markov models) that have been proposed for modeling Web
browsing. Section 7.5 concludes the chapter by looking at how we issue queries to
search engines, discussing data from various empirical studies as well as a number of
probabilistic models that capture various aspects of how we use Web search engines.

7.2 Web Data and Measurement Issues

7.2.1 Background
As with any data analysis endeavor, before we even look at our data it is critically
important to understand how the data are collected. Web data are typically collected
in a completely automated fashion via software logging tools. The basic features of
Web server logging were discussed in Chapter 2. Automated logging is quite useful
in the sense that the data are acquired continuously without any manual supervision
being required on our part to collect the data (in comparison, for example, with the
effort required to collect census data, or to collect patient data in a medical study).
However, the downside of automated logging is that we can never be completely
certain that the collected data faithfully represents what we believe it to be. For
example, the server log les for page requests at a website will contain page requests to
the Web server from both human users and automated robots (e.g. crawler programs
as discussed in Chapter 2). In any analysis of the data it will be important to separate
these two different sources of page requests (if possible), since otherwise the robot data
may signicantly skew the statistics for human users. For example, robots typically
explore the entire website in a very systematic manner and, thus, tend to issue far
more page requests (and often at a much quicker rate) than a typical human user. Tan
and Kumar (2002) discuss general techniques for automatically detecting sessions
generated by robots.
For example, Figure 7.1 shows three plots of the number of page requests per
hour logged at the www.ics.uci.edu website during the rst week of April 2002.
HUMAN BEHAVIOR AND THE WEB 173

10000
Saturday Sunday Monday Tuesday Wednesday Thursday Friday

5000

0
0 20 40 60 80 100 120 140 160

10000
HUMAN-GENERATED TRAFFIC

5000

0
0 20 40 60 80 100 120 140 160

10000 ROBOT-GENERATED TRAFFIC

5000

0
0 20 40 60 80 100 120 140 160
Time (in hours)
Figure 7.1 Number of page requests per hour as a function of time from page requests
recorded in the www.ics.uci.edu Web server logs during the rst week of April 2002. The
solid vertical lines indicate boundaries between the different days (at midnight). The top plot
is total trafc, the center plot is human-generated trafc, and the lower plot is robot-generated
trafc.

Robot page requests were identied by classifying page requests using a variety of
heuristics for example, many robots self-identify themselves in the server logs. The
top plot is the total number of page requests per hour, the center plot is the estimated
number generated by humans per hour, and the lower plot is the estimated number of
page requests generated by robots per hour. The top plot is the sum of the center and
lower plots.
We can clearly see that the human-generated trafc is quite different from the
robot-generated trafc. The robot trafc consists of two components: periodic spikes
in trafc combined with a lower-level relatively constant stream of requests. The more
constant component of the robot trafc can be assumed to be from robots from good
citizen search engines and cache sites, which attempt to distribute over time their
crawler requests to any single site so as not to cause any signicant bursts of activity
(and potential deterioration in quality of service) at the Web server. The spiky trafc
on the other hand can be assumed to be coming from a less well-intentioned source
(or sources), whose crawler periodically oods websites with page requests and can
174 WEB DATA AND MEASUREMENT ISSUES

potentially overload a server. In Figure 7.1 the dominant burst of page requests is
arriving at the same time each night, just after midnight.
The human-generated trafc has quite a different characteristic. There is a noisy,
but quite discernible, daily pattern of trafc, particularly during Monday to Friday.
Human-generated trafc tends to peak around midday, and tends to be at its lowest
between 2 am and 6 am (not surprisingly). This suggests that much of the trafc to the
site is being generated in synchrony with the daily Monday to Friday work schedule
of university staff and students.
The main point here is to illustrate that Web server log data need careful interpreta-
tion. The presence of robot data in this example is a signicant contributor to overall
trafc patterns on the website. For example, if we were to mistakenly assume that the
data in the top plot in Figure 7.1 were all human-generated, this assumption might
lead us to rather different (and false) hypotheses about time-dependent patterns in
human trafc to the site.

7.2.2 Server-side data

Data logging at Web servers
As discussed in Chapter 2 a primary function of a Web server (such as that running
at www.ics.uci.edu) is to accept requests for particular pages at that website
from individual clients (Web browser software running on machines anywhere on the
Web) and to transmit the content of that page to the requester if it is a legal page
request (e.g. after it checks if the page actually exists). In addition to serving Web
pages, Web server software is typically congured to archive all page requests in a log
le, recording the URL of the le requested, the time and date of the request, the IP
address from which the request originated, information about the browser requesting
the le, and an error code if the requested le could not be returned (e.g. if there is
no such le). The URL of the referrer page may also be recorded if applicable; if a
page request for page B is generated by a user clicking on a link on page A, then A is
the referrer page for page B.
Server-side log les can potentially provide a wealth of information about how
a website is being used, but require considerable care in interpretation. A detailed
discussion of these issues can be found in Cooley et al. (1999), Mena (1999), and
Shahabi et al. (2001).

Page requests, caching, and proxy servers

As described in Chapter 2, when a user requests a particular Web page using their
browser, a message is sent via the HTTP protocol to the Web server, typically via the
GET method. In theory, each such request is received by the Web server and both
the request and the action of the server are logged in the server access log le with
a timestamp. The server also records the IP address of the computer from which the
users request originated. The server can return the page if the URL is recognized, it
can return an error message if the requested URL cannot be found on the website, or
HUMAN BEHAVIOR AND THE WEB 175
User Other Users

Local Computer
Page Caching, Multiple Users

Local Network
Dynamic Addressing

Proxy Server
Page Caching

Web Server

Figure 7.2 A graphical summary of how page requests from an individual user can be masked
at various stages between the users local computer and the Web server, via page caching at
the browser level, multiple different users on the local computer, dynamic network address
assignment at the local network level, and more page caching by proxy servers.

it can simply reply with a message that it is too busy to process the request at that
time.
Thus, in theory, the Web server sees every individual request for pages on the site. In
practice the situation is a little more complex, as illustrated graphically in Figure 7.2.
Browser software, for example, often stores in local memory the HTML content of
pages that were requested earlier by a user. This is known as caching and is a generally
useful technique for improving response time for pages that are requested more than
once over a period of time. For example, if the content of a particular Web page does
not change frequently, and a user uses the back button on their browser to revisit the
page during a session, then it makes sense for the browser to redisplay a stored or
cached copy of the page rather than requesting it again from the Web server. Thus,
although the user viewed the page a second time, this page viewing is not recorded
in the Web server log because of local caching.
Caching can also occur at intermediate points between the users local computer
and the Web server. For example, an Internet service provider (ISP) might cache Web
pages that are particularly popular and store them locally. In this manner, requests
from users for these pages will be intercepted by the ISP cache and the cached version
of the page is displayed on the users browser. Once again the page request by the
user for the cached page will not be recorded in the Web server logs.
Proxy servers can be thought of as a generalization of caching. A proxy server is
typically designed to serve the Web-surng needs of a large set of users, such as the
customers of an ISP or the employees of a large organization. The proxy exists on
the network at a point between each individuals computer and the Web at large. It
acts as a type of intermediary for providing Web content to users. For example, the
176 WEB DATA AND MEASUREMENT ISSUES

contents of entire websites that are frequently accessed might be downloaded by the
proxy server during low-trafc hours, in order to reduce network trafc at peak hours.
Various security mechanisms and ltering of content can also be implemented on the
proxy server.
From a data analysis viewpoint the overall effect of caching and proxy servers is
to render Web server logs somewhat less than ideal in terms of being a complete and
faithful representation of individual page views.
In practice, in analyzing Web data, one can use various heuristics to try to infer the
true actions of the user given the relatively noisy and partial information available
in the Web server logs at the server. For example, one can try to combine pairs of
referrer-requested pages with knowledge of the link structure of the website, to try
to infer the actual path taken by an individual through a website. For example, if a
user were to request the sequence page A, page B, page C, page B, page F, the Web
server log would likely only record the sequence ABCF, since page B would likely be
cached and not requested from the server the second time it was viewed. The logged
referrer page for F would be page B, and using the timing information for the requested
pages (say, for example that page F was requested relatively quickly after page C)
one could then infer that the actual sequence of requests was ABCBF. In simple cases
these heuristics can be quite useful (see, for example, Anderson et al. (2001) for a
denition of Web trails in this context). However, in the general case, one can rarely be
certain about what the user actually viewed. Statistical methods are likely to be quite
useful for handling such uncertainty. At the present time, however, relatively simple
heuristics, such as only using actual logged page requests in subsequent data analysis,
and not making any assumptions about missed requests, are common practice in Web
usage analysis.

Identifying individual users from Web server logs

In analyzing server log data over a specic period of time we would like to be able to
associate specic page requests with specic individual users. We can certainly group
requests from the same IP address. However, IP addresses do not necessarily have a
one-to-one mapping to individual users (see for example Figure 7.2). One IP address
might correspond to multiple different users, such as different individuals logged in
to a multi-user system each running different browsers simultaneously, or to a single-
user machine used by different people over time (e.g. in an academic laboratory).
Even if there is only one user using a particular machine, the IP address may change
over time, since it might be dynamically assigned by the individuals local network
and, thus, we cannot link different page requests from this individual over time.
One approach to get around some of these problems is to use cookie information.
A cookie is small text le that is temporarily created on a users hard disk by a Web
server. This text le allows the Web server to uniquely identify and track requests
from that individual user on that machine over time. When a user returns to a site
they visited previously (e.g. a day earlier) the information in the cookie le can be
accessed by the Web server and used to identify the user as the same user who visited
HUMAN BEHAVIOR AND THE WEB 177

on the previous day. Consequently, actions by the same user, such as page requests and
product purchases, during different sessions can be linked together. Even if a users IP
address is dynamically re-assigned between sessions, the cookie information remains
the same.
Thus, we can replace the IP address in the server log le with a more reliable
identier as provided by the cookie le. Commercial websites use cookies extensively
and they can be quite effective in improving the quality of matching page requests to
individual users, or at least matching to specic login accounts on specic machines.
Although browser software allows individual users to disable the storing of cookies
on their machines, many individual users either choose not to do so voluntarily ( or
are unaware of the fact that cookies can be disabled). Many commercial websites
require that cookies are enabled in order for a user to access their site. It is estimated
that in general well over 90% of Web surfers have the cookie feature permanently
enabled on their browsers, and cookies generally tend to be quite effective for this
reason. Of course, cookies in effect rely on the implicit cooperation of the user and,
thus, raise the issue of privacy. For this reason some non-commercial sites (such as
those at university campuses) often choose not to activate cookie creation on their
Web servers in the belief that cookies can be viewed as somewhat of an invasion of
an individuals right to privacy.
Another option for user identication is to require users to voluntarily identify
themselves via a login or registration process, e.g. for websites such as newspapers
that charge fees for accessing content. As with cookies, this can provide much higher
reliability in terms of being able to accurately associate particular requests with par-
ticular individuals, but again comes with the cost of imposing additional requirements
on the user, which in turn may discourage potential visitors to the site.

7.2.3 Client-side data

Collecting data at the client side, for example, by recording a users browser actions
with software on the users computer, solves many of the problems associated with
server-side data logging. Advantages include
direct recording of page requests, thus, eliminating masking due to caching;
recording of all browser-related actions by a user, including visits to multiple
websites;
more-reliable identication of individual users (e.g. by login ID for multiple
users on a single computer).
The richness of client-side data means that it is the preferred mode of data col-
lection for large-scale detailed studies of general navigation behavior on the Web.
For example, companies such as comScore and Nielsen use client-side software to
track home computer users browser usage on an ongoing basis. Users are selected
randomly using statistical experimental design methods to ensure a representative
178 WEB DATA AND MEASUREMENT ISSUES

Table 7.1 Estimates of monthly average user activity as reported from client-side data
at www.netratings.com for September 2002.

N. America Japan Ireland France

Sessions per month 22 21 8 18

Unique sites visited 47 59 24 51
Total time (h) 11.6 11.6 3.6 8.5
Time per session (min) 31.6 33.3 26.7 28.9
Page-view duration (s) 55 35 44 42

cross-section. For example, as of October 2002, comScores MediaMetrix software

was actively monitoring browser actions of 60 000 Web users in North America.
These data sets can provide relatively accurate estimates of which search engines and
e-commerce sites are most popular, as well as providing the basis for studies that try
to estimate the effect of user demographics on user behavior (we will see an example
of such a study later, in Section 7.4.8).
Table 7.1 shows the estimated average monthly activity of a typical Web user
in different countries. This type of client-side data is typically recorded on home
computers and, thus, does not reect a complete picture of Web browsing behavior,
since it does not necessarily include the substantial volume of surng data generated
at business locations.
Statistics such as Time per session and Page-view duration in Table 7.1 can
provide valuable information for user modeling in general. Server-side estimates of
these same numbers would be very unreliable, if not impossible to obtain. Page-view
duration for example, if estimated at the server-side, would be confounded by the
fact that a particular Web server log only measures requests to that site. Thus, in
between any two page requests to a particular site it is possible that the same user
requested multiple other pages from other sites. Of course, even at the client-side, a
statistic such as page-view duration does not directly measure how long a user views
the page. Rather it merely measures the time duration between one page request and
the next, and instead of viewing the page during that time the user might in fact be
fetching a cup of coffee. One would in principle need to use video cameras and eye-
tracking software to accurately monitor page viewing patterns at the client-side, but
the expense and inconvenience of such setups have been a barrier to their use in any
large-scale studies to date. However, as sensor technology continues to decline in size
and cost it seems likely that user-monitoring systems (such as cameras that passively
monitor and track a users behavior) will become more prevalent.
Web surng data can also be collected at other intermediate points in the network
between an individuals computer and specic Web servers. For example, ISPs and
proxy servers can log all outgoing page requests and browser actions from their cus-
tomers. These data are often analyzed and used for a variety of commercial purposes.
For example, by analyzing the sites that an ISP customer typically visits, a prole of
that user can be constructed over time and used in real time for targeted advertising.
HUMAN BEHAVIOR AND THE WEB 179

Handling massive Web server logs

Web server logs can be very large in size. Even a relatively small university depart-
ment website can receive millions of page requests per month. Popular commer-
cial websites, such as www.amazon.com or www.google.com, can receive tens
of millions of page requests in a single day. If each page request requires roughly
100 bytes to store, then one months worth of page-request data could be on the
order of a terabyte or more. These data volumes pose signicant practical challenges
to conventional data analysis methods, since they clearly exceed the main memory
capacities of current computers. Instead the server data are typically be stored in sec-
ondary memory on disk. Random access to records on a disk is typically orders of
magnitude slower than random access to records in main memory (up to 105 times
slower!). Even pipelined sequential scanning of data on disk is limited to about two
Mb per second with current technology, so that to load one Gb of data from disk to
main memory will take on the order of 8 min. These time-costs for data access place
signicant constraints on what types of data analyses we can carry out. In practice
we must often resort to analyzing subsets of the data, e.g. selecting a single days
worth of data, perhaps ltering out events and elds that are not of direct interest, and
focusing on the remainder of the data. In what follows we will ignore these issues of
data access and in effect implicitly assume that the server log data of interest can be
accessed in a time-efcient manner (e.g. assuming that we are in effect focusing on
a subset of the log les and the data of interest reside in main memory). Techniques
and algorithms for analyzing large data sets that reside outside of main memory are
an interesting data mining topic in themselves, but somewhat outside the scope of
this text. For further reading on such techniques and algorithms see Han and Kamber
(2001).

7.3 Empirical Client-Side Studies of Browsing

Behavior
A variety of empirical studies have examined in detail how groups of selected users use
their Web browsers to navigate the Web. Data for these studies are typically collected
at the client-side over a period of time so that, for example, reliable information on
Web page revisitation patterns can be gathered.
Because the data sets are collected at the client-side, explicit permission needs to
be requested from each user whose data are used in the study. For research studies
conducted at universities (rather than large-scale commercial efforts) this often limits
the number of individuals in each study to a relatively small number. Furthermore,
there can be an implicit bias in the nature of the population being studied individuals
in an academic client-side study are typically all from the same university department
and consist of computer science faculty, staff, and graduate students. Thus, some
caution must be exercised in generalizing to the Web-using population at large from
data that are gathered on small samples on individuals.
180 EMPIRICAL CLIENT-SIDE STUDIES OF BROWSING BEHAVIOR

Client-side studies nonetheless provide relatively important clues about human

behavior in the context of the Web. These clues can, for example, help design better
network algorithms such as improved Web-caching and prediction strategies as well
as support the development of better human interfaces for Web navigation such as
more effective tools to allow a user to nd pages previously visited.

7.3.1 Early studies from 1995 to 1997

Two of the earliest and most widely cited studies on client-side Web-browsing behav-
ior are Catledge and Pitkow (1995) and Tauscher and Greenberg (1997). In both
studies, data were collected by logging Web browser commands (page requests, use
of the back and forward buttons, bookmarking, etc.) for a specic set of users over
a period of time. In both studies the population consisted of faculty, staff, and students
in a computer science department. Catledge and Pitkow collected data on 107 users
over three weeks in 1994 and recorded 31 134 navigation commands, corresponding
to roughly 14 page requests per user per day. Tauscher and Greenberg collected data
on 23 subjects over a six-week period in 1995, with a total of about 19 000 navigation
commands, and about 21 page requests per user per day.
Both studies found that clicking on hypertext anchors to navigate to another page
was the most common Web-browser action, accounting for about 50% of all actions
in each study. Using the back button was the second most common action, 41% of
all actions in the 1994 study and about 30% of actions in the 1995 study. The other
categories of actions (such as bookmarking, navigating using a history list, etc.) were
roughly evenly distributed among the remaining 1020% of possible actions.
The Tauscher and Greenberg study focused in particular on how often pages are
revisited by users. They found that the probability that a page is revisited (i.e. that the
URL requested by a user is a URL that the user has requested in the past during the
study) was about 0.58 across all users in the study. They also re-analyzed data from
55 users in the Catledge and Pitkow study and estimated the revisitation probability
to be 0.61 from those data. Note that these estimates are likely to be underestimates of
the actual revisitation rate, since any pages requested prior to the start of each study
are not accounted for (and thus are not be counted as revisitedwhen requested during
the study). Nonetheless, a lower bound of 61% indicates that the overall revisitation
rate is quite high, suggesting that we are indeed creatures of habit. In practice, many
of these revisits could be to pages whose contents change regularly (such as news
pages) so that although the URL is the same, the content is different.
Tauscher and Greenberg also found a strong recency effect, namely that a page that
is revisited is often a page that has been visited very recently rather than in the distant
past. This is correlated with the empirical observation that users tend to use the back
button when navigating far more frequently than they use bookmarks or history lists.
It is interesting to note that repetition of user actions is not unique to Web-browsing
and has been found to be a strong characteristic of user behavior in several other
domains. Greenberg (1993) describes empirical data that estimated how many tele-
HUMAN BEHAVIOR AND THE WEB 181

phone numbers dialed are repeated (57%) and how many Unix command lines are
repetitions of earlier commands (75%).

7.3.2 The Cockburn and McKenzie study from 2002

Although the 1994 and 1995 studies described above are widely cited in the research
literature on empirical analysis of Web use, they are relatively out of date given the
rate at which both the Web and our use of the Web has changed in the past few years.
The more recent study of Cockburn and McKenzie (2002) provides a more up-to-date
snapshot of browsing behavior. In their study, they analyzed the daily history.dat les
produced by the Netscape browser for 17 users between October 1999 and January
2000. The population being studied consisted of faculty, staff, and graduate students
at the University of Canterbury in New Zealand.
Over the 119 days of logging, 84 841 separate page requests were issued by the
17 users, which yields an approximate rate of 42 page requests per day per user. This is
substantially higher than the roughly 14 and 21 page requests per day in the 1994 and
1995 studies, and is likely a conrmation of the fact that Web-browsers have become
an increasingly integral part of our daily work patterns (again with the caveat that all
of these studies are based on relatively small sample sizes in terms of the number of
individuals involved).
Cockburn and McKenzie estimated that the revisitation rate was 0.81 in their study,
averaged across users, ranging from a low of 0.61 for one user to a high of 0.92 for
another user. These rates are signicantly higher than the rates of about 0.6 obtained in
the earlier studies. Two factors seem plausible as explanations for this increase. First,
as mentioned earlier, the estimation of revisitation rates from a nite time-window of
data is likely to be an underestimate of the true revisitation rate. Since the Cockburn
and McKenzie time-window of four months was over three times longer than either
of the other two studies, the longer time-window may have less of a downward bias.
The second and potentially more interesting factor is that our usage of the Web may
have evolved from a more exploratory mode in 19941995, to a more utilitarian mode
by 1999, a mode where we revisit certain sites such as general news, stock market
information, and digital libraries on a regular basis.
Figure 7.3a shows a scatter plot of the number of distinct pages requested per
user versus the total number of pages requested (over the 119 days). As we might
expect there is a strong linear dependence: the more you surf the Web, the more
new pages you tend to visit. It is also informative to look at the same data on a
loglog plot (Figure 7.3b), where we now see that the marginal distributions of unique
pages and total pages are somewhat less skewed and more symmetric on a log-scale.
This suggests that if we wished to model the probability distribution of the average
number of total requests per user (or the rate of requests per unit time) over a whole
population, a log-scale distribution such as a log-normal might be a useful approach,
since the data appear to have more symmetry and regularity on a log-scale (a log-
normal distribution for a real-valued variable X is dened by assuming that log X has
a normal distribution).
182 EMPIRICAL CLIENT-SIDE STUDIES OF BROWSING BEHAVIOR

4500

4000
Number of unique pages visited per user

3500

3000

2500

2000

1500

1000

500

0
0 0.5 1 1.5 2 2.5
Total number of pages visited per user 4
x 10
4
10
Number of unique pages visited per user

3
10

2
10

1
10
2 3 4 5
10 10 10 10
Total number of pages visited per user

Figure 7.3 The number of distinct pages visited versus page vocabulary size for each of
17 users in the Cockburn and McKenzie (2002) study, on both (a) standard, and (b) loglog
scales.
HUMAN BEHAVIOR AND THE WEB 183
0.45

0.4

Fraction of most frequent page visits to total page visits

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18
User IDs

Figure 7.4 Bar chart of the ratio of the number of page requests for the most frequent page
divided by the total number of page requests, for 17 users (sorted by ratio) in the Cockburn
and McKenzie (2002) study.

Cockburn and McKenzie also found several other interesting and rather pronounced
characteristics of client-side browsing. For example, the most frequently requested
page for each user (e.g. their home page) can account for a relatively large fraction of
all page requests for a given user. Figure 7.4 shows a bar chart of this fraction for each
of the 17 users, ranging from about 50% of all requests for one user down to about
5%. This data also indicates that, while there is often considerable regularity in Web-
browsing behavior for a particular individual, there can be considerable heterogeneity
in behavior across different individuals.

Video-based analysis of Web usage

In addition to the studies mentioned above, other types of research studies have been
used to improve our understanding of client-side Web behavior. For example, Byrne et
al. (1999) analyzed video-taped recordings of eight different users as they used their
Web browsers over a period of 15 min to about 1 h. The resulting audio descriptions by
the users of what they were doing and the video recordings of their screens were then
analyzed and categorized into a taxonomy of tasks. Among the empirical observations
were the fact that users spent a considerable amount of time scrolling up and down
Web pages (40 min out of 5 h). Another observation was that a considerable amount
of time was spent by users waiting for pages to load, approximately 50 min in total,
or 15% of the total time. This gure is of course highly dependent on bandwidth and
other factors related to the network infrastructure between the user and the website
being visited.
184 PROBABILISTIC MODELS OF BROWSING BEHAVIOR

7.4 Probabilistic Models of Browsing Behavior

Moving beyond empirical studies, a natural next step is to build models that describe
the browsing behavior of users on the Web. Modeling can both generate insight into
how we use the Web as well as provide mechanisms for making predictions for a
variety of applications such as pre-fetching of Web pages and personalization of Web
content. In the next few sections we will examine some specic examples of various
types of stochastic models that have been proposed for Web browsing data.

7.4.1 Markov models for page prediction

Consider building a model to describe how users navigate around a specic website.
One general approach is to model their behavior by a nite-state Markov chain, where
the states in the chain correspond to specic Web pages or general categories of Web
pages. If we are only interested in the order in which users visit Web pages, and
not in the time between page requests, then we can model each new page request
as a transition between states. If each state corresponds to a Web page, it might
be reasonable to assume that there are no self-transition probabilities, i.e. that all
transitions go from one page to a different page. However, if the states correspond to
general categories of pages and contain multiple pages, then we can allow nonzero
self-transition probabilities in the model to allow navigation between pages that are
assigned to the same state in the model.
If we wish to explicitly include time in the model then we could (for example)
use a discrete-time Markov chain (where the unit of time could be one second, for
example) where we record at each tick of the clock what state the user is in. Another
option would be to use a continuous-time Markov process.
For simplicity we will focus on the order-dependent time-independent model where
each transition corresponds to a page request. We will also assume that we can asso-
ciate each page request with a specic individual and, thus, decompose a Web server
log into individual user sessions.
A nite-state Markov chain is characterized by a set of M states (see Appendix A).
Let s be a sequence of observed states of length L, e.g. such a sequence might
be ABBCAABBCCBBAA with three states A, B, and C. st denotes the state at
position t in the sequence, 1
t
L. In general, for a sequence of xed length L,
the probability of generating a specic sequence s can be written as

L
P (s) = P (s1 ) P (st | st1 , . . . , s1 ).
t=2

Under a rst-order Markov assumption, we assume that the probability of each state
st given the preceding states st1 , . . . , s1 only depends on state st1 , i.e.

L
P (s) = P (s1 ) P (st | st1 ). (7.1)
t=2
HUMAN BEHAVIOR AND THE WEB 185

This provides a simple generative model for producing sequential data, in effect a
stochastic nite-state machine: we select an initial state s1 from a distribution P (s1 )
then, given s1 , we choose state s2 from the distribution P (s2 | s1 ), and so on. Denoting
Tij = P(st = j | st1 = i) as the probability of transitioning from state i to state j ,
where M j =1 P (st = j | st1 = i) = 1, we dene T as an M M transition matrix
with entries Tij , assuming a stationary Markov chain (see Appendix A).
This type of rst-order Markov model makes a rather strong assumption about
the nature of the data-generating process, namely that the next state to be visited is
only a function of the current state and is independent of any previous states visited.
The virtue of such a model is that it provides a relatively simple way to capture
sequential dependence. Thus, although the true data-generating process for a particular
problem may not necessarily be rst-order Markov, such models can provide useful
and parsimonious frameworks for analyzing and learning about sequential processes
and have been applied to a very wide variety of problems ranging from modeling
of English text to reliability analysis of complex systems (see, for example, Ross
2002).
Returning to the issue of modeling how a user navigates a website, consider a
website with W Web pages, where we wish to describe the navigation patterns of
users at this website. Since W may often be quite large (e.g. on the order of 105 or
106 for a large computer science department at a university), it may not be practi-
cal to represent each Web page with its own state, since a W -state model requires
O(W 2 ) transition probability parameters to specify the Markov transition matrix T .
To alleviate this problem we can cluster or group the original W Web pages into a
much smaller number M of clusters, each of which is assigned a state in the Markov
model. The clustering into M states could be carried out in several different ways,
e.g. by manual categorization into different categories based on content, by grouping
pages based on directory structure on the Web server, or by automatically clustering
Web pages using any of the clustering techniques described earlier in Chapters 1
and 4.
Transition probabilities P (st = j | st1 = i) in such a model represent the
probability that an individual users next page request will be from category j , given
that their most recent page request was in category i. We can add a special end-state to
our model (call it E), which indicates the end of a session. For example, a commonly
used heuristic is to declare the end of a session once a certain time duration, such as
20 min, has elapsed since the last page request from that user.
As an example, if we have three categories of pages plus an end state E the transition
matrix for such a model can be written as

P (1 | 1) P (2 | 1) P (3 | 1) P (E | 1)
P (1 | 2) P (2 | 2) P (3 | 2) P (E | 2)
T = P (1 | 3)
. (7.2)
P (2 | 3) P (3 | 3) P (E | 3)
P (1 | E) P (2 | E) P (3 | E) 0
This model can simulate a set of nite length sequences, where the boundaries between
sequences are denoted by the invisible symbol E. The state E has a self-transition
186 PROBABILISTIC MODELS OF BROWSING BEHAVIOR

probability of zero, to denote that after we get a single occurrence of E the system
restarts and begins generating a new sequence, with probabilities P (1 | E), P (2 | E),
and P (3 | E) of starting in each of the three states. We could if we wish also dene
a xed starting state S of a sequence, which is then followed by one of the other
states however, in effect we only need the end-state E here, since E is sufcient to
denote the boundaries between the end of one sequence and the start of another.
Markov models have their limitations. The assumption of the rst-order Markov
model that we can predict what comes next given only the current state does not take
into account long-term memory aspects of Web browsing. For example, users can
use the back button to return to the page from which they just came, imparting at
least second-order memory to the process. For example, see Fagin et al. (2000) for
an interesting proposal for a stochastic model that uses back buttons.
We can try to capture more memory in the process by using a kth-order Markov
chain, dened in general by the assumption that

P (st | st1 , . . . , s1 ) = P (st | st1 , . . . , stk ), (7.3)

where k is some positive integer. Further model generalizations can be obtained by

using various forms of stochastic grammars. However, tting these more complex
models to data can often require inordinate (and impractical) amounts of training
data, and thus, we focus our discussion on the lowly rst-order Markov model and
later move on to various extensions of this model.

7.4.2 Fitting Markov models to observed page-request data

We can t Markov models to observed data by collecting server-side page-request

data from Web server logs. As discussed earlier in this chapter, server-side data have
a number of limitations in terms of being an accurate representation of the actual
page requests issued by a user. For example, caching by a users browser software
or by an intermediate proxy server can hide certain page requests from the servers
viewpoint. There are several ways to try to deal with such problems (see, for example,
Anderson et al. 2001; Cooley et al. 1999). We implicitly assume in what follows that
these issues have been sufciently resolved that we can proceed with tting models
to the session data that has been gathered.
Assume that we have collected such data in the form of N sessions, where the ith
session si , 1
i
N, consists of a sequence of Li page requests, categorized into
M 1 states and terminating in an invisible symbol E. We denote the entire data
set as D = {s1 , . . . , sN }.
Let denote the set of parameters for the Markov model that we wish to t to the
data. Using the representation in Equation (7.2), consists of the M 2 1 entries in the
transition matrix T , and we can dene ij as the estimated probability of transitioning
from state i to state j .
HUMAN BEHAVIOR AND THE WEB 187

The likelihood function L( ) is dened as the probability of the data conditioned

on a particular set of parameters for this model, i.e.


N
L() = P (D | ) = P (si | ). (7.4)
i=1

We assume on the right-hand side that page requests from different sessions are
conditionally independent of each other given the model (a reasonable assumption in
general, particularly if the sessions are being generated by different individuals). The
statistical principle of maximum likelihood can now be used to nd the parameters ML
that maximize the likelihood function L( ), i.e. the parameter values that maximize
the probability of the observed data D conditioned on the model (see Chapter 1 for a
general discussion of maximum likelihood).
To nd the ML parameters requires that we determine the maximum of L( ) as
a function of . Under our Markov model, it is straightforward to show that the
likelihood above can be rewritten in the form
 nij
L() = ij , 1
i, j
M, (7.5)
i,j

where nij is the number of times that we see a transition from state i to state j in the
observed data D. Note that the initial starting states for each sequence are counted as
transitions out of state E in our model.
We use the log-likelihood l() = log L() for convenience:

l( ) = nij log ij .
ij

As shown earlier in Chapter 1, we can directly maximize this expression by taking

partial derivatives with respect to each parameter ij and incorporating (via Lagrange
multipliers) the constraint
 that the sum of the transition probabilities out of any state
must sum to one, i.e. j ij = 1. This leads straightforwardly to ML solutions
nij
ijML =
ni
for the transition probabilities , where ni is the total number of page requests for
state i in the observed data. This is similar to the result obtained for the die model in
Chapter 1. The ML solutions turn out to be the intuitive frequency-estimates in this
case: the estimated transition probability is the number of times a transition from i to
j is observed divided by the number of observations of state i.

7.4.3 Bayesian parameter estimation for Markov models

In practice, if M is large (e.g. in the range of 102 or 103 ) we will end up estimating
a very large number of transition probabilities, namely M 2 . Even though our Web
188 PROBABILISTIC MODELS OF BROWSING BEHAVIOR

server logs may be quite large and, thus, D may contain potentially millions or even
hundreds of millions of sequences, there is still a good chance that some of the
observed transition counts nij may be zero in our data D, i.e. we will never have
observed such a transition. For example, if we have a new e-commerce site we might
wish to t a Markov model to the data for prediction purposes but in the rst few
weeks of operation of the site the amount of data available for estimating the model
might be quite limited.
As discussed in Chapter 1, a useful approach to get around this problem of lim-
ited data is to incorporate prior knowledge into the estimation process. Specically,
we can specify a prior probability distribution P ( ) on our set of parameters and
then maximize P ( | D), the posterior distribution on given the data, rather than
P (D | ).
The prior distribution P ( ) is supposed to reect our prior belief about it is
a subjective probability. The posterior P ( | D) reects our posterior belief in the
location of , but now informed by the data D. For the particular case of Markov
transition matrices, it is common to put a distribution on each row of the transition
matrix and to further assume that each of these priors are independent,

P ( ) = P ({i1 , . . . , iM }),
i

where j ij = 1. Consider the set of parameters {i1 , . . . , iM } for the ith row in the
transition matrix T . A useful general form for a prior distribution on these parameters
is the Dirichlet distribution, dened in Appendix A as


M
(qij 1)
P ({i1 , . . . , iM }) = Dqi = C ij , (7.6)
j =1

where , qij > 0, j qij = 1, and C is a normalizing constant to ensure that the
distribution integrates to unity (see Appendix A and Chapter 1 for details). The MP
posterior parameter estimates are
nij + qij
ijMP = . (7.7)
ni +
If there is a zero count nij = 0 for some particular transition (i, j ) then rather than
having a parameter estimate of 0 for this transition (which is what ML would yield),
the MP estimate will be qij /(ni + ), allowing prior knowledge to be incorporated.
If nij > 0, then we get a smooth combination of the data-driven information (nij )
and the prior information represented by and qij .
For estimating transitions among Web pages the following is a relatively easy way
to set the relative sizes of the priors.
(1) Let be the weight of the prior. One intuitive way to think of this is as
the effective sample size required before the counts for a particular state ni
balance the prior component in the denominator in Equation (7.7).
HUMAN BEHAVIOR AND THE WEB 189

(2) Given a value for , partition the states into two sets: set 1 contains all states
that are directly linked to state i via a hyperlink, and set 2 contains all states
that have no direct link from state i.
(3) Assign a total prior probability mass of , 0

1, on transitions out of any
state i into the states in set 2. We could further assume that transitions to all
states in set 2 are equally likely a priori, so that qij = /K for states j in set 2
(those not linked by a direct hyperlink to any page in state i). This assumes that
we have no specic information about which states are more likely in set 2.
The total probability  reects our prior belief that a site visitor will transition
to a Web page by a non-hyperlink mechanism such as via a bookmark, by use
of the history information in a browser, or by directly typing in the URL of the
page. Observational data (e.g. Tauscher and Greenberg 1997) suggest that  is
typically quite small, i.e. users tend to follow direct hyperlinks when surng.
(4) For the remaining transitions, those to states in set 1 (for which hyperlinks
exist from state i), we have a total of probability mass of 1  to assign. We
can either assign this uniformly, or we could weight it nonuniformly in various
ways, such as (for example) weighting by the number of hyperlinks that exist
between all of the Web pages in state i and those in set 1 (if the states represent
sets of Web pages).
(5) The prior probabilities both into and out of the end-state E can be set based on
our prior knowledge of how likely we think a user is to exit the site from any
particular state, and our prior belief over which states users are likely to use
when entering the site.
This assignment of a prior on transition probabilities is intended as a general guide of
how Bayesian ideas can be utilized in building models for Web trafc, rather than a
denitive prescription. For example, the model would need to be extended to account
for other frequently used navigational mechanisms such as the back button. Nor does
it account for dynamically generated Web pages (such as those at e-commerce sites),
where the hyperlinks on a page vary over time dynamically in a non-predictable
fashion.
The general idea of using a small probability  for nonlinked Web pages is quite
reminiscent of the methodology used to derive the PageRank algorithm earlier in
Chapter 5 (we might call this the Google prior). Sen and Hansen (2003) describe
the use of Dirichlet priors for the problem of estimating Markov model parameters,
with applications to characterizing and predicting Web navigation behavior.

7.4.4 Predicting page requests with Markov models

Markov models in a variety of avors have been applied to the problem of predicting
the next page, or a future page, that a user will request, given the sequence of pages that
he or she has requested up to that point (Sarukkai 2000; Zukerman et al. 1999). Being
190 PROBABILISTIC MODELS OF BROWSING BEHAVIOR

able to make such predictions provides a basis for numerous different applications,
such as pre-fetching and caching of Web pages and personalizing the content of a
Web page by dynamically adding hyperlink short-cuts to other pages.
As mentioned earlier, for a typical website the number of pages on the site can be
quite large. For example, for the UC Irvine website (Figure 7.1) it is estimated that
there are on the order of 50 000 different pages on the site. Thus, it is typical to group
pages into M categories and build predictive models for these groups with an M-state
Markov chain. The groups can be dened, for example, in terms of page functionality
(Li et al. 2002), page contents (Cadez et al. 2003), or the directory structure of the
site (Anderson et al. 2001).
Simple rst-order Markov models have generally been found to be inferior to
other types of Markov models in terms of predictive performance on test data sets
(Anderson et al. 2001; Cadez et al. 2003; Deshpande and Karypis 2001; Sen and
Hansen 2003). The most obvious generalization of a rst-order Markov model is the
kth-order Markov model. A kth-order model is dened so that the prediction of state
st depends on the previous k states in the model:
P (st | st1 , . . . , s1 ) = P (st | st1 , . . . , stk ), k {1, 2, 3, . . .}. (7.8)
This model requires the specication of O(M k+1 )
transition probabilities, one for
each of the possible combinations of M k histories multiplied by M possible current
states. An obvious problem with this model is that the number of parameters in the
model increases combinatorially as a function of both k and M. Thus, for example,
if we use simple frequency-based (maximum-likelihood) estimates of the transition
probabilities we face the problem of having many subsequences of length k + 1 that
will not have been seen in the historical data. For example, if M = 20 and k = 3,
then we need to have at least M k+1 = 204 = 160 000 symbols in the training data set.
Even this number is optimistically low some subsequences will be more frequent
than others and in addition we will need to see multiple instances of each subsequence
to get reliable estimates of the corresponding transition probabilities.
There are a number of ways to get around this problem. For example, Deshpande
and Karypis (2001) describe a number of different schemes for pruning the state
space of a set of kth-order Markov models such that histories with little predictive
power are pruned from the model. Empirical results with page-request data from two
different e-commerce sites showed that these pruning techniques provided systematic
but modest improvements in predictive accuracy compared to using the full model.
Techniques from language modeling are also worth mentioning in this context, since
language modeling (e.g. at the word level) can involve both very large alphabets
and dependencies beyond near-neighbors in the data. Of particular relevance here
are empirical smoothing techniques that combine different models from order one to
order k, with weighting coefcients among the k models that are (generally speaking)
estimated to maximize the predictive accuracy of the combined model. See Chen
and Goodman (1996) for an extensive discussion and empirical evaluation of various
techniques and MacKay and Peto (1995a) for a description of a general Bayesian
framework for this problem.
HUMAN BEHAVIOR AND THE WEB 191

Cadez et al. (2003) and Sen and Hansen (2003) propose mixtures of Markov chains,
where we replace the rst-order Markov chain

P (st | st1 , . . . , s1 ) = P (st | st1 ) (7.9)

with a mixture of rst-order Markov chains


K
P (st | st1 , . . . , s1 ) = P (st | st1 , c = k)P (c = k), (7.10)
k=1

where C is a discrete-valued hidden variable taking K values, k P (c = k) = 1,
and P (st | st1 , c = k) is the transition matrix for the kth mixture component. One
interpretation of this model is that users consist of K different groups, each with
different navigation behaviors described by K different Markov chains. Cadez et
al. (2003) use this model to cluster sequences of page requests into K groups. The
parameters of the models are learned using the EM algorithm (see Chapter 1). Sen and
Hansen (2003) use the same type of model motivated by the problem of pre-fetching
of Web pages to reduce user access times.
Consider the problem of predicting the next state, given some number of states t
that have been observed for a given user, where t {1, 2, 3, . . .}. This is the one-
step ahead prediction problem. Let s[1,t] = {s1 , . . . , st } denote the sequence of t
states that have been observed up to t. Under a standard rst-order mixture model,
the predictive distribution on the next state is

P (st+1 | s[1,t] ) = P (st+1 | st ). (7.11)

For a mixture of K Markov models, the predictive distribution is


K
P (st+1 | s[1,t] ) = P (st+1 , c = k | s[1,t] )
k=1

K
= P (st+1 | s[1,t] , c = k)P (c = k | s[1,t] )
k=1

K
= P (st+1 | st , c = k)P (c = k | s[1,t] ). (7.12)
k=1

The last line follows from the fact that, conditioned on component c = k, the next
state st+1 only depends on st . Thus, the Markov mixture model denes the probability
of the next symbol as a weighted convex combination of the transition probabilities
P (st+1 | st , c = k) from each of the component models. The weights are the mem-
bership probabilities P (c = k | s[1,t] ) based on the observed history s[1,t] :
P (s[1,t] | c = k)P (c = k)
P (c = k | s[1,t] ) =  , 1
k
K, (7.13)
j P (s[1,t] | c = j )P (c = j )
192 PROBABILISTIC MODELS OF BROWSING BEHAVIOR

where

L
P (s[1,t] | c = k) = P (s1 | c = k) P (st | st1 , c = k). (7.14)
t=2
Intuitively, these membership weights evolve as we see more data from the user,
i.e. as t increases. In fact, it is easy to show that as t and if the model is
correct, then one of the weights will converge to one and the others to zero. In other
words, if a data sequence becomes long enough it will always be possible to perfectly
identify which of the K components is generating the sequence. In practice, sequences
representing page requests are often quite short. Furthermore, it is not realistic to
assume that the observed data are really being generated by a mixture of K rst-order
Markov chains the mixture model is a useful approximation to what is likely to
be a more complex data-generating process. Thus, in practice, several of the weights
P (c = k | s[1,t] ) will be nonzero and the corresponding models will all participate
in making predictions for st+1 via Equation (7.12).
Comparing Equations (7.11) and (7.12) is informative it is clear from the func-
tional form of the equations that the mixture model provides a richer representation
for sequence modeling than any single rst-order Markov model. The mixture can in
effect represent higher-order information for predictive purposes because the weights
encode (in a constrained fashion) information from symbols that precede the current
symbol.
A value for K in this type of mixture model can be chosen by evaluating the out-
of-sample predictive performance of a number of models on unseen test data. For
example, the one-step ahead prediction accuracy could be evaluated on a validation
data set for different values of K and the  most accurate model chosen.
Alternatively, the log probability score, L1t=1 log P (st+1 | s[1,t] ), provides a some-
what different scoring metric, where a model is rewarded for assigning high predictive
probability to the state that occurs next and, conversely, is penalized for assigning
probability to the other states. This is suggestive of a prediction game, where pre-
dictive models compete against each other by spreading a probability distribution
over the M possible values of the unseen state st+1 . If the model is too condent in
its predictions and assigns a probability near unity to one of the states, then it must
assign a very low probability to other states, and may be penalized substantially (via
log P ) if one of these other states actually occurs at t + 1. On the other hand, if the
model is very conservative and always hedges its bets by assigning probability 1/M
to each of the states, then it will never gain a high log P score. Good models provide
a trade-off between these two extremes.
If we take the negative of the log probability score and normalize by the number
of predictions made, 1/L L t=1 log P (st+1 | s[1,t] ), we get a form of entropy that
is bounded below by zero. The model that always makes perfect predictions gets a
score of log P = log 1 = 0 for each prediction and a predictive entropy of zero,
while the model that always assigns probability 1/M to each possible outcome will
have a predictive entropy of log M. Cadez et al. (2003) describe the evaluation of
predictive entropy for a variety of Markov and nonMarkov models using a test data
HUMAN BEHAVIOR AND THE WEB 193

set of roughly 100 000 page request sequences logged at www.msnbc.com with
M = 17 page categories. The entropy ranged from about four bits (using log to the
base two) for a single histogrammodel down to about 2.2 bits for the best-performing
mixture of Markov chains with about K = 60 components.
A number of other variations of Markov models have been proposed for page-
request predictions, including variable-order Markov models (Sen and Hansen 2003)
and position-dependent Markov models where the transition probabilities depend on
the position of the symbol in the sequence (Anderson et al. 2001). This latter model
can be viewed as a nonstationary Markov chain. The position-dependent model makes
sense for websites that have a pronounced hierarchical structure, where (for example)
a user descends through different levels of a tree-structured directory of Web pages.
Several other variations of Markov models for prediction are explored in Zukerman
et al. (1999), Anderson et al. (2001, 2002), and Sen and Hansen (2003).

7.4.5 Modeling runlengths within states

A notable characteristic of a Markov model is that the run-length distribution (the

distribution on the length of runs of the system in a given state) follows a geometric
distribution. Specically, letting r be the runlength for state i,

P (r) = Tiir1 (1 Tii ), r = 1, 2, . . . ,

since, after one enters state i, the probability of exiting to another state after exactly r
subsequent state transitions requires r 1 self-transitions with probability Tii followed
by a transition out of state i (with probability 1 Tii ). The mean of this geometric
distribution is 1/1 Tii , so that the mean runlength is inversely proportional to the
exit probability.
This geometric form alerts us to a potential limitation of rst-order Markov mod-
els. In practice the geometric model for run-lengths might be somewhat restrictive,
particularly for certain types of states and websites. For example, for a news website
or an e-commerce website, we might nd that observed runlengths in a state such as
weather or business news have a mode at r = 2 or more page requests (since users
may be more likely to request multiple pages rather than just a single page) rather
than the monotonically decreasing shape of a geometric distribution with a mode at
r = 1.
We can relax the geometric distribution on runlengths implied by the Markov
model by using a semi-Markov model. A semi-Markov model operates in a generative
manner as follows: after we enter state i we draw a runlength r from a state-dependent
distribution Pi (r), and then after r time-steps stochastically transition to a different
state according to a set of transition probabilities from state i to the other states. (The
more usual variant of semi-Markov models uses continuous time, where on entering
a state i a time duration is drawn from a distribution on time for that state, > 0).
194 PROBABILISTIC MODELS OF BROWSING BEHAVIOR
= 0.5 =1
0.7 0.4

0.6
0.3
0.5

0.4
P(r)

P(r)
0.2
0.3

0.2
0.1
0.1

0 0
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
Runlength r Runlength r
=2 =5
0.35 0.2

0.3
0.15
0.25

0.2
P(r)

P(r)
0.1
0.15

0.1
0.05
0.05

0 0
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
Runlength r Runlength r

Figure 7.5 Modied Poisson distributions with parameters

(a) = 0.5, (b) = 1, (c) = 2, and (d) = 5.

For example, we could have Pi (r) be a modied Poisson distributed random vari-
able,
r1
Pi (r) = e , (7.15)
(r 1)!
where r is a positive integer, > 0 is a parameter of the Poisson model, and the
distribution has mean + 1 and variance . The standard Poisson distribution (see
Appendix A) has r = 0, 1, 2, . . ., whereas here we use a slightly modied denition
that has r starting at one, so that the mean is also shifted by one (in the standard model
the mean and the variance are both (see Appendix A)).
The most likely run-length (the mode) is the largest integer k such that k
+ 1.
Thus, for example, if we choose > 1 for this model, the most likely runlength will be
two or greater, providing a qualitatively different shape for our runlength distribution
compared to the geometric model which has its mode at one (see Figure 7.5 for
examples). For example, a Poisson distribution with = 0.5 might be appropriate for
websites where visits tend to be relatively short and the most likely number of page
requests is one, while a Poisson distribution with = 5 might be a better model for
websites where most visitors request multiple pages.

7.4.6 Modeling session lengths

We now move from modeling run-lengths within states to modeling overall session
lengths for individual users. In general, an M-state Markov model with an end-state
HUMAN BEHAVIOR AND THE WEB 195
0
10

1
10
Empirical Frequency of L

2
10

3
10

4
10

5
10

6
10
0 1 2
10 10 10
Session Length L

Figure 7.6 Histogram on a log-scale of session length for 200 000 different sessions at the
www.ics.uci.edu website in early 2002. Session boundaries were determined by 20 min
or longer gaps between page requests, and site visitors that were clearly identiable as robots
were removed.

allows us to generate a variety of distributions on the lengths of strings (sessions), by

varying the parameters of the transition matrix T . While such a model may be quite
adequate in practice from a purely predictive viewpoint, in terms of understanding
the underlying data-generating process such a Markov model may not provide much
insight. With this in mind, there is motivation to explore general parametric models
for session lengths and this is the focus of this section.
Let L represent the session length, L {1, 2, 3, . . .}. Figure 7.6 shows observed
session length data from the www.ics.uci.edu website, plotted on a loglog
scale. This gure is suggestive of a linear relationship between log P (L) and log L,
which in turn suggests a power-law distribution
P (L | ) = CL (7.16)
as introduced
 in Chapter 1, where > 1 is the single parameter for this model, and
C = 1/ L is a normalization constant, known as the Riemann zeta function,
which in general can only be evaluated in closed form for even integer values of .
This distribution is known in statistics as the zeta distribution, dened on positive
integers L = 1, 2, 3, . . . .
We can use the technique of ML estimation once again to estimate the parameter
from observed data (e.g. from the data in Figure 7.6). Since we are only estimating
a single parameter and have a relatively large amount of data available, we expect
that ML and Bayesian estimation would yield roughly similar results so we use the
simpler ML technique here. The likelihood function, L( ) is simply N i=1 P (Li | ),
196 PROBABILISTIC MODELS OF BROWSING BEHAVIOR
5
x 10

3.325

3.33
Log-Likelihood

3.335

3.34

3.345
1.9 1.95 2 2.05 2.1
Parameter in the power-law model

Figure 7.7 The log-likelihood as a function of the unknown parameter ,

under a power-law model, for the session-length data in Figure 7.6.

given an observed set of session lengths {L1 , . . . , LN } that are assumed to be con-
ditionally independent given the model, where P (Li | ) is dened as in Equa-
tion (7.16). Differentiating this expression (or the log-likelihood) with respect to
, and setting it to zero to nd the maximum, doesnot yield a closed-form solu-
tion for the ML value for , since the term C = 1/ L cannot be reduced to a
closed-form function of .
The solution can, however, be obtained by standard numerical methods or, since
we only have a single parameter to t, we can just visually examine the log-likelihood
as a function of and read-off an approximate value for ML . Figure 7.7 shows the
log-likelihood function, under the power-law distribution model, as a function of ,
for the empirical data shown in Figure 7.6, where we see that the maximum value
occurs at 2 ( = 1.98 to be precise). Under the power-law model, maximum
likelihood suggests that P (L) L1.98 . Figure 7.8 shows this model superposed on
the empirical data.
Note that an alternative technique for tting the parameter of a power-law dis-
tribution to empirical data is to use linear regression directly, i.e. to estimate the best-
tting slope for the line in Figure 7.6 with a least-squares penalty function. While this
may seem like a reasonable approach, it can provide somewhat different results than
maximum likelihood, since it in effect places equal weight on the error in predicting
the count for each value of L, even though (for example) small values of L with large
counts may account for a much larger fraction of the data than large values of L with
very small counts. Maximum likelihood on the other hand seeks parameter values
that maximize the probability of the data as a whole, rather than equally weighting
the t for different values of L. In many respects this is a more suitable criterion than
HUMAN BEHAVIOR AND THE WEB 197
0.7

0.6

0.5
Probability of r

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10 12 14 16 18 20
Session Length r
0
10

1
10

2
10
Probability of r

3
10

4
10

5
10

6
10
0 1 2
10 10 10
Session Length r

Figure 7.8 Comparing the tted power-law model (dotted line) using an ML parameter
estimate of = 1.98 with the observed empirical session lengths (dots), on both (a) standard
scales, and (b) loglog scales.

the implicit equal-weighting criterion of least-squares. Thus, in assessing the t in

Figure 7.8 (on the loglog plot in particular) one should pay more attention to the
quality of the t for the smaller values of L (to the left), since this is the region that
accounts for the bulk of the sessions. One could of course also use a weighted least
squares penalty function, where the weight on the error for predicting the count at L
198 PROBABILISTIC MODELS OF BROWSING BEHAVIOR

is multiplied by the size of the count at L, encouraging smaller prediction errors at

the places where there are more data.

7.4.7 A decision-theoretic surng model

Huberman et al. (1998) take a decision-theoretic approach to modeling navigation
behavior at a website, and session lengths in particular. In their model, they consider
the decision of a site-visitor visiting a particular page on whether to continue surng
to another page or to end the session. The utility to the user of viewing another page
is modeled as the utility of the current page plus an additive Gaussian random term.
Under this model, the decision on whether to continue or not depends on whether the
expected cost of continuing is greater or less than the discounted expected utility of
information to be gained from continuing. When the utility of the current page falls
below some (negative) threshold, it is no longer optimal to continue. This leads to a
stochastic random-walk type of model for user navigation. Under these assumptions
on utilities and decision-making, Huberman et al. (1998) show that the distribution of
session lengths will follow a particular functional form known as an inverse Gaussian
distribution:   
(L )2
P (L) = exp , (7.17)
2L3 22 L
where L 1, 2, 3, . . . is the session length as before, > 0 is the mean value and
is a scale parameter (the variance of the distribution is inversely proportional to ).
The inverse-Gaussian distribution arises in statistical modeling of various physical
problems and, strictly speaking, it denes a probability density function for all positive
values of L, whereas here we have positive integers L for session lengths. Nonetheless,
Huberman et al. argued that it provided a better t to session-length data from two
different websites than alternatives such as a simple geometric model or a power law.
We can use maximum likelihood once again to estimate the parameters of this
inverse Gaussian model. It is straightforward to show that, given a data set D =
{L1 , . . . , Ln } of n session lengths, the likelihood is maximized at the values

1
n
ML = Li , (7.18)
n
i=1

i.e. the ML estimate of the mean under this model is the average session length, and
  n  
1 1 1 1
ML = . (7.19)
n Li
i=1

In Figure 7.9 we show the t of the inverse Gaussian distribution and three of the
other models we have discussed earlier to the session length data from Figure 7.6.
From this plot we see that the power-law t matches the general shape of the empirical
data distribution more closely than any of the other models. The Poisson model has the
HUMAN BEHAVIOR AND THE WEB 199
0
10

POISSON

1
10

GEOMETRIC
2
10
Probability of L

INVERSE GAUSSIAN

3
10 POWER-LAW

4
10

5
10

6
10
0 1 2
10 10 10

Session Length L
Figure 7.9 Comparing a variety of simple parametric session-length models with the observed
empirical session lengths (dots), on a loglog scale, for the session-length data in Figure 7.6.

worst t, with a mode at about r = 3. The geometric and inverse-Gaussian models both
fail to capture the tail properties of the distribution, underestimating the probabilities
of long session lengths. The power-law model is far from a perfect t, but nonetheless
it does capture the general characteristics of the session-length distribution with only
a single parameter. Sen and Hansen (2002) discuss other approaches to modeling
session lengths, such as higher-order Markov models.

7.4.8 Predicting page requests using additional variables

In the previous sections we looked at models that characterize the distribution of
session lengths on a website and that try to predict the page a user will go to next
based on pages visited in the past. It is natural to consider the incorporation of other
predictor variables (sometimes referred to as covariates) into our characterization of
navigation behavior. Relevant covariates might include
(a) the number of times a user has visited the site before (users with prior visits
might tend to navigate the site more quickly and have shorter sessions);

(b) various demographic characteristics of users such as age, education level; and

(c) the bandwidth available to the user.

200 PROBABILISTIC MODELS OF BROWSING BEHAVIOR

Bucklin and Sismeiro (2003) propose a covariate model of this form for site nav-
igation behavior. As with the Huberman et al. model, they also consider a decision-
theoretic framework for modeling the utility of continuing to a new page. However, in
the Bucklin and Sismeiro model, this utility is modeled as a stochastic function of var-
ious covariates rather than just being a stochastic function of the utility of current and
previous pages (the Huberman et al. approach). The covariates in the model include
the number of previous site visits, the visit depth(number of pages viewed on the site
during this session up to this point), as well as various system variables that take into
account the complexity of a page, how long it takes to download, whether it contains
dynamic content, and so forth. In the proposed model, both the utility of requesting an
additional page and the logarithm of the page-view duration (conditioned on being at
a particular page in a particular session), are modeled as linear functions of individual
and page-specic covariates, plus an additive random noise term.
A useful general concept in the work of Bucklin and Sismeiro is heterogeneity of
behavior across users. Rather than estimating a single set of model parameters that
describe all site-users, a more general approach is to allow each user to have their own
parameters. If we had a large amount of site-visit data for each individual user, then
in principle we could estimate the parameters for each user independently using their
own data. However, in practice we are usually in the opposite situation in terms of
browsing behavior, i.e. we typically have very few data for many of the site visitors. In
this situation, estimating model parameters independently for each user (using only
their own data) would result in highly noisy parameter estimates for each individual.
A general statistical approach to this problem is to couple together the parameter
estimates in a stochastic manner, based on the simple observation that the behavior of
different individuals will be dependent although different. In classical statistics this is
known as a random effects model individual behavior is modeled as a combination
of a common systematic effect with an individual-specic random effect.
In Bayesian statistics a somewhat more general approach is taken by specifying
what is called a hierarchical model. In the hierarchical model we can imagine for
simplicity that there are just two levels to the hierarchy: at the upper level we have
a distribution that characterizes our belief in how the parameters of the model are
distributed across the population as a whole; at the lower level we have a set of n
distributions describing our belief in the distributions of parameters for each of n
individuals. Both the population distribution and the n individual distributions can be
learned from the observed data in a Bayesian manner. The upper level distribution
in effect acts as a population prior (learned from the data) that constrains the infer-
ences we make on specic individuals. This population distribution typically acts to
constrain the parameter estimates that we would get from looking at individual is
data alone, so that they are closer to those of the population as a whole. A proper
description of Bayesian hierarchical models is beyond the scope of this text. A good
introductory treatment can be found for example in Gelman et al. (1995).
For Web data analysis the use of hierarchical models (or random effects) is quite
natural given that Web browsing behavior from populations of users can have both
common patterns as well as signicant individual differences. Bucklin and Sismeiro
HUMAN BEHAVIOR AND THE WEB 201

Table 7.2 Toy example of query data from a search log.

User ID Time stamp Query

US12635123 1998 09 16 18 8102 World Cup soccer

YR19281101 1998 09 16 18 8312 Markov
YR19281101 1998 09 16 18 8410 Markov model
YR19281101 1998 09 16 18 8557 Markov state
FT29172531 1998 09 16 18 8876 Canary
FT29172531 1998 09 16 18 8992 Canary Islands

adopt a Bayesian hierarchical model approach in estimating the parameters of their

covariate-dependent model. Their study was based on logged server-side data from
5000 visitors to a commercial website in the automotive industry, with 6630 sessions
consisting of two or more page visits, and 40 560 page requests. These data were
randomly sampled from a larger Web server log of 170 000 visitors in October 1999.
Using a Bayesian parameter estimation approach with the covariate linear models
mentioned earlier, the authors were able to make a number of interesting inferences
from the data about site-navigation behavior. For example, they found that there was
signicant variation in the parameters of the models t at the individual level, support-
ing the hypothesis that signicant heterogeneity exists among different individuals
in terms of how they browse a website (see also Ansari and Mela 2003). They also
found that repeat visits to a website have the effect of reducing the number of pages
viewed but do not affect the length of time spent viewing a page. This can be thought
of as a type of learning effect over multiple visits as users become more familiar
with a site their sessions tend to get shorter in length.
Bucklin and Sismeiro also found evidence of a within-site lock-in effect, as evi-
denced by a positive correlation in their model between visit depth and the mean
length of time spent viewing a page users who explore more pages tend to spend a
longer time reading them. Such inferences from empirical data must be treated with
some caution, given the relatively noisy nature of server data (as discussed earlier)
and the fact that only a single website was used in the study. Nonetheless as both the
quality and quantity of Web data continue to improve, these types of analytical studies
are increasingly likely to provide substantial insights into site navigation behavior.

7.5 Modeling and Understanding Search Engine

Querying
Up to this point in the chapter we have focused primarily on navigational behavior.
An implicit driver for much of our online navigation is the search for some specic
information goal or set of goals. In the nal section of this chapter we look in depth
at one aspect of online search, namely, how users issue queries to search engines.
202 MODELING AND UNDERSTANDING SEARCH ENGINE QUERYING

7.5.1 Empirical studies of search behavior

We can gain a broad understanding of the general characteristics of how search queries
are issued by analyzing the search query logs of large commercial Web-based search
engines. A number of different studies have followed this general approach, including
Jansen et al. (1998), Silverstein et al. (1998), Lau and Horvitz (1999), Spink et al.
(2002), and Xie and OHallaron (2002). These studies are typically based on the
search query logs recorded by the search engine server and usually consist of three
pieces of information per individual search query: a timestamp for when the query
was received, the text string corresponding to the query, and a unique user ID (see,
for example, Table 7.2).
The user ID is often a proxy used to replace the actual IP address from which the
query is issued, for anonymity purposes. The term user could be a little mislead-
ing since, for example, multiple individuals could be issuing queries from a single
IP address. Xie and OHallaron (2002) used the user-agent eld in the Web server
log les for a particular search engine to check how many queries were coming from
dial-up services such as America Online (AOL). They found that out of about 111 000
queries in total, roughly 2.7% were originating from AOL, a percentage which is rela-
tively small but which nonetheless does indicate that data from dynamically assigned
IP addresses can represent a nontrivial portion of log data and could in principle
somewhat skew the overall statistics. Thus, as discussed earlier in the context of nav-
igational studies, some caution in the interpretation of the empirical data is warranted
since a set of queries from a particular user might in reality be a set of individuals
all sharing a particular IP address.
Similarly, since some search services answer search queries by querying other
search engines and aggregating the results, a user need not be a human searcher.
The presence of non-human users (software programs such as robots) can signicantly
skew observed counts in the raw data, as illustrated earlier in Figure 7.1. For example,
the Excite trace in Xie and OHallaron (2002) contained 130 200 queries from one
particular IP address over 8 h it is clearly not possible for a human to generate that
many search queries in that amount of time. This type of outlier can make summary
statistics such as means and standard deviations somewhat unreliable and emphasizes
the need to examine the raw data, by plotting histograms for example, before any
serious analysis or modeling.
The size and temporal extent of the search engine query logs analyzed in the
different studies varies greatly.

Lau and Horvitz (1999) analyze a sample of 4690 queries from a set of approx-
imately one million queries collected by the Excite search engine in September
1997.

Silverstein et al. (1998) analyze approximately one billion entries in the query
log for the AltaVista search engine collected over six weeks in August and
September 1998.
HUMAN BEHAVIOR AND THE WEB 203

Spink et al. (2002) describe a series of studies over multiple years involving
the Excite search engine, with query logs collected at different time-points in
September 1997, December 1999, and May 2001. The query logs analyzed at
each time-point typically each contain about one million queries.
Xie and OHallaron (2002) analyze 110 000 queries from the Vivisimo search
query logs collected over a 35-day period in January and February of 2001, and
1.9 million queries collected over 8 h in a single day in December 1999 from
the Excite query logs.
We will discuss below the main results from these four studies. Other empirical
studies have also been reported generally speaking, the inferred characteristics of
query patterns across all of these studies are in broad agreement.

General characteristics of search query logs

The following general characteristics consistently appear across all of these studies.
The average number of terms in a query is quite small, ranging from a low of
2.2 to a high of 2.6 across the studies.
The most common number of terms in a query (the mode) in all four of the
studies is two. In fact the distribution of the number of terms in a query is
quite consistent across the different studies. Figure 7.10 plots the empirical
histograms of the number of terms per query for both the Excite and Vivisimo
data sets in the study by Xie and OHallaron. The histograms have been nor-
malized by the total number of terms across all queries so that for each query
length L = 1, 2, 3, . . . the height of each bar represents an empirical estimate
of the probability P (L).
The modied Poisson model, dened in Equation (7.15) has also been tted to
each of the data sets as a simple hypothetical model for the query length data.
Recall that the modied Poisson distribution has a single parameter which
we estimate here as
1 
Li 1,
N
where N is the total number of queries and Li is the length of the ith query. The
model provides a reasonable t to both data sets quite close to the observed
values for the Vivisimo data, not quite as close for the Excite data. We can see
that for both search engines the distribution of query lengths is qualitatively
similar (i.e. mode at L = 2, signicant mass for queries that are three or shorter,
etc.). The tails of the empirical distributions and the Poisson models are in less
agreement and diverge somewhat from each other if we plot the data on a
log-scale on the Y -axis (not shown).
The majority of users do not rene their query based on feedback from the
results obtained with the initial query (e.g. by adding terms or modifying the
syntax of the query). Silverstein et al. (1998) report that 77.6% of users issue
204 MODELING AND UNDERSTANDING SEARCH ENGINE QUERYING
0.4

0.35
Vivisimo Query Lengths

0.3 Estimated Poisson parameter = 2.2249

0.25
Probability, P(L)

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18 20
L = Query Length
0.4

0.35
Excite Query Lengths

0.3 Estimated Poisson parameter = 2.633

0.25
Probability, P(L)

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18 20
L = Query Length

Figure 7.10 Comparison of empirical query length distributions (bars) and tted Poisson
models (dots and lines) for both Vivisimo query data (top) and Excite query data (bottom). The
data were provided by Xinglian Yie and David OHallaron.

only a single query in a particular session, and Lau and Horvitz (1999) found
similar results with only about 25% of the users in their study modifying queries.
Spink et al. (2002) report that the number of users who viewed only a single page
of Excite results (the top 10 websites as ranked by the search engine) increased
from 29% in 1997 to 51% in 2001. In the AltaVista study of Silverstein et al.
(1998) fully 85% of users were found to view only the rst page of search results.
HUMAN BEHAVIOR AND THE WEB 205

The difference in numbers between the two studies could reect two different
types of user populations (Excite users versus AltaVista users). It could also
reect a difference in methodology in terms of how number of pages viewed
was dened and measured in each study.
There appears to have been a shift in the distribution of query topics over
time. Both Lau and Horvitz (1999) and Spink et al. (2002) manually classied
samples of several thousand queries into a predened set of 15 information
goals and 11 general categories, respectively. Both studies found that about
one in six queries in the 1997 samples concerned adult content. In contrast,
in the 2001 sample in the Spink study this gure had dropped to 1 in 12. The
top two topics of commerce, travel, employment, or economy and people,
places or things in the 2001 sample accounted for 45% of all queries. This is a
considerable increase from the 20% share for the same two query topics in 1997,
reecting perhaps the increased use of the Web for business and commercial
purposes. Conversely the top-ranked topic in 1997 with 20% of the queries,
entertainment or recreation, had dropped to a 7% share by 2001. It should be
noted that these topic-related gures are based on a particular sample of users
(about 2500 per sample) at a particular search engine (Excite), and relied on a
subjective process of manual assignment of query terms to categories.
Despite differences in sample populations, search engines, and dates of the studies,
these four studies produced a generally consistent set of ndings about user behavior
in a search-engine context. For example, it is clear from the data that most users view
relatively few pages per query and most users do not use advanced search features.

Power-law characteristics of repeated queries

Consider a set of observed queries Q = {q1 , . . . , qN }, and let n(qi ), 1
i
N ,
be the number of observations (or count) of query qi . If we sort the queries by their
counts, we can dene C(r) as the count for the query of rank r, e.g. C(1) is the
number of times the most frequent query occurred, and so on. Equivalently we can
dene f (r) as the empirical frequency of queries with rank r, calculated as C(r),
the number of times the query of rank r occurs, divided by the total number of
queries.
Clearly f (r) is a non-decreasing function of r; it varies inversely with r by deni-
tion. As discussed earlier in Chapter 1, for a variety of different empirical phenomena,
f (r) is often found to be well described by a power-law, i.e. f (r) r .
Xie and OHallaron (2002) ranked queries by their frequency of occurrence for both
the 110 000 queries from Vivisimo and the 1.9 million queries from Excite. Thus, the
rst query is the query that is most commonly issued, with a frequency count of
approximately 100 in the Vivisimo data, and 10 000 in the Excite data, and so on.
Plotting the log-frequency, log f (r), on the y-axis versus the log of the rank r on the
x-axis reveals a near linear relationship in Figure 7.11, so that log f (r) a log r + b.
Once again we see power-law phenomena cropping up in another aspect of Web
206 MODELING AND UNDERSTANDING SEARCH ENGINE QUERYING
2
10

Empirical Probability per Rank Vivisimo Query Data

3
10

4
10

5
10
0 1 2 3 4 5
10 10 10 10 10 10
Rank of the Query
2
10

Excite Query Data

3
10
Empirical Probability per Rank

4
10

5
10

6
10
0 1 2 3 4 5 6
10 10 10 10 10 10 10
Rank of the Query

Figure 7.11 Frequency of different queries issued as a function of rank, for both Vivisimo
query data (top) and Excite query data (bottom). The data were provided by Xinglian Yie and
David OHallaron.

data analysis. Furthermore although the two curves are from two different search
engines they are quite similar to each other. This again suggests that there are strong
regularities in terms of patterns of behavior in how we search the Web, and these
patterns appear to be relatively independent of the particular search engine being
used.
HUMAN BEHAVIOR AND THE WEB 207
Informational Goals

Current Search Action

Time Interval

Next Search Action

Figure 7.12 A graphical representation of the simple Bayesian network model

used by Lau and Horvitz (1999) to model a users search query actions over time.

7.5.2 Models for search strategies

In order to design search engines that can better serve the information needs of indi-
vidual users it is important to understand the process by which a typical user navigates
through search space when looking for information using a search engine. The four
studies described earlier provide a general idea in terms of summary statistics of
general characteristics of user behavior in an aggregate sense. Ideally, however, we
would like to understand the search process in more detail. Lau and Horvitz (1999)
describe a simple probabilistic model (a directed acyclic graph, or Bayesian network
as dened in Chapter 1) for modeling how a users search actions evolve over time.
The four nodes in the network are:
(1) current search action, the rst of two adjacent (in time) search actions, tak-
ing values new query, reformulation, generalization, specialization, additional
results and interruption;
(2) time interval, a discrete-valued variable representing the elapsed time in seconds
between the rst user action and the second (if any), ranging in value from
1020 s to more than 20 min;
(3) next search action, taking the same values as user search action;
(4) informational goals, taking 15 different values such as current events and
education.
The structure of the model proposed by Lau and Horvitz is illustrated in Figure 7.12
in the form of a Bayesian network. The conditional probability tables for each variable-
208 EXERCISES

value pair, given the variable-value combinations of all of its parents, were estimated
by frequency counts (maximum likelihood) directly from the sample of 4690 queries in
the study. This particular model imposes a specic (and somewhat natural) ordering
on the variables (e.g. informational goals precedes all others) but otherwise does
not make any conditional independence assumptions. More generally, if there were
more variables in a model like this, a sparser graph (reecting various conditional
independence assumptions) would likely be used.
Although quite a simple model, Lau and Horvitz found that this network was able
to produce interesting and potentially useful predictions. For example, according to
the model, if a query specialization is followed by another action 1020 s later then
the user is most likely to be searching for entertainment-related information. The
conditional probability for entertainment given the user action and time-interval
information is signicantly higher than the marginal probability of entertainment.
Such correlations between sequences of actions, time between actions, and informa-
tion goals, could in principle allow the search engine to make inferences about the
likely search trajectory of individual users based on their action history. In turn,
these inferences could then be used to provide more relevant feedback to the user in
real time and could also be used for marketing purposes such as real-time targeted
advertising.

7.6 Exercises
Exercise 7.1. Obtain Web server logs from your organizations Web server (if you
can) for one days worth of data and carry out the following analyses:
estimate the fraction of page requests that are coming from robots;
segment the data into sessions, using a 20 minute time-out rule for declaring
the end of sessions, and plot the empirical distribution of session lengths;
t at least two of the following models to the session-length data: geometric,
Poisson, inverse Gaussian, and power-law. Plot and comment on the results.
Exercise 7.2. Derive the mean posterior estimate in Equation (7.7) from rst princi-
ples.
Exercise 7.3. Prove that the ML estimate of the parameter for the Poisson distri-
bution is the empirical mean (sample average) of the data.
Exercise 7.4. Construct three different session-length distributions by designing a
transition matrix T for a four-state Markov chain (with one of the states being an
end state), where each of the distributions should be quite different from the other
two. Write a program to simulate data from each of the Markov models and construct
histograms of the session lengths for each model from your simulations.
Exercise 7.5. Discuss how the Lau and Horvitz (1999) Bayesian network model for
search queries and the Bucklin and Sismeiro (2003) model for site navigation might
HUMAN BEHAVIOR AND THE WEB 209

be combined to provide a more general model that encompasses both navigation and
searching.
Exercise 7.6. We mentioned that for large M (e.g. a large number of pages M) esti-
mating the O(M 2 ) parameters of a rst-order Markov chain may be quite impractical.
Specify in detail two different algorithms for clustering M pages into K states, for
the purposes of estimating a reduced category K-state Markov chain for modeling
navigation behavior.
Exercise 7.7. Discuss whether the rst-order Markov models discussed in Section 7.4
can adequately model the process of users using the back button in their browsers
when browsing. If you believe that the model is not adequate, discuss how one might
develop a stochastic model that can handle this aspect of user behavior.
Exercise 7.8. Derive from rst principles the ML estimates for the inverse Gaussian
distribution in Equations (7.18) and (7.19).
Exercise 7.9. None of the models in Figure 7.9 t the data perfectly well, and indeed
to the right of the data (for session lengths greater than 102 ) there may be some outliers
present. Comment on what aspects of real data sets these simple models may not be
accounting for.
Exercise 7.10. In the techniques described in this chapter we did not discuss the issue
of an individuals right to privacy. Research the issue of privacy in the context of Web
data analysis and write an essay on the current state of affairs in your country in terms
of privacy and the individual surfer. For example, if a surfer visits your website what
does the law say about who owns the logged data? What are the laws (if any) that
govern how Web servers can issue cookies? And so forth.
Exercise 7.11. Imagine that you are conducting a study on Web navigation behavior
10 years from now, with client-side data and demographic information for 100 000
individuals. Predict how you think both navigation and search behavior will have
changed by then, compared to the data in the studies described in this chapter.
This Page Intentionally Left Blank
8

Commerce on the Web:

Models and Applications

8.1 Introduction
In the late 1990s the Web quickly changed from being primarily an academic pursuit
to an increasingly commercial one. This rapid commercialization of the Web, in
areas such as e-commerce, subscription news services, and targeted advertising, has
led to the infusion of the Web into modern daily life. While commercialization can
have its negative aspects, on the plus side it presents signicant new challenges and
opportunities for academic researchers. Some examples of these are given in the
following list.
Can we design algorithms that can help recommend new products to site-
visitors, based on their browsing behavior?
Can we better understand what factors inuence how customers make purchases
on a website?
Can we predict in real time who will make purchases given their observed
navigation behavior?
In this chapter we will investigate these and related questions. Once again, as in
Chapter 7, what is driving this research is the availability of vast amounts of raw data.
For a standard bricks-and-mortar retail store, the only information about a customers
behavior within a store is usually in the form of scanner data or market basket data,
namely, the list of items and prices that were scanned when the customer paid for
their purchases. In contrast, for a store on the Web, a sequence of clickstream data
is typically available for each customer prior to making a purchase, in addition to
the purchase data (if the customer makes a purchase). This provides a much richer
description of a customers behavior over time. Naturally this clickstream data can
itself be quite noisy, but nonetheless the volume of data available is often large enough
(e.g. millions of customers per day visiting very large e-commerce sites such as
www.amazon.com) that certain patterns clearly emerge even from noisy data.

Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
212 CUSTOMER DATA ON THE WEB

8.2 Customer Data on the Web

Data relevant to e-commerce can be obtained at the server side, at the client side,
or at points in between (for example, by ISPs). Server-side data for e-commerce is
similar in form to standard Web logs (i.e. a list of events that are timestamped, that
record the page requested, the IP address requesting the page, and so forth), with
the addition of purchase information if the customer made a purchase, as well as
personal information on the customer such as address and credit card information.
The purchase information is of course the primary event of interest to the retailer we
can, for example, try to estimate in real time whether a customer is likely to make a
purchase or not as they begin to navigate the site. A predictive classication model
for this purpose could be built from historical data on visitors to the site. We could
also correlate personal information, such as city or country of origin, age, income
(if known), with purchase behavior to try to get a better understanding of who is
purchasing different types of products.
E-commerce logs are often more reliable by necessity than standard Web logs. For
example, many e-commerce sites require that a site-visitor have cookies enabled,
so that sessions can be reliably tracked and customers reliably identied. Many
e-commerce sites personalize the information being presented to the site-visitor, and
to do so they must identify who the customer is early on in the session and link
them to a database of prior information. This identication can happen implicitly via
cookies, or explicitly via customer login to the site. The overall effect is much more
reliable identication of the site-visitor from session to session than with a typical
non-commercial site.
Marketing researchers also nd client-side data quite useful for modeling customer
behavior. For example, how many websites (for a particular class of product) do
customers typically visit before making a purchase? As discussed in Chapter 7, one
source of such client-side data is companies such as Nielsen and comScore that provide
industry data on Web usage. These companies select a representative random sample
of Web users (e.g. perhaps 50 000 users in the United States), who agree to have their
Web usage recorded over a period of time, usually by installing special software on
their computers. This type of data, where the behavior of a set of selected individuals
is tracked over time, is referred to as panel data in economics and marketing. We will
look in detail at a sample of such data in Section 8.5 later in this chapter.

8.3 Automated Recommender Systems

A particular topic which has attracted considerable research attention in e-commerce
in recent years is the problem of building automated recommender systems for Web
applications. There are two main variants of this problem. In both variants we typically
have a large number n of individuals and a large number m of items.
In the rst variation of the problem, the observed data can be thought of as a very
sparse binary n m matrix V , where vi,j = 1 indicates that individual i purchased
item j , and the entry is assumed to be zero otherwise. See Table 8.1 for a toy example.
COMMERCE ON THE WEB 213

Table 8.1 A toy matrix R of binary votes, where items correspond to columns and users
correspond to rows. If entry (i,j ) is 1 it means that the ith user voted positively for the j th
item. A blank entry means that no positive vote was recorded. Note the sparsity of the data.

User 1 1 1
User 2 1 1 1
User 3 1 1
User 4 1 1
User 5 1
User 6 1
User 7 1
User 8 1
User 9 1 1
User 10 1 1
User 11
User 12 1 1 1 1
User 13 1
User 14 1
User 15 1 1

In the second variation of the problem the items are not products that are purchased,
but instead are items that are rated by the user, either explicitly (e.g. movies rated on
a scale of one to ve) or implicitly (e.g. documents in a digital library rated by how
long the user is estimated to have spent viewing the document).
In either case the matrix V is usually very sparse, since the typical user will only
have purchased or rated a very small fraction of the overall number m of items. For
example, Popescul et al. (2001) report a study of recommender systems applied to
the ResearchIndex online scientic literature database (Lawrence et al. 1999), with
n = 33 050 users accessing m = 177 232 documents. Each user accessed on average
only 0.01% 18 documents in the database, so that 99.99% of the possible useritem
pairs are in effect not present.
An interesting issue is the treatment of missing data in the matrix V for items j
that were not voted on by user i, where missing is considered to be all the zeros.
These missing data are not missing completely at random but instead are likely to
be affected by a form of negative selection bias, where users may be more likely to
not vote on items that they do not like (Breese et al. 1998). In much of the work
on recommender systems this bias is not explicitly dealt with. We also choose to
conveniently ignore this issue in the discussion below, under the assumption that it
may be somewhat of a second-order factor in modeling the data compared to the issue
of how to handle the positive votes.
In our discussion we will generically refer to the useritem pairs vi,j as votes,
keeping in mind that votingmight represent purchasing an item on a website, access-
ing a document on a website, or rating an item such as a piece of music or a movie
on a website. All of these actions can be loosely interpreted as positive votes for the
item in question, whether implicit or explicit.
214 AUTOMATED RECOMMENDER SYSTEMS

The recommender problem is to be able to automatically recommend and rank a

list of new items to a user based both on items this user has already voted on, as
well as past voting patterns of other users. This can be viewed as a high-dimensional
prediction problem where the training data from which to build a model is very sparse.
Let a be the active user for whom we wish to make predictions and let Ia be the set
of items that user a has already voted on (e.g. the items already in as electronic
basket, while purchasing books at an online bookstore). I is the total set of items.
The prediction problem is to predict which of the items in the set I \ Ia the user is
most likely to provide a positive vote for, if they were asked to vote on that item.
Typically the number of items in the set I \ Ia is quite large. For example, in
the document database example earlier, the average user only voted on or accessed
18 documents, leaving 177 214 other documents that we need to rank in terms of how
the user might like them. Thus, the computational aspects of recommender systems
can be quite challenging, especially given that on a website the ranking needs to be
computed in real time and for potentially thousands of different users who are visiting
the site at any given time. Schafer et al. (2001) discuss various other important practical
aspects of implementing recommender systems for real e-commerce applications.

8.3.1 Evaluating recommender systems

An important issue is that of evaluating the effectiveness of any particular recom-
mender algorithm. Large e-commerce retailers conduct live experiments with real
customers to test different algorithms, but academic researchers do not usually have
the ability to attract large numbers of users to websites for experimental purposes.
Unfortunately, from a research viewpoint the results of e-commerce experiments are
rarely published for competitive reasons. For this reason recommender systems in
research papers are evaluated based on historical data rather than actual recommen-
dations to real customers. There is an important issue here: a system that recommends
item X for individual a will be rewarded if in the test data individual a did indeed
purchase that item. However, its entirely possible that the individual would have
purchased the item anyway (whether recommended or not), or indeed the perverse
situation might arise where the recommendation might actively discourage individual
a from purchasing a product that they would have otherwise purchased!
This is an important caveat to keep in mind in terms of how recommender systems
are evaluated, particularly when the votes represent purchases. The point is aptly
summarized by Zhang and Iyengar (2002):

the true value of a recommender system can only be measured by con-

trolled experiments with actual users. Such an experiment could measure
the improvement achieved by a specic recommendation algorithm when
compared to, say, recommending the most popular item. Experiments
with historical data have been used to estimate the value of recommender
algorithms in the absence of controlled live experiments . . . .
COMMERCE ON THE WEB 215

Many different ideas have been proposed as the basis for recommender algorithms,
the main ones being nearest-neighbor and model-based collaborative ltering. We
discuss both of these approaches as well as some related techniques in the next few
sections.

8.3.2 Nearest-neighbor collaborative ltering

Basic principles
The term collaborative lteringgenerally refers to the use of prior votes of individuals
to predict new votes and make new recommendations. The nearest-neighbor approach
to collaborative ltering (Resnick et al. 1994; Shardanand and Maes 1995) uses a
memory-based approach for prediction: for user a, nd the most similar users to a
in the training data matrix V , and then use the votes of these neighbors to ll in
predictions for a on the items that a has not yet voted on. This is somewhat similar
in concept to the K-nearest neighbor classier described in Chapter 4.
A variety of different approaches have been proposed for (a) nding neighbors,
(b) combining their histories (often using some form of weighting), and (c) making
predictions. A typical scheme is as follows: let
1 
vi = vi,j , j I \ Ia ,
|Ii |
j Ii

be the mean vote for user i, 1
i
n, where Ii is the set of items that user i has
voted on (i.e. vi,j > 0 for j Ii , and vi,j = 0 otherwise). It is customary to calibrate
user is votes by subtracting the mean vote, to create an adjusted vote matrix

vi,j = vi,j vi
so that if (for example) one user tends to vote high and another tends to vote low, this
systematic mean difference is removed.
We can then make predictions for the vote of new user a on items j that he or she
has not voted on by using

1 
n

va,j = va + wa,i vi,j ,
C
i=1
n
where C = i=1 |wa,i | is a normalizing constant. The weights wa,i are dened by
a function that estimates how similar user i is to user a. These weights are usually
calculated based on the similarity of the votes that a has already made on items in
Ia and the votes of other users on this same set of items. The overall effect of the
prediction equation is to generate a predicted vote for items j , where the prediction is
more heavily weighted toward votes of individuals who are have similar past voting
histories to a. Once the predicted vote for each individual item in I \ Ia is calculated,
these predictions can then be ranked and the highest ranked items can be presented
to user a.
216 AUTOMATED RECOMMENDER SYSTEMS

Dening weights
One way to differentiate between different collaborative ltering algorithms is in
terms of how the similarity weights are dened in the prediction equation above. One
widely used weighting scheme (Resnick et al. 1994) is the correlation coefcient
between users i and a, dened as
1 
wa,i = (va,j va )(vi,j vi ),
C2
j

where  1/2

C2 = (va,j va )2 (vi,j vi )2
j j

is another normalizing constant, so that 1
wa,i
1. The sums over items j

in the expressions above are on items that both user a and i have voted on. If there
are no common items in a and is voting histories, then wa,i = 0 by default. This
alerts us to a potential limitation of nearest-neighbor methods namely, when the
intersection of Ia and Ii is small the weight will be based on relatively few matching
items and may not provide reliable predictions. One technique to get around this
problem is to introduce default votes on certain items that neither a nor i have voted
on (e.g. popular items). We can view this as lling in default values for missing data.
From a statistical modeling perspective this may be somewhat suboptimal and a more
principled alternative may be to estimate the missing values in a model-based fashion
(Little and Rubin 1987). For example, we could use a Dirichlet prior to smooth out
the votes in a manner analogous to the way we discussed the use of such a prior for
generating reliable estimates of transition probabilities in Chapter 7. We will discuss
model-based recommender algorithms later in this chapter.
Another example of a weighting scheme is to use vector similarity from information
retrieval, the cosine of the angle formed by two vectors of word frequencies (see
Chapter 4), where the vote vectors for individual i takes the place of word frequencies
(Breese et al. 1998):
 va,j vi,j
wa,i =   .
2 2
j kIa va,k j Ii vi,k

The denominator has the effect of normalizing the contribution of user i relative to
the total number of votes that user i has made. Thus, in an extreme example, if a has
voted on only two items, but there is a user i who has voted on all items with the same
value as user a on the two items in common, then the correlation weight as dened
earlier would have value 1, even though user a and user i may really have very little
in common. In contrast, the vector similarity function would have a weight wa,i that
is much closer to zero than to one, since the contribution of user i to as items would
be downweighted by the fact that i has so many votes in general.
In addition to the weighting scheme, there are a signicant number of other factors
that can be tweaked in terms of constructing an automated algorithm for predictions.
COMMERCE ON THE WEB 217

For example, for a user a with many votes, there may be a very large number of users i
that will match with a but that have relatively small weights. Including these matches
in the prediction equation may actually lead to worse predictions than simply leaving
them out, and thus, one can impose a threshold on wa,i such that only users with
similarity weights above the threshold are used in the prediction, or only select the
top k most similar users, and so forth. The paper by Sarwar et al. (2000) discusses a
variety of such extensions.

The curse of dimensionality

One general problem with similarity-based approaches is the sparseness of the data
due to the high-dimensionality m of the space of items. Recall that typically only a
tiny fraction of all useritem pairs are nonzero (Table 8.1). As an example of the effect
that this can have, consider a user a who has a positive vote for a single book about
mountain climbing in the Himalayas. There is a good chance that this person will be
interested in other mountain-climbing books. With enough data from other users, we
will eventually be able to infer a general preference by user a for mountain-climbing
books in general. However, it may take a very large training data set of users to pick
up these correlations by nearest-neighbor methods.
An alternative approach is to search for correlation structure among the items
directly and to incorporate this into the prediction model. In essence, this idea is a
model-based approach, where we try to model the structure of the item dependencies
directly in some manner (and which we discuss in the next section in more detail). One
such approach is dimensionality reduction via singular value decomposition (SVD)
of the vote matrix V (Sarwar et al. 2000). In this approach we approximate the set
of row-vectors of votes as a linear combination of r basis columns, or equivalently,
we approximate the matrix V with a rank-r approximation (see Appendix A). r is
typically chosen to be much smaller than m (e.g. r = 100 while m = 10 000). This
general approach is also sometimes referred to as latent semantic indexing (LSI) when
applied to text or recommender data (see Chapter 4).
The basis columnscan be estimated by standard numerical techniques for deriving
SVD representations from sparse matrix data (Berry 1992; Letsche and Berry 1997)
these computations can be carried out ofine. The general approach is to nd the set
of k columns such that the distance between the approximated versions of the rows
and the original rows is minimized in a least-squares sense. The basis columns can
be thought of as dening a set of meta-items, each of which represent a set of items
that are highly correlated with each other. In this way, we can perform the nearest-
neighbor types of calculations for collaborative ltering in the r-dimensional space,
and then project out the predictions to all m items using the relevant linear algebra
from the SVD representation.
This type of technique can be quite useful for alleviating the sparsity problem, since
it can exploit item correlations directly. For example, Billsus and Pazzani (1998) used
SVD to reduce item dimensionality before applying machine-learning algorithms
to vote data, and found that this led to signicantly higher predictive accuracies
218 AUTOMATED RECOMMENDER SYSTEMS

compared to simpler memory-based methods based on direct correlation of users

votes. An important practical issue, particularly for online recommender systems, is
updating the basis vectors for an SVD representation as new data arrive. Ideally one
would like to be able to do this calculation in an incremental manner given the current
representation rather than recomputing a new SVD from scratch incremental SVD
computation is still somewhat of a research issue (Brand 2002).

Computational issues and clustering

In the memory-based approach to collaborative ltering, given a new user we must
calculate the weights of potential neighbors in the training data V to make a prediction.
For very large values of n and with many users on the site at any given time (e.g. large
e-commerce websites) the time required to compute the weights for all users may be
much greater than that required for a real-time response. As a result, nearest-neighbor
approaches to collaborative ltering do not scale well to large n.
One approach to this problem is to cluster the historical data contained in V into
K clusters, where K is much smaller than n (Breese et al. 1998; Ungar and Foster
1998). For prediction, a user is mapped into the cluster that they are most likely to
belong to (an operation that scales as O(K) rather than O(n) per user) and properties
of the cluster (e.g. the centroid or the mean prediction vector) are then used to make
predictions. A potential weakness of clustering is that any individual user might really
belong to multiple different clusters simultaneously. For example, in book buying, a
user a might have interests in kayaking, Italian history, and blues music. While there
might well be clusters representing each of these topics individually, it is unlikely that
there would be a cluster that has all three topics within it. Thus, by forcing user a to
be represented by a single cluster model, we will inevitably lose information about
the heterogeneity of his or her interests. Breese et al. (1998) found that cluster-based
collaborative ltering was outperformed by both nearest-neighbor and probabilistic
model-based collaborative ltering, the second of which we discuss next.

8.3.3 Model-based collaborative ltering

General concepts
Model-based approaches to recommender systems make predictions by building
(ofine) an explicit model of the relationship among items, as inferred from histori-
cal voting patterns represented by V . This model is then used (online) in prediction
mode for making predictions for a new user a. This is quite different in principle
from the nearest-neighbor collaborative ltering approach discussed in the previous
section, which relies on memory-based computations to compute predictions in real
time, without ever explicitly constructing a global model.
Model-based approaches have an advantage in real-time computation requirements.
Typically, once a model is constructed, the time to make a prediction is independent
of n (the number of training examples used to construct the model) in contrast to
memory-based approaches that scale directly with n.
COMMERCE ON THE WEB 219

Model-based approaches to collaborative ltering typically fall into a few general

classes. The rst general approach constructs a full joint probability distribution over
the m items in I, and then uses this joint distribution to make predictions online. The
challenge here is to construct an accurate high-dimensional joint distribution from a
very sparse data set. The second general approach builds m conditional models, where
each model predicts the likelihood of an individual item given any combination of
observed votes for the remaining m 1, in contrast to a single joint model. Using
this multiple conditional model paradigm a variety of different predictive modeling
techniques have been investigated, including dependency networks (Heckerman et
al. 2000) and linear classiers (Zhang and Iyengar 2002). A third general approach
is to directly learn a model than can produce ranked lists, rather than models which
produce scores that are then ranked (see, for example, Cohen et al. 1999).
We next examine some proposed methods for model-based recommender systems,
specically the joint density model proposed by Hofmann and Puzicha (1999) and
the conditional density approach of Heckerman et al. (2000).

Joint density modeling

An unconstrained joint probability distribution for the full m-dimensional set of items
requires the specication (or estimation) of 2m independent parameters (for binary
votes), which is clearly impossible for the ranges of m one typically encounters in
recommender systems, e.g. m 103 and upwards. Thus, a practical strategy is to
seek more parsimonious representations for item dependencies.
One approach is to model the joint distribution as a nite mixture (a weighted
combination) of simpler distributions, e.g.

K
P (v1 , . . . , vm ) P (v1 , . . . , vm | c = k)P (c = k),
k=1

where the sum is over K mixture components, P (c = k) is the probability

 that a ran-
domly chosen row in the data set belongs to component k (where k P (c = k) = 1),
and P (v1 , . . . , vm | c = k) is a probability model for each component. If we think
of this model in a generative fashion (for simulating the data) we can imagine for
each user rst selecting one of the K models with a K-sided die having face prob-
abilities P (c = k), and then generating a set of votes for that user using the model
P (v1 , . . . , vm | c = k) given that the votes are coming from model k, 1
k
K
(see Chapter 1).
Of course to gain anything in terms of reducing the number of parameters needed
in the overall model, we need to make the component models much simpler than a
full unconstrained joint distribution on m items. A common choice for component
models in mixtures for high-dimensional data is to assume that the votes for items
are generated independently within a component

m 
m
v
P (v1 , . . . , vm | c = k) P (vj | c = k) = j kj (1 j k )1vj ,
j =1 j =1
220 AUTOMATED RECOMMENDER SYSTEMS

where vj 0, 1 indicates whether the vote in the j th column is zero or one and
j k = P (vj | c = k). This is known as a conditional independence or Naive Bayes
model in statistics and machine learning discussed earlier in Chapters 1 and 4. More
generally one could use various forms of Bayesian networks or log-linear models for
each component.
The number of parameters of the Naive Bayes model per component is linear in
dimensionality m, so overall we only have O(Km) parameters compared to O(2m )
for a full joint distribution (for binary votes). The parameters of the model can be esti-
mated from training data using a simple application of the expectation maximization
algorithm that iteratively adjusts the parameters of the hiddenmodel to maximize the
likelihood of the observed data (see, for example, the earlier discussion in Chapter 1
and more generally in Hand et al. (2001)).
The mixture of conditional independence models may be too simple to fully
capture many aspects of the real data. For example, it ignores any dependencies
between items within a cluster, i.e. all pairs P (vj , vl | c = k) are modelled as
P (vj | c = k)P (vl | c = k). Nonetheless, it does capture unconditional marginal
dependence of items, in that the model does allow P (vj , vl ) to be different from
P (vj )P (vl ). Intuitively, we can imagine that for each component there are sets of
items that are relatively likely to have on votes, where P (vj = 1 | c = k) is much
greater than P (vj = 1), and the rest of the items are more likely to have votes vj = 0.
A limitation of this type of model for collaborative ltering is that it implies that
each user can be described by a single component model we are assuming that each
user was generated by one (and only one) of the K components. This is the same
assumption made in the clustering approach discussed earlier, and indeed our mixture
model is essentially a form of clustering with probabilistic semantics. Thus, as before
with clustering, if a users interests span multiple different types of topics (e.g. books
on kayaking, Italian history, and blues music) there may be no single cluster that
represents the combination of these three topics. However, there may well be mixture
components that represent each of the individual topics, e.g. clusters of books for
each of water sports, European history, and music.
Hofmann and Puzicha (1999) proposed an interesting extension of the conditional
mixture model above that directly addresses the problem of multiple interests, based
on a more general model proposed earlier by Hoffman (1999). Their model can be
interpreted as allowing each users interests to be generated by a superposition of K
different underlying simpler component models. Thus, instead of assuming that each
users row vector of votes is generated by a single component model P (v1 , . . . , vm |
c = k), in the Hofmann and Puzicha model each row vector of votes can be generated
by a combination of up to K of the component models. This is a potentially powerful
idea in modeling high-dimensional data. To represent arbitrary combinations of K
different interests we do not need 2K different models but can instead combine K
models appropriately.
Parameters of the Hofmann and Puzicha model can be t to the observed data
matrix V using the EM algorithm. In the original formulation of the model each
user has his/her own set of parameters, leading to potential overtting when using
COMMERCE ON THE WEB 221

EM in a standard ML framework. Another limitation of the original model is that

it is not a complete probabilistic model in the usual sense: it only denes a distri-
bution over users in the training data V and does not directly provide a mechanism
for making predictions on a new user a who is not in the training data. More recent
work by Blei et al. (2002a,b) extends the framework to address both of these prob-
lems, demonstrating a systematic improvement in prediction power over conditional
independence mixture models, when the model is used as the basis for making rec-
ommendations on the widely used EachMovie ratings data set (available online at
www.research.compaq.com/SRC/eachmovie/).

Conditional Distribution Models

An alternative to modeling the full joint density P (v1 , . . . , vm ) is to build m different
conditional density models, P (vj | S \ vj ), 1
j
m, where S is the full set of m
random variables. The intuition behind this idea is that it may be more effective to
directly predict the probability of each item j conditioned on the other items, rather
than the less direct approach of modeling the full joint density and then using the
joint density to infer predictions on how likely individual items are given information
about other items.
One such approach proposed by Heckerman et al. (2000) is to use probabilistic
decision trees (Buntine 1992) for each of the m conditional models. Trees are widely
used in classication for high-dimensional problems, since they automatically per-
form variable selection and (in principle) only select the variables relevant to the
prediction task at hand. They can also be used to approximate a high-dimensional
conditional distribution for one variable vj conditioned on many others. The tree is
typically estimated from the data in a greedy fashion, by choosing a root node and
recursively growing a binary tree beneath this node.
In Heckerman et al. (2000) each internal node corresponds to the addition of a
single predictor binary variable vk , where one branch corresponds to a particular
value of vk and the other branch corresponds to all other values of vk . The node
(variable plus split) that is greedily added to the current tree is the one that leads to
the greatest increase in a Bayesian score. This score approximates the probability of
a particular tree structure T given the data, P (T | D) P (D | T )P (T ), where D
is the data, P (D | T ) is the likelihood of the data under the current tree model T ,
and P (T ) is a suitably dened prior distribution over tree structures. If there are no
variable-split combinations available that can be added to any node to increase the
score then tree-growing halts. Conditional probability vectors for vj are estimated at
the leaves using only variable-value pairs in the branches from the root to each leaf.
In this manner only variable-value pairs that are relevant to the prediction of vj are
included in the tree model, and thus, it is hoped that a relatively parsimonious tree
structure can be found from the data that provides a useful approximation to the true
conditional density.
The probability tree approach is relatively straightforward to apply to recommender
data sets. m different probability trees are constructed (as described above) to predict
222 AUTOMATED RECOMMENDER SYSTEMS

the probability of each of the m different items in the data conditioned on the other
items, using the vote matrix V . To make predictions for user a, if the votes are implicit
(e.g. product purchases or Web page requests) we can use all of the other m1 implicit
votes as inputs to predict the vote of interest. Predictions are made for each item that
has an implicit vote of zero, i.e. that has not been purchased or visited, and then the
resulting set of probabilities are ranked to provide a recommendation to the user a.
Heckerman et al. (2000) applied this technique to recommender data sets involving
website visits (a visit to a particular page by a user is considered an implicit binary
vote for that page, and items correspond to pages) and TV viewing records consisting
of data collected by companies that estimate TV viewership here the items corre-
spond to programs watched by viewers. The tree models were compared to a method
proposed earlier by Breese et al. (1998) that constructs a full Bayesian network as
a model for the joint distribution P (v1 , . . . , vm ) and then uses the network to make
predictions on each vj that does not have a positive vote, using all the others. Breese et
al. found that the Bayesian network model empirically outperformed memory-based
and cluster-based collaborative ltering on different ratings data sets, and thus, it
provided a useful comparative benchmark for the tree models.
The models were trained on training subsets and evaluated on disjoint out-of-sample
test subsets. In the test data sets, each users positive votes were randomly partitioned
into two subsets of items: an input set, where the votes were assumed known and
used as inputs to each model, and a measurement set, where the votes were assumed
unknown and used to test the models predictive power. Predictions were made for a
variety of scenarios using this train/test arrangement. Under one scenario all but one
of the positive votes for user a were placed in the input set of items and then used
to predict the other item (this corresponds to knowing as much as possible about the
user). Under the other scenarios only a xed number k 2, 5, 10 of positive votes
were put in the input set for prediction, so that smaller k values corresponded to less
information about the user.
A particular type of objective function was constructed to estimate the utility of
each ranked list of ratings to a user a full description of this function is provided
in Heckerman et al. (2000). For the Web and TV viewing data the probabilistic trees
were generally slightly less accurate than Bayesian networks in terms of this objective
function, but the differences were quite small.
Table 8.2 summarizes a variety of other aspects of the experiments, across the three
data sets, for the specic scenario of all but one prediction results for the other
scenarios were quite similar. The trees are signicantly faster in terms of making
predictions (e.g. 23.5 versus 3.9 per second on the second Web data set), which is an
important feature for real-time prediction on websites. Presumably both algorithms
were implemented efciently so that these numbers provide fair comparisons between
the two. The probabilistic trees also have advantages ofine: they require less time
and an order of magnitude less memory to train than the Bayesian network models.
For both techniques it is quite impressive that models for high-dimensional data can
be constructed so quickly, e.g. in about 100 s for the rst Web data set based on
1000 dimensions (items) and 10 000 users.
COMMERCE ON THE WEB 223

Table 8.2 Summary of data sets and experimental results for all but one prediction experi-
ments from Heckerman et al. (2000). BN is the belief network model and PT is the probability
tree model.

Web data 1 Web data 2 TV data

Users in training data 10 000 32 711 1 637

Users in test data 5 000 32 711 1 637
Number of items 1 001 294 203
Mean positive votes per row 2.7 3.0 8.6

Predictions per second (BN) 7.1 3.9 23.5

Predictions per second (PT) 11.8 23.5 37.4

Training time [s] (BN) 105.8 144.6 7.7

Training time [s] (PT) 98.9 98.3 6.5

Training memory [MB] (BN) 43.0 42.4 3.3

Training memory [MB] (PT) 3.7 5.3 2.1

In practical settings, e.g. for a real-world e-commerce application, both the number
of users and the number of items will often be signicantly larger than those in the data
sets described above (Schafer et al. 2001). Nonetheless, these experiments provide
useful guidelines on some of the tradeoffs and options available when using model-
based techniques for recommender systems.

8.3.4 Model-based combining of votes and content

Another variation on recommender systems is where we have additional content infor-
mation about each item to be recommended. For example, for documents we usually
know the terms that are in each document, and for movies we may have informa-
tion about the type of movie, the actors, the rating of the movie, the director, and so
forth. This type of content information can be used to estimate how similar certain
items are to each other. For instance, we could represent the content information as
a vector of values and use any of a variety of vector-based similarity measures. In
principle, this type of content-based similarity alone could be used as the basis for a
recommender system (Mooney and Roy 2000). For example, a user model can be
automatically constructed for each user based on the content descriptions of items
purchased or rated, and then these user models are matched in terms of similarity to
infer ratings and likely purchases for unseen items. Search engines can also be thought
of as recommender systems that are purely based on content, where Web pages are
recommended based on matches to a user query.
An advantage of content-based systems is that they can make recommendations
for entirely new items for which there is no history, such as a new book or movie that
no one has yet rated or purchased. Approaches based on historical patterns of ratings
and purchases, such as collaborative ltering, cannot generalize in this manner to new
224 NETWORKS AND RECOMMENDATIONS

items. On the other hand, systems that are only based on content are clearly ignoring
potentially valuable information that lurks in historical rating and purchasing patterns
of the community.
An example of a model-based approach that combines both content and collabora-
tive ltering is that proposed in Popescul et al. (2001). Their approach is an extension
of the hidden variable model of Hofmann and Puzicha (1999) discussed earlier in Sec-
tion 8.3.3, where now the features of the items are included in the model as well votes
from users. A specic model was developed for the case of users browsing documents
at an online digital library (the NEC ResearchIndex document database), where items
correspond to documents, content consists of the words in each document, and a
positive vote corresponds to user u requesting a particular document m. In this model,
a joint density is constructed by assuming the existence of a hidden latent variable z
that renders users u, documents m, and words w conditionally independent, i.e.

P (u, d, w) P (u | z)P (d | z)P (w | z)P (z). (8.1)
z

As in the Hofmann and Puzicha (1999) approach, the hidden variable z represents
different (hidden) topics for the documents, and multiple topics can be active within
a single document d or for a single user u. The inclusion of the term P (w | z) allows
the inclusion of content information in a natural manner. The EM algorithm can be
used to estimate the conditional probability parameters that relate the hidden variable
z to the observed data. Popescul et al. (2001) found that this particular model had
problems with overtting due to the sparsity of the data, even based on a relatively
active set of 1000 users accessing 5000 documents, where the density of ones in the
data matrix was 0.38% versus 0.01% for randomly selected users. To combat this
overtting, they also proposed a simpler model P (u, w) based on content alone that
in effect infers preferences in word-space, which is much more dense than document-
space since there are far more words, and found empirically that this model produced
better predictions than the original model. Thus, while relatively sophisticated models
can be proposed for combining content and votes, tting these models to sparse high-
dimensional data remains a signicant challenge.

8.4 Networks and Recommendations

8.4.1 Email-based product recommendations
While email-based recommending evokes images of being inundated with junk or
spam email messages, there are a number of situations where the concept works
rather well in favor of the consumer. We are all familiar with the concept of word-
of-mouth, where commercial products (such as movies) gain wide acceptance with
very little explicit advertising. Instead, users of the product recommend the product
to their circle of friends, who in turn recommend it to other friends, and so forth. For
example, the Google search engine became (in 2002) the most widely used search
COMMERCE ON THE WEB 225

engine on the Web, primarily based on word-of-mouth recommendations and with

very little direct advertising. From a marketing viewpoint this is an ideal way to gain
market share without having to pay for expensive advertising campaigns.
Perhaps the most well-known and famous example of marketing by email is the
case of Hotmail. The general idea was to provide free email service to users, with
revenue for Hotmail being generated by selling various forms of advertising and
additional fee-based services. The Hotmail free email service was introduced in July
1996 and had 20 000 subscribers by the end of the month. By September 1996 it
had 100 000 subscribers, by January 1997 it had over one million subscribers, and
18 months after starting it had 12 million subscribers. By April 2002 the number of
Hotmail subscribers (now part of Microsoft) was reported to be 110 million. What is
particularly impressive about this rate of growth is the fact that Hotmail spent very
little on conventional advertising during the initial growth period. Instead, they gained
customers by embedding a link to the Hotmail sign-up page at the bottom of every
email being sent out by their current subscribers. In a network-based system such as
the Web this has a tremendous spreading activation effect. In effect, compared to
our discussion earlier in this chapter on collaborative ltering, the emails being sent
by each individual can be thought of as an implicit recommendation of the Hotmail
service, e.g. if my colleague is using this service perhaps I should look into it.
Intuitively, we can imagine the Web as a very large dynamic network, with Hot-
mail emails emanating from an initially tiny fraction of nodes in the graph, e.g. the
initial 20 000 Hotmail subscribers in July 1996. As Hotmail emails are sent from one
individual to other individuals, each receiver will see the advertisement at the bottom
of the email and some fraction (which need not be very large) will sign up for the
service. At the next time-step (say the next day), there is now a larger set of Hot-
mail subscribers sending out emails, and thus, an even greater number of individuals
receiving Hotmail emails. It is easy to see under fairly reasonable assumptions that
large portions of the network will soon have received emails from someone who is a
Hotmail subscriber, and some fraction of this large number will have signed up for
the service.
This is precisely the type of spreading activation mechanism that appears to have
occurred with Hotmail. Even if only a very small fraction of individuals receiving
Hotmail emails decided to sign up for the service (say 0.1% or less), because of the
effective connectivity of email-based networks, and the number of individuals who
send and receive email on a daily basis, there is an excellent chance that the product
recommendations will spread like wildre.
The success of network-based recommendations relies of course on the assumption
that the underlying product or service is something that is of a certain quality and of
general benet to consumers. For example, we can contrast the implicit recommen-
dations from Hotmail subscribers with the now prevalent and unfortunate practice of
spam email, where emails that advertise products and services of dubious quality
are broadcast to millions of email addresses. It is likely that most of these emails are
not even read by the recipients and are almost never forwarded on to other colleagues.
226 NETWORKS AND RECOMMENDATIONS

Thus, there is clearly no network effect at play here but rather a brute force attempt
to use the Web for direct advertising.

8.4.2 A diffusion model

Montgomery (2001) provides an interesting discussion of the Hotmail effect using
diffusion models from the marketing literature our discussion closely follows the
treatment in this paper. These types of models were used by Bass (1969) to model
the word of mouth network effect in sales of consumer products such as televisions,
refrigerators, and so forth. The model predicts how many individuals k(t) at time t
will adopt the product from a total population of N individuals.
The model is based on two simple assumptions.
(1) At time t, N k(t) individuals have not adopted the product. It is assumed that
a constant proportion  0 of these individuals will adopt the product at the
next instant in time because of advertising. This part of the model represents
the effect of direct advertising.
(2) The second part of the model represents the network effect. At time t there are
k(t)(N k(t)) potential links between pairs of individuals who have already
adopted the product and those that have not. (For Hotmail this could be thought
of as potential emails between pairs of individual email accounts.) It is assumed
that the rate at which new individuals adopt the product is proportional to this
number, with weighting coefcient  0.
From these two assumptions, the rate of change of k(t) can be expressed as
k(t)
= (N k(t)) + k(t)(N k(t)), (8.2)
t
which has a solution in the form
 
1 e(+N)t
k(t) = N (8.3)
1 + (N/)e(+N)t
that converges to N as t .
Montgomery (2001) t this model to the weekly Hotmail subscriber numbers for the
rst year of its operation (using data from Roberts and Mahesh (1999)), by estimating
the parameters , , and N . The resulting estimates were = 0.0012, = 0.008,
and N = 9.67 (in millions of individuals) with t measured in weeks. It is noteworthy
that the value for N (which measures the network effect) is much larger than the
coefcient which measures the direct advertising effect.
This model of course has its limitations. It ignores effects such as individuals who
stop using their Hotmail accounts (perhaps after initially trying it out), and the fact
that the number of individuals using email in general is not constant but is growing as
a function of time t. Nonetheless, the model provides quite a good t to the observed
weekly data (see Figure 8.1) as reported by Montgomery (2001), although it should be
COMMERCE ON THE WEB 227
5

Bass Model: = 0.0012, = 0.008, N = 9.76 million

4.5

3.5
Millions of Subscribers

2.5

1.5

0.5

0
0 10 20 30 40 50 60
Weeks

Figure 8.1 Fitted Bass distribution model for the rst 52 weeks of adoption of Hotmail, where
the curve represents the tted model and the dots represent ve of the 52 data points used to
estimate the parameters of the model. (Adapted from Montgomery (2001).)

noted that these particular data could also be well-approximated by a simple quadratic
or exponential for this initial growth period.
If we extrapolate the model to a later time-period we see that the diffusion model has
a characteristic S-shaped curve (Figure 8.2), that asymptotes at N, the estimated
ultimate number of adopters. In Figure 8.2 we have added the number of Hotmail
subscribers after 72 weeks (12 million) and after six years (110 million). We can see
that the coefcients estimated using only the rst 52 weeks of data do not extrapolate
well. It is perhaps not surprising that a model tted to the rst years worth of data do
not extrapolate well to predicting what will happen ve years later. Figure 8.2 also
shows a different set of hand-chosen parameters for the diffusion model, where the
asymptotic value of N = 110 is assumed known and the values of and are each
reduced, resulting in a much better overall t of course in hindsight it is always
somewhat easy to explain the data!
The diffusion model as presented above is certainly too simple to fully explain the
success of Hotmail or other similar Web phenomena. Such models nonetheless serve
the useful purpose of providing a starting point for generative modeling of network-
based recommendations on the Web. In related work on this topic, Domingos and
Richardson (2001) proposed a Markov random eld model based on social networks
for estimating an individuals network value, which depends on how much a particular
individual can inuence other individuals in the network. In a broader context, Daley
228 WEB PATH ANALYSIS FOR PURCHASE PREDICTION
120

100

Bass Model:
80
Millions of Subscribers

= 0.0002, = 0.0004, N = 110 million

60

40

20 Bass Model:
= 0.0012, = 0.008, N = 9.76 million

0
0 50 100 150 200 250 300 350
Weeks

Figure 8.2 Bass distribution models over six years of Hotmail adoption.

and Gani (1999) provides a comprehensive overview of a variety of deterministic

and stochastic models that have been investigated for characterizing the spread of
epidemics and rumors in a population.

8.5 Web Path Analysis for Purchase Prediction

Analysis of individual navigation paths through e-commerce websites can provide
valuable insights into customer behavior and provide clues about whether improve-
ments in site design might be useful. From a marketing viewpoint it is often desirable
to be able to predict early in a session whether an individual is likely to make a pur-
chase or not. In this section we look at how an individuals browsing behavior can be
related to purchase probability, using illustrative data from a study by Li et al. (2002).
When consumers search for products their behavior can be generally categorized
into (1) a goal-directed search, or (2) an exploratory search (Janiszewski 1998). In a
goal-directed search a consumer is typically looking for information about a specic
product and is directly interested in purchasing that product. In exploratory search
a consumer is just browsing and may purchase any of a variety of products if he
or she decides to do so on the spur of the moment, or may choose not to make
any purchase at all. For consumers browsing an e-commerce website, this simple
categorization seems quite reasonable. We might expect the goal-directed consumers
COMMERCE ON THE WEB 229

Table 8.3 Estimated transition probabilities among eight states from sessions where items
were purchased. The value for row i, column j indicates the probability of going from a page
in category i to a page in category j . The marginal probabilities (the probability of a page being
in a particular category, as estimated from the data) are in the rightmost column.

H A L P I S O E Marginal

H 0.21 0.03 0.36 0.00 0.30 0.07 0.00 0.03 0.01

A 0.01 0.67 0.01 0.00 0.04 0.25 0.01 0.01 0.13
L 0.01 0.00 0.63 0.21 0.08 0.06 0.00 0.01 0.27
P 0.01 0.01 0.32 0.44 0.15 0.06 0.00 0.01 0.16
I 0.02 0.05 0.11 0.12 0.56 0.06 0.03 0.05 0.23
S 0.01 0.19 0.05 0.02 0.13 0.45 0.14 0.01 0.15
O 0.00 0.01 0.00 0.00 0.71 0.12 0.00 0.17 0.03
E 0.50 0.00 0.00 0.00 0.50 0.00 0.00 0.00 0.02

to have different navigation patterns through a site than the consumers who are just
browsing in an exploratory manner.
We next examine this hypothesis further using a data set collected by Li et al.
(2002), consisting of the browsing patterns of 1160 different individuals who visited
the online bookstore www.barnesandnoble.com between 1 April and 30 April
2002. The data were collected at the client side using the Comscore Media Metrix
system that records the page requests and page-viewings of a randomly selected set
of computer users. While relatively small in terms of the total number of individuals,
this data set is nonetheless quite useful in terms of providing a general idea of the
characteristics of online shopping behavior.
The 1160 individuals generated 1659 different sessions at the Barnes & Noble
website. The sessions in total consisted of 9180 page requests and 14 512 page view-
ings, where page viewings count all pages viewed by the user and page requests
only record requests that went to the server website (thus, pages that were viewed
twice in a session and redisplayed by the caching software would be counted as page
viewings, but not as page requests). The end of a session was declared if no page had
been viewed for 20 min.
The mean session length (in terms of page viewings) was 8.75, the median was
ve, the standard deviation was 16.4, the minimum length was one, and the maximum
was 570. 7% of sessions ended in purchases, a rate that is relatively high purchase
rates of between 1% and 2% have been reported in the media as being more typical
for e-commerce sites in general (Tedeschi 2000).
Demographic data were also available for the 1160 individuals. They had an average
age of 46, 53% were female, 77% were white, 40% had children, 29% were married,
82% had some college education, and 32% had an annual income in excess of $50 000.
This suggests a relatively well-educated and afuent set of consumers.
One of the difculties in analyzing Web navigation data is the very large number
of possible pages that can be presented to a site visitor. In essence the number is
unbounded, since in addition to there being a page for each product, the pages them-
230 WEB PATH ANALYSIS FOR PURCHASE PREDICTION

selves can change dynamically (in terms of presentation, information, promotions,

etc.) for different visitors and even for the same visitor at different times. To address
this, Li et al. (2002) classied each page into one of seven different categories:
home (H), the main home page and entry point for most visitors;
account (A), containing account information for individuals;
list (L), for pages with lists of items in certain categories or results of searches;
product (P), with information on a specic product;
information (I), with general information about shipping, order status, etc.;
shopping cart (S), with current shopping cart information; and
order (O), indicating conrmation of a purchase.
In addition an entry/exit category (E) was added to indicate the page prior to the rst
page on the site or the page after the last page visited on the site.
An example of an actual session where a purchase did not occur is IHHIIILIIE
(taken from Li et al. (2002)). We see that the user visited several information pages, but
did not visit any specic product page. An example of a session where a purchase did
occur is IPPPPSASSSSSOIAAAHLLPLLLLLLE, where the user looks at specic
product pages (P), visits the shopping cart (S), goes to the order conrmation page
(O), then the account page (A), before nishing the session with some viewing of
lists of information on category pages (L). For these two examples, the session with
a purchase is much longer than that with no purchase. In fact this is true in general
for this data set the average length of sessions with purchases is 34.5 page views,
versus only 6.8 page views for sessions with no purchases. While this is interesting
retrospectively, it is not necessarily useful from a predictive viewpoint, since ideally
we would like to be able to predict whether a visitor will make a purchase or not
after only the rst few page requests. One has to be careful here not to infer that
there is necessarily a causal connection between the number of pages viewed and the
probability of making a purchase. It might well be that any causal relationship (if it
exists) is in the other direction, i.e. if a visitor is inclined to make a purchase then this
might cause the visitor to view more pages than usual during this session.
From the data Li et al. (2002) estimated sample Markov transition matrices for
the eight categories, both for sessions where purchases occurred, indicated by the
presence of the state O for order conrmation in a session sequence, and for sessions
where no purchases occurred. The estimated matrices are shown in Tables 8.3 and 8.4.
We note for example that the probability of visiting the home page (per pageview)
is eight times higher at P (H | no purchase) = 0.08 for non-purchasers than it is for
purchasers (P (H | purchase) = 0.01) indicating that non-purchasers spend much
more time looking at general aspects of the website. We also see that the probability
of exiting a session is also much higher for the non-purchasers than for the purchasers,
P (E) = 0.14 versus P (E) = 0.02, respectively, which translates into much longer
sessions on average for the purchasers, as noted earlier. Transition probabilities into
COMMERCE ON THE WEB 231

Table 8.4 Estimated transition probabilities among eight states from sessions where items
were not purchased. The value for row i, column j indicates the probability of going from a
page in category i to a page in category j . The marginal probabilities (the probability of a page
being in a particular category, as estimated from the data) are in the rightmost column.

H A L P I S O E Marginal

H 0.36 0.01 0.16 0.01 0.25 0.01 0.00 0.19 0.08

A 0.03 0.68 0.04 0.01 0.08 0.06 0.00 0.11 0.04
L 0.02 0.01 0.59 0.14 0.12 0.01 0.00 0.11 0.30
P 0.03 0.01 0.25 0.39 0.19 0.01 0.00 0.12 0.14
I 0.08 0.01 0.14 0.08 0.53 0.02 0.00 0.15 0.28
S 0.06 0.13 0.03 0.06 0.33 0.32 0.00 0.07 0.02
O 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
E 0.25 0.06 0.19 0.00 0.50 0.00 0.00 0.00 0.14

the Shopping Cart (S) and Account (A) states are much higher in the purchase group
than in the non-purchase group, and of course the transition probabilities into the
Order (O) state are all zero for the non-purchase group.
Li et al. tted a series of relatively sophisticated statistical models to this data,
the details of which are somewhat beyond the scope of this text. The most complex
models included several components:
(1) latent variables that modelled the utility of individual i selecting a page from
category c at time t;
(2) time-dependence via a hidden Markov model that allows switching between
states over time;
(3) covariates based on both properties of the pages and individual demographics;
and
(4) heterogeneity across different individuals using a hierarchical Bayes approach,
as briey discussed in Chapter 7.
The models were tted using a widely used stochastic sampling technique for
Bayesian estimation known as the Monte Carlo Markov chain. The different models
were then evaluated in terms of their ability to provide out-of-sample predictions of
both
(a) the next page category requested by an individual given the sequence of pages
requested by that individual up to that time-point and given the individuals
demographic data; and
(b) whether or not that person would make a purchase or not (given the same
information).
In terms of predicting the next category, the results indicated that Markov models
with memory were signicantly more accurate than models that did not have memory,
232 EXERCISES

with the best memory-based models achieving 64% accuracy in predicting the next
category (out of eight) versus only about 20% for non-memory models. It was also
found that hidden Markov models with two states consistently had slightly better
accuracy out-of-sample than models with only a single state, although it should be
cautioned that this was based on only a single train/test split of the available data.
The authors interpreted one of the states as being an exploratory browsing state and
the other as being purchase-oriented, lending some support to the hypothesis earlier
about a simple two-state model for consumer behavior in an e-commerce context.
Demographics did not appear to have signicant predictive power in the models, and
browsing behavior seemed to be relatively independent of the duration of page views
according to the model. Interestingly, the composition of the page appeared to have an
inuence on browsing behavior. Having prices on a page seemed to discourage visitors
in the browsing-oriented state from continuing to browse, while the same information
had a positive effect on visitors in the purchase-oriented state. Conversely, promotional
items such as advertised discounts had the reverse effect, encouraging visitors in
the browsing state to continue browsing, but discouraging users in the purchasing
state from continuing in that state. Naturally, we need to keep in mind that all such
interpretations of the data are being ltered through a particular model and as such may
reect artifacts of the model-tting process and the random sample at hand rather than
necessarily reecting a true property of the processes generating the data. Nonetheless,
these types of inferences are quite suggestive and lead naturally to hypotheses about
consumer behavior that could be properly tested in further experimental studies.
In terms of predicting whether a visitor makes a purchase or not, the models tted
by Li et al. were somewhat less accurate than they were in predicting the next category.
Recall that 7% of all sessions in this data set result in a purchase. The two-state model
was used to predict whether a visitor would make a purchase or not, based on the rst
k page views, for site visitors whose sessions were not used in building the model
(but whose demographic data were assumed known). After two page views, the model
predicted that 12% of the true purchasers would purchase a product and only 5.3%
of the true non-purchasers. After six page views, the model predicted that 13.1% of
the true purchasers would purchase a product and made the same prediction for 2.9%
of the true non-purchasers. Note that a perfect forecasting system would predict a
purchase for 100% of the purchasers and 0% of the non-purchasers. More-accurate
forecasts can be made as we receive more information six page views are more
accurate than two page views but overall there is signicant room for improvement,
since some 87% of true purchasers went undetected by the model after six page
views.

8.6 Exercises
Exercise 8.1. Consider a simple toy model that we could use to simulate a large
sparse binary data set with n rows (users) and m columns (items). The model operates
in the following way.
COMMERCE ON THE WEB 233

(1) Each user acts independently of all other users, using the same model (below).
(2) For a given user i, he or she considers each of the m items in turn, and indepen-
dently makes a decision with probability p as to whether to purchase the item
or not. If they decide to purchase the item then they enter a 1 for that item,
otherwise a 0.
(3) We set p = k/d, where k is a parameter of the model and represents the
expected number of items purchased by each user. Typically k is much smaller
than m, e.g. k = 4, m = 1000.
Calculate the probability, as a function of k, m, and n, that a new user acting under
the same model will have no items at all in common with any of the n users in an
n m data matrix generated under the model above. Plot this probability for different
values of k, m, and n, and comment on the results.
Exercise 8.2. An extension of the model in Exercise 8.1 allows each of the items to
have a different probability pj , 1
j
m, of being purchased. A further extension
allows limited dependencies among items, e.g. the probability of a certain item j being
purchased depends on whether another item i was purchased or not, represented by
conditional probabilities P (j | i). Clearly the most complex model we could consider
in terms of item dependencies would be a full joint density on all m items.
How many parameters are needed to specify this full joint density? How many
are required for the independence model in Exercise 8.1? Clearly the independence
model is too simple, but the full joint density is impractical once m becomes large.
Describe as many types of probabilistic models you can think of that are in between
the independence model and the full joint density model (e.g. a Markov model where
each item depends on other item, in a specied order, would be one such model).
Exercise 8.3. A different probabilistic model than the independence model in Exer-
cise 8.1 for generating sparse binary data operates as follows. Each user still generates
data independently. To generate a row of data, the ith user now tosses an m-sided die
some number of times. The number of tosses could also be drawn from a random dis-
tribution, or could be xed for each user. The probability of the m-sided die coming
up on side j is represented by pj , 1
j
m, where pj = 1, and for simplicity
we could assume that all pj are equal, i.e. pj = 1/m. If the die comes up on side j ,
the user puts a 1 in column j , otherwise a 0. This type of data generation model is
sometimes referred to as a multinomial model as discussed in Chapter 4.
Compare this model with the independence model in Exercise 8.1, in terms of the
type of data that will be generated. For example, assume that in our model here that
the die is tossed exactly k times for each user. How does the probability distribution
of number of ones per row (we can think of this as the distribution of basket sizes)
under the die model compare to the probability distribution of basket sizes under
the independence model in Exercise 8.1? (Hint: reviewing the material on Markov
chains in Chapter 1 may help.)
You can also consider the problem of trying to model two items that have an
exclusive-OR relationship: the user is likely to purchase one or the other, but not both
Appendix A

Mathematical Complements

In this appendix, we rst begin with a short and rather informal introduction to some
of the basic concepts and denitions of graph theory used primarily in Chapter 3 but
also in the denition of Bayesian networks in Chapter 1. We then review a number
of useful distributions that are used in various sections of this book, and provide
a short introduction to Singular Value Decomposition (SVD), Markov chains, and
the basic concepts of information theory. The list of distributions that are routinely
encountered in probabilistic models and analysis of real world phenomena is not
meant to be exhaustive in any way. In fact, a more detailed list can be found, for
instance, in Gelman et al. (1995).

A.1 Graph Theory

A.1.1 Basic denitions
A graph G = (V , E) represents a binary relationship and is dened by a set of vertices
(or nodes) V and a set of edges E consisting of unordered pairs (v, w) of vertices.
Depending on the problem, loops of the form (v, v) may or may not be allowed.
Likewise, multiple edges between two nodes may or may not be allowed in different
situations. With |V | = n vertices numbered from 1 to n, the graph can be represented
by an n n 0-1 adjacency (or incidence) matrix A, where aij = 1 if and only if there
is an edge between vertices i and j . If multiple edges between vertices are allowed,
this can be reected in the matrix A by allowing nonzero integer entries reecting the
multiplicity of the corresponding edge. The degree d(v) of a vertex is the number of
edges adjacent to that vertex. In a directed graph G = (V , E) the edges are directed, or
equivalently the corresponding pairs are ordered. The adjacency matrix can be dened
in a similar way and, unlike what happens in the directed case, in the directed case the
matrix is usually not symmetric. In a directed graph we can distinguish the indegree
din (v) and outdegree dout (v) of a vertex, by considering the number incoming and
outgoing directed edges, respectively.

Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
236 APPENDIX A

A.1.2 Connectivity
A path of length k in a graph is an alternating sequence of edges and vertices
v0 , e1 , v1 , . . . , ek , vk with each edge adjacent to the vertices immediately preced-
ing and following it. A directed path is dened in the obvious way by adding the
requirement that all edges be oriented in the same direction. The shortest directed
or undirected path between any two vertices in a graph can be found using standard
dynamic programming/breadth rst methods (Cormen et al. 2001; Dijkstra 1959;
Viterbi 1967). The distance between two vertices can be measured by the length of
the shortest path. A loop is a path where v0 = vk and a cycle is a directed loop. A
directed acyclic graph, or DAG, is a directed graph which contains no directed cycles.
A (strongly) connected component of a (directed) graph is a maximal set of vertices
such that there is a (directed) path joining any two nodes. A graph is connected if it has
only one connected component. The diameter of a component is the maximal length
of the shortest path between any vertices in the component. The diameter of a graph is
the largest diameter of its connected components. A clique of G is a fully connected
subgraph of G (in some denitions, clique are also required to be maximal complete
subgraphs). The mathematical denition of diameter corresponds to a worst-case
denition. For practical purposes, it is often more informative to know what is the
average distance between any two vertices rather than the maximal distance. This is
also called average diameter or average distance in the literature and throughout the
book. Finally, it is easy to see how these connectivity notions can be modied and
adapted as needed in the context of directed graphs.

A.1.3 Random graphs

In Chapter 3, we consider graphs that can be associated with the Internet and the
Web, in particular the Web graph where the vertices are Web pages and the (directed)
edges are hyperlinks, as well as several classes of other related graphs. To the extent
that these graphs and their evolution have a stochastic component, it is useful to
compare and contrast them with various models of random graphs. Several models of
random graphs exist but, as a basic starting point, the reader should be familiar with
two classical models the uniform random graph models. The rst model G(n, e)
(Bollobs 1985; Erdos and Rnyi 1959, 1960) is associated with a uniform distribution
over all graphs with n vertices and e edges. In the second model G(n, p) (Bollobs
1985; Gilbert 1959), each possible edge occurs with a probability p, independently
of all other edges. It is well known that for large n and e = pn(n 1)/2 that
these two models have very similar properties. In particular, the vertex degrees have
approximately a Normal or Poisson distribution with mean 2e/n. The classical theory
of random graphs deals, in particular, with the asymptotic properties of such classes
of random graphs, for instance, for large n as p is varied. For many properties, it has
been observed and proved that there exist threshold functions (or phase transitions),
i.e. values of p = p(n) around which the asymptotic properties vary abruptly. For
instance, it is well known that the connectivity of random uniform graphs undergoes a
APPENDIX A 237

phase transition around p = log n/n, so that the graph tends to be disconnected below
the threshold value and has a giant connected component above the threshold (see,
for instance, Bollobs 1985). For reasons described in Chapter 3, the random uniform
graphs sometimes are also called random exponential graphs, due to the exponentially
decaying tail of their degree distribution.

A.2 Distributions
A.2.1 Expectation, variance, and covariance
For completeness, we provide a brief reminder on the denition of expectation, vari-
ance, and covariance. Let P (X) be a probability density function for a real-valued
variable X. The expected value of X is dened as

E[X] = xP (x) dx

and the variance is dened as


var[X] = (x E[X])2 P (x) dx = E[(X E[X])2 ] = E[X2 ] (E[X])2 .

The covariance of two random variables Xi and Xj is dened by

cov[Xi , Xj ] = E[(Xi E[Xi ])(Xj E[Xj ])].

It should be clear how these expressions can also be applied to discrete random
variables, or to vector random variables, on a component-by-component basis.

A.2.2 Discrete distributions

Binomial and multinomial distributions
The binomial distribution counts the number of heads or successes in a sequence of
n independent and identically distributed (i.i.d.) coin ips (Bernouilli trials). If p is
the probability of a single success, the probability of having k successes is given by
 
n k
P (X = k | p, n) = p (1 p)nk . (A.1)
k
The mean is E[X] = np and the variance var[X] = np(1 p).
The multinomial distribution is the generalization of the binomial distribution,
when each trial has m different possible outcomes. Thus instead of a coin, we have
a die with m faces, with probabilities p1 , . . . , pm . In this case, the probability
 of
observing each one of the m categories k1 , . . . , km times in n trials, with i ki = n,
238 APPENDIX A

is given by
n!
P (X1 = k1 , . . . , Xm = km ) = pk1 . . . pm
km
. (A.2)
k1 ! . . . km ! 1
The mean of each component is E[Xi ] = npi , the variance var[Xi ] = npi (1 pi ),
and the covariance cov[Xi , Xj ] = pi pj for i  = j .

Poisson distribution
The Poisson distribution corresponds to rare events. The Poisson distribution with
parameter is given by
k
P (X = k | ) = e . (A.3)
k!
The mean and the variance are equal to . When p is small, the binomial distribution
can be approximated by the Poisson distribution.

Geometric distribution
The geometric density with parameter p (0 < p < 1) is described by

P (X = k) = p(1 p)k , (A.4)

for k = 0, 1, 2, . . . . In the framework of Bernouilli trials, this is the probability that

heads appears for the rst time on the kth trial. The geometric density has mean
E[X] = 1/p and variance var[X] = (1 p)/p 2 . It is a special case of the negative
binomial density.

A.2.3 Continuous distributions

Gaussian distribution
The Gaussian or normal density is probably the most important continuous density.
A one-dimensional Gaussian density has the form
1
e(x)
2 /2 2
N (x | , ) = , (A.5)
2
where the parameters and are the mean and the standard deviation. The pre-
dominance and ubiquity of the Gaussian density has multiple origins. For instance, it
can be shown that the sum of a large number of i.i.d. random variables approaches a
Gaussian random variable. This is the well known Central Limit Theorem. The bino-
mial distribution can be approximated by a Gaussian distribution when n is large. The
Gaussian density is also its own conjugate in models parameterized only by the mean.
Furthermore, when the only information available about a continuous density is its
mean and its variance 2 , then the Gaussian density N (, ) is the one achieving
maximal entropy (Cover and Thomas 1991). Finally, if log X is normally distributed,
APPENDIX A 239

then X is said to have a log-normal distribution. In particular if log X is normally

distributed with mean and variance 2 , then the density of X is
1 1 (log x)2 /(2 2 )
f (x | , 2 ) = e (A.6)
2 x
2 /2)
for x > 0. The mean satises E[X] = e+( and the variance
2(+ 2 ) 2
var[X] = e e2+ .
Because the sum of two independent normal distributions is normal, the product of
two independent log-normal distributions is log-normal.

Exponential distribution
The exponential density with parameter is given by
f (x | ) = ex (A.7)
for x  0. The mean is E[X] = 1/ and the variance var[X] = 1/2 . The exponential
distribution is a special case of Gamma distribution.

Gamma distribution
The gamma density (Feller 1971) with parameters and is given by
1 x
(x | , ) = x e (A.8)
()


for x > 0, and zero otherwise. () is the Gamma function () = 0 ex x 1 dx.
The mean is E[X] = / and the variance var[X] = /2 . If X has a Gaussian
distribution with mean zero and variance 2 , then X 2 has a Gamma density with
parameters = 1/2 and = 1/2 2 . The exponential density with parameter is the
Gamma density with parameters = 1and . The Chi-square density with k degrees
of freedom, for instance, is the Gamma density with parameters = k/2 and = 1/2.

Dirichlet distribution
Finally, in the context of multinomial distributions that play an important role in this
book, an important class of distributions are the Dirichlet distributions (Berger 1985).
By denition, a Dirichlet density on the probability vector p = (p1 , . . . , pm ), with
parameters and q = (q1 , . . . , qm ), has the form
q 1
()  qi 1  pi i
m m
Dq (p) =  pi = , (A.9)
i (qi ) i=1 i=1
Z(i)
 
with , pi , qi  0 and pi = qi = 1. For such a Dirichlet distribution, E(pi ) =
qi , var[pi ] = qi (1 qi )/( + 1), and cov[pi pj ] = qi qj /( + 1). Thus q is the
mean of the distribution, and determines how peaked the distribution is around its
240 APPENDIX A

mean. Alternatively, the distribution can be parameterized using = (1 , . . . , m )

with i = qi . One reason Dirichlet distributions are important is that they are
the natural conjugate prior for multinomial distributions, as seen in Chapter 1. In
the two-dimensional case (m = 2), the Dirichlet distribution is also called the Beta
distribution.

A.2.4 Weibull distribution

The Weibull distribution is used in Chapter 6 in connection with the life span of
Web pages. More generally, it is used to model reliability and the distribution of life
spans in a wide range of applications ranging from mechanical devices to biological
organisms. The Weibull density function is dened by
 
x 1 (x/)
f (x | , ) = e , (A.10)

where > 0 is a scale parameter, > 0 is a shape parameter, and x > 0. The mean
(also called MTBF, mean time between failures) is given by
       
1 2 1 2
E[X] = 1 + , var[X] = 1 + 1+ .

The cumulative distribution function is simply

F (x | , ) = 1 e(x/) .

(A.11)

A.2.5 Exponential family

The exponential family (Brown 1986) is the most important family of probability
distributions. It has a wide range of applications and unique computational properties:
many fast algorithms for data analysis have some version of the exponential family at
their core. Many general theorems in statistics can be proved for this particular family
of parameterized distributions. The density in an exponential family with parameter
must have the form
 m 
f (x | ) = h(x)c( ) exp wi ( )ti (x) , (A.12)
i=1

where h(x)  0, ti (x) are real-valued functions of the observation x that do not
depend on , and c( ) and wi ( ) are real-valued functions of the parameter vector
that do not depend on x.
Most common distributions belong to the exponential family, including the nor-
mal (with either mean or variance xed), gamma (e.g. Chi square and exponential),
Dirichlet (e.g. Beta), in the continuous case, and binomial and multinomial, geomet-
ric, negative binomial, and Poisson distributions in the discrete case. A characteristic
APPENDIX A 241

of all exponential distributions is the exponential decay to zero for large values of x.
This is in contrast with the polynomial decay of power-law distributions studied in
Chapter 1 and often encountered in the other chapters of this book. Among the impor-
tant general properties of the exponential family is the fact that a random sample
from a distribution in the one-parameter exponential family always has a sufcient
statistic S. Furthermore, the sufcient statistic itself has a distribution that belongs to
the exponential family.

A.2.6 Extreme value distribution

Finally, the extreme value distribution is an important distribution when consider-
ing extreme effects and the distribution of the maximum
xof a sample of n points
sampled from a density g(x). If G(x) = P (X
x) = g(v) dv is the distri-
bution function, then the distribution of the maximum in the sample is given by
F (x) = P (max
x) = G(x)n with corresponding density f (x) = nGn1 (x)g(x).
The EVD (extreme valued density) is the limit of f (x) when n . The EVD
distribution has many applications to the study of many extreme events ranging from
catastrophes, to assessing the signicance of a maximum score in data mining and
information retrieval.
In the case where g is the exponential density with g(x) = ex straightforward
integration gives G(x) = 1 ex . If we dene z = x y with n = ey a simple
calculation shows that
nG(x)n1 g(x) ez exp(ez ) (A.13)
when n . This yields the distribution P (max
x) = exp(ez ), which is
called the Gumbel distribution (Galambos 1987). A remarkable fact is that the Gumbel
distribution is the EVD for many other densities besides the exponential density,
including the case where g is Gaussian.

A.3 Singular Value Decomposition

It is a standard fact in linear algebra that any m n real matrix X can be written as
X = U SV T , (A.14)
where
S is a diagonal m n matrix having the form




if m  n,
S= 0

! "

0 if m < n,

being = diag(1 , . . . , min(m,n) ), and i are called the singular values of X;

242 APPENDIX A

U is an m m matrix such that U T U = Im whose columns are called the left

singular values of X;

V is an n n matrix such that V T V = In whose columns are called the right

singular values of X.

Without loss of generality we may assume 1  2   min(m,n) . If

r = rank(X) < min(m, n),

then only the rst r singular values are nonzero.

SVD is often studied in conjunction with multivariate regression problems y =
f (x), y R, x Rn . In this case, assuming a linear model for f : y = f (x) =
x T w, the vector of parameters w can be estimated from a data set of m observations
{(x(i), y(i)), i = 1, . . . , m} using the least squares method

min 21 Xw y2 , (A.15)

w

where y = [y(1), . . . , y(m)]T and X = [x(1), . . . , x(m)]T .

The solution of (A.15) is

w = (XT X)1 X T y, (A.16)

which can be rewritten as

w = V S 1 U T y (A.17)

after plugging in Equation (A.14).

Another interesting interpretation of the SVD is in terms of principal component
analysis (PCA) or Karhunen Love transform. If X represents a data set as in the least
squares problem, then XT X/n is the sample covariance matrix and it is proportional
to
X T X = V S T SV T . (A.18)

The eigenvectors vi are known as principal components directions of X. The vector

u1 1 is the principal component of X and satises

u1 1 = arg max Ex [(w T x)2 ], (A.19)

w=1

i.e. it is the projection along the direction having maximum variance, which equals to
12 /n. Recursively, the kth principal component maximizes the projection variance
subject to the constraint that is must be orthogonal to the previous k 1 principal
components. PCA is often used in statistics and machine learning for reducing the
dimensionality of data. In this case, only the coordinates of data points with respect
to the rst K < n principal components are conserved.
APPENDIX A 243

A.4 Markov Chains

A nite-state discrete-time Markov chain consists of a set of random variables X =
{Xt , t {0, 1, 2, . . .}}. The random variable at any time t, Xt , takes values in a nite
set which we can index as {1, . . . , M} without loss of generality. The key property of
a Markov chain is that the probability of the next state Xt+1 , given the values of all
previous states X0 , . . . , Xt , can be written as

P (Xt+1 = j | X0 = i0 , . . . , Xt1 = in1 , Xt = i) = P (Xt+1 = j | Xt = i).

This is the standard rst-order Markov assumption: the probability of the future is
independent of the past given the current state.
If the probabilities P (Xt+1 = j | Xt = i) are independent of time t, then we say
that the Markov chain is stationary and denote the corresponding set of probabilities
as Pij , 1
i, j
M. The set of probabilities Pij can be conveniently described by
an M M transition matrix T , where the rows sum to unity, i.e.

Pij = 1.
j

For example, a three-state stationary Markov chain could have the transition matrix:

0.7 0.2 0.1
T = 0.1 0.8 0.1 . (A.20)
0.2 0.3 0.5

A.5 Information Theory

The concept of information (Aczel and Daroczy 1975; Chaitin 1987; Cover and
Thomas 1991) is central to science, technology, and many other human endeavors,
hence to the Internet and the World Wide Web. Information ought to play also a fun-
damental role in connection with the Web. Our starting point in this section is the
denition provided by Claude Shannon over half a century ago (Shannon 1948a,b),
which is at the center of modern telecommunication systems. As we shall see, how-
ever, the concept of information is complex and multi-faceted. Shannons denition
does not capture all aspects of information and, for Web applications in particular,
there is room for renements.
Before we review Shannons denition of information, we must rst review three
mathematical concepts that are essential in information theory and also other elds
such as statistical mechanics, and which are used in several chapters of this book.
These are the entropy, the mutual information, and the relative entropy. These con-
cepts are essential for the study of how information is transformed through a variety
of operations such as information coding, transmission, and compression. The rela-
tive entropy is the most general concept, from which the other two can be derived.
However, here we begin with the slightly simpler concept of entropy.
244 APPENDIX A

A.5.1 Mathematical background

Entropy
The three most basic concepts and measures of information are the entropy, the mutual
information, and the relative entropy (Blahut 1987; Cover and Thomas 1991; Shannon
1948a,b). The entropy H (p) of a probability vector p = (p1 , . . . , pn ) is dened by

n
H(p) = E( log p) = pi log pi . (A.21)
i=1

The units used to measure entropy depend on the base used for the logarithms. When
the base is two, the entropy is measured in bits. The entropy measures the prior
uncertainty in the outcome of a random experiment described by p, or the information
gained when the outcome is observed. It is also the minimum average number of bits
(when the logarithms are taken base 2) needed to transmit the outcome in the absence
of noise. The corresponding concept in the case of a continuous random variable X
with density p(x) is called the differential entropy
 +
H (X) = p(x) log p(x) dx (A.22)

and likewise when x is vector valued.

Of all the densities with variance 2 , the Gaussian N (, ) is the one with the
largest differential entropy. The differential entropy of a Gaussian distribution with any
mean and variance 2 is given by [log 2 e 2 ]/2. More generally, in n dimensions,
consider a random vector X, with vector mean , covariance matrix C, and density
p. Then the differential entropy of p satises
H (p)
1
2 log(2 e)n |C| = H(N (, C)) (A.23)
with equality if and only if X is distributed according to N (, C) almost everywhere.
Here |C| denotes the determinant of C.
The concept of entropy can be derived axiomatically. Indeed, consider a random
variable X that can assume the values x1 , . . . , xn with probabilities p1 , . . . , pn . The
goal is to dene a quantity H(p) = H (X) = H (p1 , . . . , pn ) that measures, in a
unique way, the amount of uncertainty represented in this distribution. It is a remark-
able fact that three commonsense axioms, really amounting to only one composition
law, are sufcient to determine H uniquely, up to a constant factor corresponding to
a choice of scale. The three axioms are as follows.
(1) H is a continuous function of the pi .
(2) If all pi are equal, then H (p) = H (n) = H (1/n, . . . , 1/n) is a monotonic
increasing function of n.
(3) Composition law: partition all the events xi into k disjoint sets. Let Ai
represent
the indices of the events associated with the ith set, so that qi = j Ai pj
APPENDIX A 245

represents the corresponding probability. Then


k  
pi
H (p) = H (q) + qi H , (A.24)
qi
i=1

where pi denotes the set of probabilities pj for j Ai . Thus, for example, the
composition law states that, by grouping the rst two events into one,

H ( 13 , 16 , 21 ) = H ( 21 , 21 ) + 21 H ( 23 , 31 ) + 21 H (1). (A.25)

From the rst condition, it is sufcient to determine H for all rational cases where
pi = ni /n, i = 1, . . . , n. But from the second and third conditions,
 n  
n
H ni = H (p1 , . . . , pn ) + pi H (ni ). (A.26)
i=1 i=1

For example,

H(9) = H ( 39 , 49 , 29 ) + 39 H (3) + 49 H (4) + 29 H (2). (A.27)

In particular, by setting all ni equal to m, from Equation (A.26) we get

H (m) + H (n) = H (mn). (A.28)

This yields the unique solution

H (n) = C log n, (A.29)

with C > 0. By substituting in Equation (A.26), we nally have


n
H(p) = C pi log pi . (A.30)
i=1

The constant C determines the base of the logarithm. Base 2 logarithms lead to
a measure of entropy and information in bits. For most mathematical calculations,
however, we use natural logarithms so that C = 1.
It is not very difcult to verify that the entropy has the following properties:
H(p)  0;
H(p | q)
H(P ) with equality if and only if p and q are independent;

H(p1 , . . . , pn )
ni=1 H (pi ) with equality if and only if pi are independent;
H (p) is convex () in p;

H(p1 , . . . , pn ) = ni=1 H (pi | pi1 , . . . , p1 );
H(p)
H (n) with equality if and only if p is uniform.
246 APPENDIX A

Relative entropy
The relative entropy between two density vectors

p = (p1 , . . . , pn ), q = (q1 , . . . , qn ),

or the associated random variables X and Y , is dened by


n
pi
H (p, q) = H (X, Y ) = pi log . (A.31)
qi
i=1

The relative entropy is also called cross-entropy, or KullbackLeibler distance

(Kullback 1968), or discrimination (see Shore and Johnson (1980), and references
therein, for an axiomatic presentation of the relative entropy). It can be viewed as
a measure of dissimilarity between p and q. The more dissimilar p and q are, the
larger the relative entropy. The relative entropy can also be viewed as the amount
of information that a measurement gives about the truth of a hypothesis compared
with an alternative hypothesis. It is also the expected value of the log-likelihood
ratio. Strictly speaking, the relative entropy is not symmetric and therefore is not a
distance. It can be made symmetric by using the divergence H (p, q) + H (q, p). But
in most cases, the symmetric version is not needed. If u = (1/n, . . . , 1/n) denotes the
uniform density, then H (p, u) = log n H(p). In this sense, except for a constant,
the entropy is equal to the relative entropy with respect to the uniform distribution.
By using the Jensen inequality (see below), it is easy to verify the following two
important properties of relative entropy:
H (p, q)  0 with equality if and only if p = q;
H (p, q) is convex () in p and q.
The relative entropy is not symmetric but has well known theoretical advantages
over other possible measures, including invariance with respect to reparameteriza-
tions.

Mutual information
The third concept for measuring information is the mutual information. Consider two
density vectors p and q associated with a joint distribution r over the product space.
The mutual information I(p, q) is the relative entropy between the joint distribution
r and the product of the marginals p and q:

I(p, q) = H (r, pq). (A.32)

As such, it is always positive. When r is factorial, i.e. equal to the product of the
marginals, the mutual information is zero. The mutual information is a special case
of relative entropy. Likewise, the entropy (or self-entropy) is a special case of mutual
information because H (p) = I(p, p). Furthermore, the mutual information satises
the following properties:
APPENDIX A 247

I(p, q) = 0 if and only if p and q are independent;


I(p1 , . . . , pn , q) = ni=1 I(pi , q | p1 , . . . , pi1 ).
It is easy to understand mutual information in Bayesian terms: it represents the
reduction in uncertainty of one variable when the other is observed, that is between
the prior and posterior distributions. If we denote two random variables by X and Y ,
the uncertainty in X is measured by the entropy of its prior

H (X) = P (X = x) log P (X = x).
x

Once we observe Y = y, the  uncertainty in X is the entropy of the posterior dis-

tribution, H (X | Y = y) = x P (X = x | Y = y) log P (X = x | Y = y). This is a
random variable that depends on the observation y. Its average over the possible ys
is called the conditional entropy:

H(X | Y ) = P (y)H (X | Y = y). (A.33)
y

Therefore the difference between the entropy and the conditional entropy measures
the average information that an observation of Y brings about X. It is straightforward
to check that
I(X, Y ) = H (X) H (X | Y ) = H (Y ) H (Y | X)
= H (X) + H (Y ) H (Z) = I(Y, X), (A.34)
where H (Z) is the entropy of the joint variable Z = (X, Y ). We leave the reader to
draw the classical Venn diagram associated with these relations.

Jensens inequality
A key theorem for the reader interested in proving some of the results in the previous
sections is Jensens inequality. If a function f is convex (), and X is a random
variable, then
E(f (X))
f (E(X)). (A.35)
Furthermore, if f is strictly convex, equality implies that X is constant. This inequality
is intuitively obvious by thinking in terms of the center of gravity of a set of points on
the curve f . The center of gravity of f (x1 ), . . . , f (xn ) is below f (x ), where x is the
center of gravity of x1 , . . . , xn . As a special important case, E(log X)
log(E(X)).
This, for instance, immediately yields the properties of the relative entropy.

A.5.2 Information, surprise, and relevance

Shannons information
While several approaches for quantifying information have been proposed, the most
successful one so far has been Claude Shannons denition introduced over half
248 APPENDIX A

a century ago (Blahut 1987; Cover and Thomas 1991; McEliece 1977; Shannon
1948a,b). According to Shannon, the information contained in a data set D is given
by log P (D), and the average information over the family D of possible data sets
is the entropy H (P (D)).
Shannons theory of information, although eminently successful for the develop-
ment of modern computer and telecommunication technologies, does not capture
subjective and semantic aspects of information that are not directly related to its
transmission. As pointed out in the title of Shannons seminal article, it is a theory of
communication, in the sense of transmission rather than information. It concentrates
on the problem of reproducing at one point either exactly or approximately a mes-
sage selected at another point regardless of the relevance of the message. But there is
clearly more to information than data reproducibility and somehow information ought
to depend also on the observer. Consider for instance the genomic DNA sequence of
the AIDS virus. It is a string of about 10 000 letters over the four-letter DNA alphabet,
of great signicance to researchers in the biological or medical sciences, but utterly
uninspiring to a layman. The limitations of Shannons denition may in part explain
why the theory has not been as successful as one would have hoped in other areas of
science such as biology, psychology, economics, or the Web.
Shannons theory fails to account how data can have different signicance for
different observers. This is rooted in the origin of the probabilities used in the denition
of information.
These probabilities are dened according to an observer or a model M (the Bell
Labs engineer (Jaynes 2003)) which Shannon does not describe explicitly, so that
the information in a data set is rather the negative log-likelihood

I (D, M) = log P (D | M) (A.36)

and the corresponding entropy is the average over all data sets

I (D, M) = H (P (D | M)) = P (D | M) log P (D | M) dD. (A.37)
D

There are situations, however, characterized by the presence of multiple models and/or
observers and where the subjective/semantic dimensions of the data are more impor-
tant than its transmission.
In a Web context, imagine surng the Web in search of a car and stumbling on a
picture of Marilyn Monroe. The Shannon information contained in the picture depends
on the picture resolution, whether it is color or black and white, etc. In this situation,
it is probably a secondary consideration. More important are the facts that the picture
is unexpected, i.e. surprising, and irrelevant to the goal of nding a car. Thus there
are at least three different aspects of information contained in data: the transmission
or Shannons information, the surprise, and the relevance. We now provide a precise
denition of surprise.
APPENDIX A 249

Surprise
The effect of the information contained in D is clearly to change the belief of the
observer from P (M) to P (M | D). Thus, a complementary way of measuring infor-
mation carried by the data D is to measure the distance between the prior and the
posterior. To distinguish it from Shannons communication information, we call this
notion of information the surprise information or surprise (Baldi 2002)

S(D, M) = d[P (M), P (M | D)], (A.38)

where d is a distance or similarity measure. It is natural to use the relative entropy so

that surprise is given by

P (M)
S(D, M) = H (P (M), P (M | D)) = P (M) log dM
M P (M | D)

= H (P (M)) P (M) log P (M | D) dM

= log P (D) P (M) log P (D | M) dM. (A.39)
M

Alternatively, we can dene the single model surprise by the log-odd ratio
P (M)
S(D, M) = log (A.40)
P (M | D)
and the surprise by its average

S(D, M) = S(D, M)P (M) dM, (A.41)
M

taken with respect to the prior distribution over the model class. Unlike Shannons
entropy, which requires integration over the space of data, surprise is a dual notion
that requires integration over the space of models.
Note that this denition addresses the TV snow paradox: snow, the most boring
of all television programs, carries the largest amount of Shannon information in terms
of exact reproducibility. At the time of snow onset, the image distribution we expect
and the image we perceive are very different and therefore the snow carries a great
deal of both surprise and Shannons information. Indeed snow may be a sign of storm,
earthquake, toddlers curiosity, or military putsch. But after a few seconds, once our
model of the image shifts toward a snow model of random pixels, television snow
perfectly ts the prior and hence becomes boring. Since the prior and the posterior
are virtually identical, snow frames carry zero surprise although carrying megabytes
of Shannons information.

Computing surprise
To measure surprise in the most simple settings, consider a data set consisting of N
binary points. The simplest class M(x) of models contains a single parameter x, the
250 APPENDIX A

probability of success. A natural prior distribution is the Dirichlet or Beta prior on x

with parameters (1 , 2) and mean 1 /(1 + 2 ). With a number n of successes, we
have seen in Chapter 1 that the posterior is also a Dirichlet distribution with parameters
1 + n and 2 + (N n). When N , and n = pN with 0 < p < 1, it can be
shown (Baldi 2002) that

S(D, M) N H(p, 1 /(1 + 2 )), (A.42)

where H represents the KullbackLeibler divergence distance between the empirical

distribution (p, 1 p) and the expectation of the prior (1 /(1 + 2 ), 2 /(1 + 2 )).
Thus, asymptotically, surprise information grows linearly with the number of data
points with a proportionality coefcient that depends on the discrepancy between
the expectation of the prior and the observed distribution. Similar equations can be
derived in the multinomial case and for simple continuous Gaussian models.

Surprise and learning

There is an immediate connection between surprise and computational learning theory
(Vapnik 1995). If we imagine that data points from a training set are presented sequen-
tially, we can consider that the posterior distribution after the N th point becomes the
prior for the next iteration (sequential Bayesian learning). In this case we can expect
on average surprise to decrease after each iteration, since as a system learns what is
relevant in a data set, new data points become less and less surprising. This can be
quantied precisely, at least in simple cases.
For instance, in the discrete binary case above, where the prior and posterior at
each step are Dirichlet distributions, it can be shown that the surprise S between step
N and N + 1 satises
1p
0
S
. (A.43)
pN
Thus surprise decreases in time with the number of examples as 1/N . Non-zero
surprise is a prerequisite to learning, viewed as an adjustment of probability distribu-
tions, and repeated recalibration of surprise in training sets is a prerequisite to prevent
boredom and sustain learning.

Relevance
Surprise is a measure of dissimilarity between the prior and posterior distributions
and as such it lies close to the axiomatic foundation of Bayesian probability. Surprise
is different from other denitions of information that have been proposed (Aczel and
Daroczy 1975) as alternatives to Shannons entropy. Most alternative denitions, such
as Rnyis entropies, are actually algebraic variations on Shannons denition rather
than conceptually different approaches.
Scoring items by surprise provides a general principle for the rapid detection and
ranking of unusual events and the construction of saliency maps, in any feature space,
that can guide the deployment of attention (Itti and Koch 2001; Nothdurft 2000;
APPENDIX A 251

Olshausen et al. 1993) and other rapid ltering mechanisms in natural or synthetic
information processing systems.
The notion of surprise, however, has its own limitations. In particular, it does not
capture all the semantics/relevance aspects of data. When the degree of surprise of the
data with respect to the model class becomes low, the data are no longer informative
for the given model class. This, however, does not necessarily imply that one has
a good model, since the model class itself could be unsatisfactory and in need of a
complete overhaul. Conversely, highly surprising data could be a sign that learning is
required, or that the data are irrelevant, as in the case of the TV snow or the Marilyn
Monroe picture.
Thus, relevance, surprise, and Shannons entropy are three different facets of infor-
mation that can be present in different combinations. The notion of relevance in
particular seems to be the least understood although there have been several attempts
(Jumarie 1990; Tishby et al. 1999). A possible direction is to consider, in addition to
the space of data and models, a third space A of actions or interpretations and dene
relevance as the relative entropy between the prior P (A) and the posterior P (A | D)
distributions over A. Whether this approach simply shifts the problem into the de-
nition of the set A remains to be seen. In any event, the quest to understand the nature
of information, and in particular of semantic relevance, goes well beyond the domain
of Web applications and is far from being over.
This Page Intentionally Left Blank
Appendix B

List of Main Symbols and

Abbreviations

General Guidelines
While we use the following guidelines, occasional exceptions are possible and clearly
indicated in the text. In general, vectors and matrices are in bold face, with matrices
represented by capital letters and vectors by lowercase letters. Unless otherwise spec-
ied, all vectors are column vectors by default. In general, Greek letters represent
parameters.

X = (xij ) matrix
x = (xi ) vector
XT transpose of X
trX trace of X
|X| = det X determinant of X
D data. In a typical unsupervised case the data are a matrix
D = (X) with one example per row. In a typical
supervised case, the data are a double matrix
D = (X, Y ), where Y denotes the targets, with one
example per row.
M (M( )) model (model with parameter )
parameters for some model M
( ML , MP , MAP ) parameter estimates (using maximum likelihood,
maximum a posteriori and mean posterior estimates)
M the universe of models under consideration
D the universe of possible data sets
n number of training examples
m number of dimensions

Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
254 APPENDIX B

x generic input vector

xj j th component of x
xi ith training example, typically an m-dimensional row
vector
xij j th component of the ith training example
ci class label for training example xi in supervised learning
K number of classes, clusters, or mixture components.
Classes are denoted by 1, 2, . . . , K.
T time horizon or sequence length
subscript t time index, in algorithmic iterations or in sequences

Probabilities
P ()(Q, R . . .) probability (probability density functions)
E[] (EQ []) expectation (expectation with respect to Q)
var[] variance
cov[] covariance
P (x1 , . . . , xn ) probability that X1 = x1 , . . . , Xn = xn .
P (X | Y )(E[X | Y ]) conditional probability (conditional expectation)
I background information
N (, ), N (, C), Normal (or Gaussian) density with mean and variance
N (, 2 ), N (x; , 2 ) 2 , or covariance matrix C
(w | , ) Gamma density with parameters and
B(n, p) Binomial distribution with n independent Bernouilli
trials each with probability p of success
P () Poisson distribution with parameter
Dq , Du Dirichlet distribution with
 parameters and q, or u
(ui = qi , qi  0, and i qi = 1)

Functions
E energy, error, negative log-likelihood or log-posterior
(depending on context)
L Lagrangian
H(p), H (X) entropy of the probability vector p, or the random
variable X/differential entropy in continuous case
H (p, q), H (X, Y ) relative entropy between the probability vectors p and q,
or the random variables X and Y
APPENDIX B 255

I(p, q), I(X, Y ) mutual information between the probability vectors p

and q, or the random variables X and Y
S surprise
Z partition function or normalizing factor (sometimes
also C)
C constant or normalizing factor
(x, y) Kronecker function equal to one if x = y, and zero
otherwise
f, f generic function and derivative of f
(x) Gamma function
B(, q) Beta function
We also use convex () to denote upward convexity (positive second derivative),
and convex () to denote downward convexity (negative second derivative), rather
than the more confusing convex and concave expressions.

Graphs and Sets

G = (V , E) undirected graph with vertex set V and edge set E
G = (V , E) directed graph with vertex set V and edge set E
u, v, w, etc. vertices
T tree
kv degree of vertex v (undirected case)
tv time at which vertex v is added
n number of vertices
e number of edges
d average diameter
c cliquishness index
N (i) neighbors of vertex i
ch[v] children of vertex v in a directed graph
pa[v] parents of vertex v in a directed graph
N (I ) neighbors or boundary of a set I of vertices
|S| size of set S

Abbreviations
DNS domain name service
EM expectation maximization
HMM hidden Markov model
256 APPENDIX B

HTML HyperText Markup Language

HTTP Hypertext Transfer Protocol
i.i.d independent, identically distributed
IP Internet Protocol
LMS least mean square
MAP maximum a posteriori
MaxEnt maximum entropy
MCMC Markov chain Monte Carlo
ML maximum likelihood
MP mean posterior
pdf probability density function
SGML Standard Generalized Markup Language
SVM support vector machine
TCP Transmission Control Protocol
URL uniform resource locator
References

Abello, J., Buchsbaum, A. and Westbrook, J. 1998 A functional approach to external graph
algorithms. In Proc. 6th Eur. Symp. on Algorithms, pp. 332343.
Achacoso, T. B. and Yamamoto, W. S. 1992 Ays Neuroanatomy of C. elegans for Computation.
Boca Raton, FL: CRC Press.
Aczel, J. and Daroczy, Z. 1975 On measures of information and their characterizations. New
York: Academic Press.
Adamic, L., Lukose, R. M., Puniyani, A. R. and Huberman, B. A. 2001 Search in power-law
networks. Phys. Rev. E 64, 046135.
Aggarwal, C. C., Al-Garawi, F. and Yu, P. S. 2001 Intelligent crawling on the World Wide Web
with arbitrary predicates. In Proc. 10th Int. World Wide Web Conf., pp. 96105.
Aiello, W., Chung, F. and Lu, L. 2001 A Random Graph Model for Power Law Graphs. Exper-
imental Math. 10, 5366.
Aji, S. M. and McEliece, R. J. 2000 The generalized distributive law. IEEE Trans. Inform.
Theory 46, 325343.
Albert, R. and Barabsi, A.-L. 2000 Topology of evolving networks: local events and univer-
sality. Phys. Rev. Lett. 85, 52345237.
Albert, R., Jeong, H. and Barabsi, A.-L. 1999 Diameter of the World-Wide Web. Nature 401,
130.
Albert, R., Jeong, H. and Barabsi, A.-L. 2000 Error and attack tolerance of complex networks.
Nature 406, 378382.
Allwein, E. L., Schapire, R. E. and Singer, Y. 2000 Reducing multiclass to binary: a unifying
approach for margin classiers. In Proc. 17th Int. Conf. on Machine Learning, pp. 916.
San Francisco, CA: Morgan Kaufmann.
Amaral, L. A. N., Scala, A., Barthlmy, M. and Stanley, H. E. 2000 Classes of small-world
networks. Proc. Natl Acad. Sci. 97, 11 14911 152.
Amento, B., Terveen, L. and Hill, W. 2000 Does authority mean quality? Predicting expert
quality ratings of Web documents. In Proc. 23rd Ann. Int. ACM SIGIR Conf. on Research
and Development in Information Retrieval, pp. 296303. New York: ACM Press.
Anderson, C. R., Domingos, P. and Weld, D. 2001Adaptive Web navigation for wireless devices.
In Proc. 17th Int. Joint Conf. on Articial Intelligence, pp. 879884. San Francisco, CA:
Morgan Kaufmann.
Anderson, C. R., Domingos, P. and Weld, D. 2002 Relational markov models and their applica-
tion to adaptive Web navigation. In Proc. 8th Int. Conf. on Knowledge Discovery and Data
Mining, pp. 143152. New York: ACM Press.

Modeling the Internet and the Web P. Baldi, P. Frasconi and P. Smyth
2003 P. Baldi, P. Frasconi and P. Smyth ISBN: 0-470-84906-1
258 REFERENCES

Androutsopoulos, I., Koutsias, J., Chandrinos, K. and Spyropoulos, D. 2000 An experimental

comparison of naive Bayesian and keyword-based anti-spam ltering with personal e-mail
messages. In Proc. 23rd ACM SIGIR Ann. Conf., pp. 160167.
Ansari, A. and Mela, C. 2003 E-customization. J. Market. Res. (In the press.)
Apostol, T. M. 1969 Calculus, vols I and II. John Wiley & Sons, Ltd/Inc.
Appelt, D., Hobbs, J., Bear, J., Israel, D., Kameyama, M., Kehler, A., Martin, D., Meyers K.
and Tyson, M. 1995 SRI International FASTUS system: MUC-6 test results and analysis.
In Proc. 6th Message Understanding Conf. (MUC-6), pp. 237248. San Francisco, CA:
Morgan Kaufmann.
Apt C., Damerau, F. and Weiss, S. M. 1994 Automated learning of decision rules for text
categorization. (Special Issue on Text Categorization.) ACM Trans. Informat. Syst. 12, 233
251.
Arajo M. D., Navarro, G. and Ziviani, N. 1997 Large text searching allowing errors. In Proc.
4th South American Workshop on String Processing (ed. R. Baeza-Yates), International
Informatics Series, pp. 220. Ottawa: Carleton University Press.
Armstrong, R., Freitag, D., Joachims, T. and Mitchell, T. 1995 WebWatcher: a learning appren-
tice for the World Wide Web. In Proc. 1995 AAAI Spring Symp. on Information Gathering
from Heterogeneous, Distributed Environments, pp. 612.
Aslam, J. A. and Montague, M. 2001 Models for metasearch. In Proc. 24th Ann. Int. ACM
SIGIR Conf. on Research and Development in Information Retrieval, pp. 276284. New
York: ACM Press.
Baldi, P. 2002 A computational theory of surprise. In Information, Coding, and Mathematics
(ed. M. Blaum, P. G. Farrell and H. C. A. van Tilborg), pp. 125. Boston, MA: Kluwer
Academic.
Baldi, P. and Brunak, S. 2001 Bioinformatics: The Machine Learning Approach, 2nd edn. MIT
Press, Cambridge, MA.
Baluja, S., Mittal, V. and Sukthankar, R. 2000 Applying machine learning for high performance
named-entity extraction. Computat. Intell. 16, 586595.
Barabsi, A.-L. and Albert, R. 1999 Emergence of scaling in random networks. Science 286,
509512.
Barabsi, A.-L., Albert, R. and Jeong, H. 1999 Mean-eld theory for scale-free random net-
works. Physica A 272, 173187.
Barabsi, A.-L., Freeh, V. W., Jeong, H. and Brockman, J. B. 2001 Parasitic computing. Nature
412, 894897.
Barlow, R. and Proshan, F. 1975 Statistical Theory of Reliability and Life Testing. Austin, TX:
Holt, Rinehart and Winston.
Barthlmy, M. and Amaral, L. A. N. 1999 Small-world networks: evidence for a crossover
picture. Phys. Rev. Lett. 82, 31803183.
Bass, F. M. 1969 A new product growth model for consumer durables. Mngmt Sci. 15, 215227.
Bellman, R. E. 1957 Dynamic Programming. Princeton, NJ: Princeton University Press.
Bengio, Y. and Frasconi, P. 1995 An inputoutput HMM architecture. Adv. Neural Inf. Process
Syst. 7, 427434.
Berger, J. O. 1985 Statistical Decision Theory and Bayesian Analysis. Springer.
Bergman, M. 2000 The Deep Web: surfacing hidden value. J. Electron. Publ. 7. (Available at
http://www.completeplanet.com/Tutorials/DeepWeb/.)
Berners-Lee, T. 1994 Universal Resource Identiers in WWW: A Unifying Syntax for the
Expression of Names and Addresses of Objects on the Network as used in the World-Wide
Web. RFC 1630. (Available at http://www.ietf.org/rfc/rfc1630.txt.)
Berners-Lee, T., Fielding, R. and Masinter, L. 1998 Uniform Resource Identiers (URI):
Generic Syntax. RFC 2396. (Available at http://www.ietf.org/rfc/rfc2396.txt.)
REFERENCES 259

Berry, M. W. 1992 Large scale singular value computations. J. Supercomput. Applic. 6, 1349.
Berry, M. W. and Browne, M. 1999 Understanding Search Engines: Mathematical Modeling
and Text Retrieval. Philadelphia, PA: Society for Industrial and Applied Mathematics.
Bharat, K. and Broder, A. 1998 A technique for measuring the relative size and overlap of
public Web search engines. In Proc. 7th Int. World Wide Web Conf., Brisbane, Australia,
pp. 379388.
Bharat, K. and Henzinger, M. R. 1998 Improved algorithms for topic distillation in a hyperlinked
environment. In Proc. 21st Ann Int. ACM SIGIR Conf. on Research and Development in
Information Retrieval, pp. 104111. New York: ACM Press.
Bianchini, M., Gori, M. and Scarselli, F. 2001 Inside Googles Web page scoring system.
Technical report, Dipartimento di Ingegneria dellInformazione, Universit di Siena.
Bikel, D. M., Miller, S., Schwartz, R. and Weischedel, R. 1997 Nymble: a high-performance
learning name-nder. In Proceedings of ANLP-97, pp. 194201. (Available at http://
citeseer.nj.nec.com/bikel97nymble.html.)
Billsus, D. and Pazzani, M. 1998 Learning collaborative information lters. In Proc. Int. Conf.
on Machine Learning, pp. 4654. San Francisco, CA: Morgan Kaufmann.
Blahut, R. E. 1987 Principles and Practice of Information Theory. Reading, MA: Addison-
Wesley.
Blei, D., Ng, A. Y. and Jordan, M. I. 2002a Hierarchical Bayesian models for applications in
information retrieval. In Bayesian Statistics 7 (ed. J. M. Bernardo, M. Bayarri, J. O. Berger,
A. P. Dawid, D. Heckerman, A. F. M. Smith and M. West). Oxford University Press.
Blei, D., Ng, A. Y. and Jordan, M. I. 2002b Latent Dirichlet allocation. In Advances in Neural
Information Processing Systems 14 (ed. T. Dietterich, S. Becker and Z. Ghahramani). San
Francisco, CA: Morgan Kaufmann.
Blum, A. and Mitchell, T. 1998 Combining labeled and unlabeled data with co-training. In
Proc. 11th Ann. Conf. on Computational Learning Theory (COLT-98), pp. 92100. New
York: ACM Press.
Bollacker, K. D., Lawrence, S. and Giles, C. L. 1998 CiteSeer: an autonomous Web agent for
automatic retrieval and identication of interesting publications. In Proc. 2nd Int. Conf. on
Autonomous Agents (Agents98) (ed. K. P. Sycara and M. Wooldridge), pp. 116123. New
York: ACM Press.
Bollobs, B. 1985 Random Graphs. London: Academic Press.
Bollobs, B. and de la Vega, W. F. 1982 The diameter of random regular graphs. Combinatorica
2, 125134.
Bollobs, B. and Riordan, O. 2003 The diameter of a scale-free random graph. Combinatorica.
(In the press.)
Bollobs, B., Riordan, O., Spencer, J. and Tusndy G. 2001 The degree sequence of a scale-free
random graph process. Random. Struct. Alg. 18, 279290.
Borodin, A., Roberts, G. O., Rosenthal, J. S. and Tsaparas, P. 2001 Finding authorities and
hubs from link structures on the World Wide Web. In Proc. 10th Int. Conf. on World Wide
Web, pp. 415429.
Box, G. E. P. and Tiao, G. C. 1992 Bayesian Inference In Statistical Analysis. John Wiley &
Sons, Ltd/Inc.
Boyan, J., Freitag, D. and Joachims, T. 1996 A machine learning architecture for optimizing
Web search engines. In Proc. AAAI Workshop on Internet-Based Information Systems.
Brand, M. 2002 Incremental singular value decomposition of uncertain data with missing
values. In Proc. Eur. Conf. on Computer Vision (ECCV): Lecture Notes in Computer Science,
pp. 707720. Springer.
Bray, T. 1996 Measuring the Web. InProc. 5th Int. Conf. on the World Wide Web, 610 May
1996, Paris, France. Comp. Networks 28, 9931005.
260 REFERENCES

Breese, J. S., Heckerman, D. and Kadie, C. 1998 Empirical analysis of predictive algorithms for
collaborative ltering. In Proc. 14th Conf. on Uncertainty in Articial Intelligence, pp. 43
52. San Francisco, CA: Morgan Kaufmann.
Brewington, B. and Cybenko, G. 2000 How dynamic is the Web? Proc. 9th Int. World Wide
Web Conf. Geneva: International World Wide Web Conference Committee (IW3C2).
Brin, S. and Page, L. 1998 The anatomy of a large-scale hypertextual (Web) search engine. In
Proc. 7th Int. World Wide Web Conf. (WWW7). Comp. Networks 30, 107117.
Broder, A., Kumar, R., Maghoul, F., Raghavan, P., Rajagopalan, S., Stata, R., Tomikns, A. and
Wiener, J. 2000 Graph structure in the Web. In Proc. 9th Int. World Wide Web Conf. (WWW9).
Comp. Networks 33, 309320.
Brown, L. D. 1986 Fundamentals of Statistical Exponential Families. Hayward, CA: Institute
of Mathematical Statistics.
Bucklin, R. E. and Sismeiro, C. 2003 A model of Web site browsing behavior estimated on
clickstream data. (In the press.)
Buntine, W. 1992 Learning classication trees. Statist. Comp. 2, 6373.
Buntine, W. 1996 A guide to the literature on learning probabilistic networks from data. IEEE
Trans. Knowl. Data Engng 8, 195210.
Byrne, M. D., John, B. E., Wehrle, N. S. and Crow, D. C. 1999 The tangled Web we wove: a
taskonomy of WWW use. In Proc. CHI99: Human Factors in Computing Systems, pp. 544
551. New York: ACM Press.
Cadez, I. V., Heckerman, D., Smyth, P., Meek, C. and White, S. 2003 Model-based clustering
and visualization of navigation patterns on a Web site. Data Mining Knowl. Discov. (In the
press.)
Califf, M. E. and Mooney, R. J. 1998 Relational learning of pattern-match rules for information
extraction. Working Notes of AAAI Spring Symp. on Applying Machine Learning to Discourse
Processing, pp. 611. Menlo Park, CA: AAAI Press.
Callaway, D. S., Hopcroft, J. E., Kleinberg, J., Newman, M. E. J. and Strogatz, S. H. 2001 Are
randomly grown graphs really random? Phys. Rev. E 64, 041902.
Cardie, C. 1997 Empirical methods in information extraction. AI Mag. 18, 6580.
Carlson, J. M. and Doyle, J. 1999 Highly optimized tolerance: a mechanism for power laws in
designed systems. Phys. Rev. E 60, 14121427.
Castelli, V. and Cover, T. 1995 On the exponential value of labeled samples. Pattern Recog.
Lett. 16, 105111.
Catledge, L. D. and Pitkow, J. 1995 Characterizing browsing strategies in the World-Wide
Web. Comp. Networks ISDN Syst. 27, 10651073.
Chaitin, G. J. 1987 Algorithmic Information Theory. Cambridge University Press.
Chakrabarti, S., Dom, B., Gibson, D., Kleinberg, J., Kumar, S. R., Raghavan, P., Rajagopalan
S. and Tomkins, A. 1999a Mining the link structure of the World Wide Web. IEEE Computer
32, 6067.
Chakrabarti, S., Joshi, M. M., Punera, K. and Pennock, D. M. 2002 The structure of broad
topics on the Web. In Proc. 11th Int. Conf. on World Wide Web, pp. 251262. New York:
ACM Press.
Chakrabarti, S., van den Berg, M. and Dom, B. 1999b Focused crawling: a new approach to
topic-specic Web resource discovery. In Proc. 8th Int. World Wide Web Conf., Toronto.
Comp. Networks 31, 1116.
Charniak, E. 1991 Bayesian networks without tears. AI Mag. 12, 5063.
Chen, S. F. and Goodman, J. 1996 An empirical study of smoothing techniques for language
modeling. In Proc. 34th Ann. Meeting of the Association for Computational Linguistics (ed.
A. Joshi and M. Palmer), pp. 310318. San Francisco, CA: Morgan Kaufmann.
REFERENCES 261

Chickering, D. M., Heckerman, D. and Meek, C. 1997 A Bayesian approach to learning

Bayesian networks with local structure. Uncertainty in Articial Intelligence: Proc. 13th
Conf. (UAI-1997), pp. 8089. San Francisco, CA: Morgan Kaufmann.
Cho, J. and Garcia-Molina, H. 2000a Estimating frequency of change. Technical Report
DBPUBS-4, Stanford University. (Available at http://dbpubs.stanford.edu/pub/2000-4.)
Cho, J. and Garcia-Molina, H. 2000b Synchronizing a database to improve freshness. In Proc.
2000 ACM Int. Conf. on Management of Data (SIGMOD), pp. 117128.
Cho, J. and Garcia-Molina, H. 2002 Parallel crawlers. In Proc. 11th World Wide Web Conf.
(WWW11), Honolulu, Hawaii.
Cho, J. and Ntoulas, A. 2002 Effective change detection using sampling. In Proc. 28th Int.
Conf. on Very Large Databases (VLDB).
Cho, J., Garcia-Molina, H. and Page, L. 1998 Efcient crawling through URL ordering. In
Proc. 7th Int. World Wide Web Conf. (WWW7). Comp. Networks 30, 161172.
Chung, E. R. K., Graham, R. L. and Wilson, R. M. 1989 Quasi-random graphs. Combinatorica
9, 345362.
Chung, F. and Graham, R. 2002 Sparse quasi-random graphs. Combinatorica 22, 217244.
Chung, F. and Lu, L. 2001 The diameter of random sparse graphs. Adv. Appl. Math. 26, 257279.
Chung, F. and Lu, L. 2002 Connected components in random graphs with given expected degree
sequences. Technical Report, Univeristy of California, San Diego.
Chung, F., Garrett, M., Graham, R. and Shallcross, D. 2001 Distance realization problems with
applications to Internet tomography. J. Comput. Syst. Sci. 63, 432448.
Clarke, I., Sandberg, O., Wiley, B. and Hong, T. W. 2000 Freenet: a distributed anonymous
information storage and retrieval system. In Designing Privacy Enhancing Technologies:
International Workshop on Design Issues in Anonymity and Unobservability, LNCS 2009
(ed. H. Federrath), pp. 311320. Springer.
Clarke, I., Miller, S. G., Hong, T. W., Sandberg, O. and Wiley, B. 2002 Protecting free expression
online with Freenet. IEEE Internet Computing 6, 4049.
Cockburn, A. and McKenzie, B. 2002 What do Web users do? An empirical analysis of Web
use. Int. J. Human Computer Studies 54, 903922.
Cohen, W. W. 1995 Text categorization and relational learning. In Proc. ICML-95, 12th Int.
Conf. on Machine Learning (ed. A. Prieditis and S. J. Russell), pp. 124132. San Francisco,
CA: Morgan Kaufmann.
Cohen, W. W. 1996 Learning rules that classify e-mail. In AAAI Spring Symp. on Machine
Learning in Information Access (ed. M. Hearst and H. Hirsh). 1996 Spring Symposium
Series. Menlo Park, CA: AAAI Press.
Cohen, W. W. and McCallum, A. 2002 Information Extraction from the World Wide Web.
Tutorial presented at 15th Neural Information Processing Conf. (NIPS-15).
Cohen, W. W., Schapire, R. E. and Singer, Y. 1999 Learning to order things. J. Artif. Intell. Res.
10, 243270.
Cohen, W. W., McCallum, A. and Quass, D. 2000 Learning to understand the Web. IEEE Data
Enging Bull. 23, 1724.
Cohn, D. and Chang, H. 2000 Learning to probabilistically identify authoritative documents.
In Proc. 17th Int. Conf. on Machine Learning, pp. 167174. San Francisco, CA: Morgan
Kaufmann.
Cohn, D. and Hofmann, T. 2001 The missing link: a probabilistic model of document content
and hypertext connectivity. In Advances in Neural Information Processing Systems (ed. T. K.
Leen, T. G. Dietterich and V. Tresp). Boston, MA: MIT Press.
Cooley, R., Mobasher, B. and Srivastava, J. 1999 Data preparation for mining World Wide Web
browsing patterns. Knowl. Informat. Syst. 1, 532.
Cooper, C. and Frieze, A. 2001 A general model of Web graphs. Technical Report.
262 REFERENCES

Cooper, G. F. 1990 The computational complexity of probabilistic inference using Bayesian

belief networks. Art. Intell. 42, 393405.
Cooper, W. S. 1991 Some inconsistencies and misnomers in probabilistic information retrieval.
In Proc. 14th Ann. Int. ACM SIGIR Conf. on Research and Development in Information
Retrieval, pp. 5761. New York: ACM Press.
Cormen, T. H., Leiserson, C. E., Rivest, R. L. and Stein, C. 2001 Introduction to Algorithms,
2nd edn. Cambridge, MA: MIT Press.
Cortes, C. and Vapnik, V. N. 1995 Support vector networks. Machine Learning 20, 125.
Cover, T. M. and Hart, P. E. 1967 Nearest neighbor pattern classication. IEEE Trans. Inform.
Theory 13, 2127.
Cover, T. M. and Thomas, J. A. 1991 Elements of Information Theory. John Wiley & Sons,
Ltd/Inc.
Cox, R. T. 1964 Probability, frequency and reasonable expectation. Am. J. Phys. 14, 113.
Crammer, K. and Singer, Y. 2000 On the learnability and design of output codes for multiclass
problems. Computational Learning Theory, pp. 3546.
Craven, M., di Pasquo D., Freitag, D., McCallum A., Mitchell, T., Nigan, K. and Slattery,
S. 2000 Learning to construct knowledge bases from the World Wide Web. Artif. Intel.
118(12), 69113.
Craven, M. and Slattery, S. 2001 Relational learning with statistical predicate invention: better
models for hypertext. Machine Learning 43(1/2), 97119.
Cristianini, N. and Shawe-Taylor, J. 2000 An Introduction to Support Vector Machines. Cam-
bridge University Press.
Daley, D. J. and Gani, J. 1999 Epidemic Modeling: an Introduction. Cambridge University
Press.
Davison, B. D. 2000a Recognizing nepotistic links on the Web. AAAI Workshop on Articial
Intelligence for Web Search.
Davison, B. D. 2000b Topical locality in the Web. In Proc. 23rd Ann. Int. ACM SIGIR Conf.
on Research and Development in Information Retrieval, pp. 272279.
Dawid, A. P. 1992 Applications of a general propagation algorithm for probabilistic expert
systems. Stat. Comp. 2, 2536.
Day, J. 1995 The (un)revised osi reference model. ACM SIGCOMM Computer Communication
Review 25, 3955.
Day, J. and Zimmerman, H. 1983 The OSI reference model. In Proc. IEEE 71, 13341340.
De Bra, P. and Post, R. 1994 Information retrieval in the World Wide Web: making client-based
searching feasible. In Proc. 1st Int. World Wide Web Conf.
Dechter, R. 1999 Bucket elimination: a unifying framework for reasoning. Artif. Intel. 113,
4185.
Deering, S. and Hinden, R. 1998 Internet Protocol, Version 6 (IPv6) Specication. RFC 2460.
(Available at http://www.ietf.org/rfc/rfc2460.txt.)
Del Bimbo A. 1999 Visual Information Retrieval. San Francisco, CA: Morgan Kaufmann.
Dempster, A. P., Laird, N. M. and Rubin, D. B. 1977 Maximum likelihood from incomplete
data via the em algorithm. J. R. Statist. Soc. B 39, 122.
Deshpande, M. and Karypis, G. 2001 Selective Markov models for predicting Web-page
accesses. In Proc. SIAM Conf. on Data Mining SIAM Press.
Devroye, L., Gyr, L. and Lugosi, G. 1996 A Probabilistic Theory of Pattern Recognition.
Springer.
Dhillon, I. S., Fan, J. and Guan, Y. 2001 Efcient clustering of very large document collections
In Data Mining for Scientic and Engineering Applications (ed. R. Grossman, C. Kamath
and R. Naburu). Kluwer Academic.
REFERENCES 263

Dhillon, I. S. and Modha, D. S. 2001 Concept decompositions for large sparse text data using
clustering. Machine Learning 42, 143175.
Dietterich, T. G. and Bakiri, G. 1995 Solving multiclass learning problems via error-correcting
output codes. J. Articial Intelligence Research 2, 263286.
Dijkstra, E. D. 1959 A note on two problem in connexion with graphs. Numerische Mathematik
1, 269271.
Diligenti, M., Coetzee, F., Lawrence, S., Giles, C. L. and Gori, M. 2000 Focused crawling
using context graphs. In VLDB 2000, Proc. 26th Int. Conf. on Very Large Data Bases,
1014 September 2000, Cairo, Egypt (ed. A. El Abbadi, M. L. Brodie, S. Chakravarthy,
U. Dayal, N. Kamel, G. Schlageter and K. Y. Whang), pp. 527534. Los Altos, CA: Morgan
Kaufmann.
Dill, S., Kumar, S. R., McCurley, K. S., Rajagopalan, S., Sivakumar, D. and Tomkins, A. 2001
Self-similarity in the Web. In Proc. 27th Very Large Databases Conf., pp. 6978.
Domingos, P. and Pazzani, M. 1997 On the optimality of the simple Bayesian classier under
zero-one loss. Machine Learning 29, 103130.
Domingos, P. and Richardson, M. 2001 Mining the network value of customers. In Proc. ACM
7th Int. Conf. on Knowledge Discovery and Data Mining, pp. 5766. New York: ACM Press.
Dreilinger, D. and Howe, A. E. 1997 Experiences with selecting search engines using
metasearch. ACM Trans. Informat. Syst. 15, 195222.
Drucker, H., Vapnik, V. N. and Wu, D. 1999 Support vector machines for spam categorization.
IEEE Trans. Neural Networks 10, 10481054.
Duda, R. O. and Hart, P. E. 1973 Pattern Classication and Scene Analysis. John Wiley &
Sons, Ltd/Inc.
Dumais, S., Platt, J., Heckerman, D. and Sahami, M. 1998 Inductive learning algorithms and
representations for text categorization. In Proc. 7th Int. Conf. on Information and Knowledge
Management, pp. 148155. New York: ACM Press.
Jones, K. S. and Willett, P. (eds) 1997 Readings in information retrieval. San Mateo, CA:
Morgan Kaufmann.
Edwards, J., McCurley, K. and Tomlin, J. 2001 An adaptive model for optimizing performance
of an incremental Web crawler. In Proc. 10th Int. World Wide Web Conf., pp. 106113.
Elias, P. 1975 Universal codeword sets and representations of the integers. IEEE Trans. Inform.
Theory 21, 194203.
Erdos, P. and Rnyi, A. 1959 On random graphs. Publ. Math. Debrecen 6, 290291.
Erdos, P. and Rnyi, A. 1960 On the evolution of random graphs. Magy. Tud. Akad. Mat. Kut.
Intez. Kozl. 5, 1761.
Everitt B. S. 1984 An Introduction to Latent Variable Models. London: Chapman & Hall.
Evgeniou, T., Pontil, M. and Poggio, T. 2000 Regularization networks and support vector
machines. Adv. Comput. Math. 13, 150.
Fagin, R., Karlin, A., Kleinberg, J., Raghavan, P., Rajagopalan, S., Rubinfeld, R., Sudan., M.
and Tomkins, A. 2000 Random walks with back buttons. In Proc. ACM Symp. on Theory
of Computing, pp. 484493. New York: ACM Press.
Faloutsos, C. and Christodoulakis, S. 1984 Signature les: an access method for documents
and its analytical performance evaluation. ACM Trans. Informat. Syst. 2, 267288.
Faloutsos, M., Faloutsos, P. and Faloutsos, C. 1999 On power-law relationships of the Internet
topology. In Proc. ACM SIGCOMM Conf., Cambridge, MA, pp. 251262.
Feller, W. 1971 An Introduction to Probability Theory and Its Applications, 2nd edn, vol. 2.
John Wiley & Sons, Ltd/Inc.
Fermi, E. 1949 On the origin of the cosmic radiation. Phys. Rev. 75, 11691174.
264 REFERENCES

Fielding, R., Gettys, J., Mogul, J., Frystyk, H., Masinter, L., Leach, P. and Berners-Lee, T.
1999 Hypertext Transfer Protocol: HTTP/1.1. RFC 2616. (Available at http://www.ietf.org/
rfc/rfc2616.txt.)
Fienberg, S. E., Johnson, M. A. and Junker, B. J. 1999 Classical multilevel and Bayesian
approaches to population size estimation using multiple lists. J. R. Statist. Soc. A 162, 383
406.
Flake, G. W., Lawrence, S. and Giles, C. L. 2000 Efcient identication of Web communities. In
Proc. 6th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, pp. 150160.
New York: ACM Press.
Flake, G. W., Lawrence, S., Giles, C. L. and Coetzee, F. 2002 Self-organization and identica-
tion of communities. IEEE Computer 35, 6671.
Fox, C. 1992 Lexical analysis and stoplists. In Information Retrieval: Data Structures and
Algorithms (ed. W. B. Frakes and R. Baeza-Yates), ch. 7. Englewood Cliffs, NJ: Prentice
Hall.
Fraley, C. and Raftery, A. E. 2002 Model-based clustering, discriminant analysis, and density
estimation. J. Am. Statist. Assoc. 97, 611631.
Freitag, D. 1998 Information extraction from HTML: Application of a general machine learning
approach. In Proc. AAAI-98, pp. 517523. Menlo Park, CA: AAAI Press.
Freitag, D. and McCallum, A. 2000 Information extraction with HMM structures learned by
stochastic optimization. AAAI/IAAI, pp. 584589.
Freund, Y. and Schapire, R. E. 1996 Experiments with a new boosting algorithm . In Proc. 13th
Int. Conf. on Machine Learning, pp. 148146. San Francisco, CA: Morgan Kaufmann.
Frey, B. J. 1998 Graphical Models for Machine Learning and Digital Communication. MIT
Press.
Friedman, N. and Goldszmidt, M. 1996 Learning Bayesian networks with local structure. In
Proc. 12th Conf. on Uncertainty in Articial Intelligence, Portland, Oregon (ed. E. Horwitz
and F. Jensen), pp. 274282. San Francisco, CA: Morgan Kaufmann.
Friedman, N., Getoor, L., Koller, D. and Pfeffer, A. 1999 Learning probabilistic relational
models. In Proc. 16th Int. Joint Conf. on Articial Intelligence (IJCAI-99) (ed. D. Thomas),
vol. 2 , pp. 13001309. San Francisco, CA: Morgan Kaufmann.
Fuhr, N. 1992 Probabilistic models in information retrieval. Comp. J. 35, 243255.
Galambos, J. 1987 The Asymptotic Theory of Extreme Order Statistics, 2nd edn. Malabar, FL:
Robert E. Krieger.
Gareld, E. 1955 Citation indexes for science: a new dimension in documentation through
association of ideas. Science 122, 108111.
Gareld, E. 1972 Citation analysis as a tool in journal evaluation. Science 178, 471479.
Garner, R. 1967 A Computer Oriented, Graph Theoretic Analysis of Citation Index Structures.
Philadelphia, PA: Drexel University Press.
Gelbukh, A. and Sidorov, G. 2001 Zipf and Heaps Laws coefcients depend on language. In
Proc. 2001 Conf. on Intelligent Text Processing and Computational Linguistics, pp. 332335.
Springer.
Gelman, A., Carlin, J. B., Stern, H. S. and Rubin, D. B. 1995 Bayesian Data Analysis. London:
Chapman & Hall.
Ghahramani, Z. 1998 Learning dynamic Bayesian networks. In Adaptive Processing of
Sequences and Data Structures. Lecture Notes in Artical Intelligence (ed. M. Gori and
C. L. Giles), pp. 168197. Springer.
Ghahramani, Z. and Jordan, M. I. 1997 Factorial hidden Markov models. Machine Learning
29, 245273.
Ghani, R. 2000 Using error-correcting codes for text classication. In Proc. 17th Int. Conf. on
Machine Learning, pp. 303310. San Francisco, CA: Morgan Kaufmann.
REFERENCES 265

Gibson, D., Kleinberg, J. and Raghavan, P. 1998 Inferring Web communities from link topol-
ogy. In Proc. 9th ACM Conf. on Hypertext and Hypermedia : Links, Objects, Time and
Spacestructure in Hypermedia Systems, pp. 225234. New York: ACM Press.
Gilbert, E. N. 1959 Random graphs. Ann. Math. Statist. 30, 11411144.
Gilbert, N. 1997 A simulation of the structure of academic science. Sociological Research
Online 2. (Available at http://www.socresonline.org.uk/socresonline/2/2/3.html.)
Gilks, W. R., Thomas, A. and Spiegelhalter, D. J. 1994 A language and program for complex
Bayesian modelling. The Statistician 43, 6978.
Greenberg, S. 1993 The Computer User as Toolsmith: The Use, Reuse, and Organization or
Computer-Based Tools. Cambridge University Press.
Guermeur, Y., Elisseeff, A. and Paugam-Mousy, H. 2000 A new multi-class SVM based on a
uniform convergence result. In Proc. IJCNN: Int. Joint Conf. on Neural Networks, vol. 4,
pp 41834188. Piscataway, NJ: IEEE Press.
Han, E. H., Karypis, G. and Kumar, V. 2001 Text categorization using weight-adjusted k-nearest
neighbor classication. In Proc. PAKDD-01, 5th PacicAsia Conferenece on Knowledge
Discovery and Data Mining (ed. D. Cheung, Q. Li and G. Williams). Lecture Notes in
Computer Science Series, vol. 2035, pp. 5365. Springer.
Han, J. and Kamber, M. 2001 Data Mining: Concepts and Techniques. San Francisco, CA:
Morgan Kaufmann.
Hand, D., Mannila, H. and Smyth, P. 2001 Principles of Data Mining. Cambridge, MA: MIT
Press.
Harman, D., Baeza-Yates, R., Fox, E. and Lee, W. 1992 Inverted les. In Information Retrieval,
Data Structures and Algorithms (ed. W. B. Frakes and R. A. Baeza-Yates), pp. 2843.
Englewood Cliffs, NJ: Prentice Hall.
Hastie, T., Tibshirani, R. and Friedman, J. 2001 Elements of Statistical Learning: Data Mining,
Inference, and Prediction. Springer.
Heckerman, D. 1997 Bayesian networks for data mining. Data Mining Knowl. Discov. 1, 79
119.
Heckerman, D. 1998 A tutorial on learning with Bayesian networks. In Learning in Graphical
Models (ed. M. Jordan). Kluwer.
Heckerman, D., Chickering, D. M., Meek, C., Rounthwaite, R. and Kadie, C. 2000 Dependency
networks for inference, collaborative ltering, and data visualization. J. Mach. Learn. Res.
1, 4975.
Hersovici, M., Jacovi, M., Maarek, Y. S., Pelleg, D., Shtalhaim, M. and Ur, S. 1998 The shark-
search algorithm an application: tailored Web site mapping. In Proc. 7th Int. World-Wide
Web Conf. Comp. Networks 30, 317326.
Heydon, A. and Najork, M. 1999 Mercator: a scalable, extensible Web crawler. World Wide
Web 2, 219229. (Available at http://research.compaq.com/SRC/mercator/research.html.)
Heydon, A. and Najork, M. 2001 High-performance Web crawling. Technical Report SRC 173.
Compaq Systems Research Center.
Hoffman, T. 1999 Probabilistic latent semantic indexing. In Proc. 22nd Ann. Int. ACM SIGIR
Conf. on Research and Development in Information Retrieval, pp. 5057. New York: ACM
Press.
Hofmann, T. 2001 Unsupervised learning by probabilistic latent semantic analysis. Machine
Learning 42, 177196.
Hofmann, T. and Puzicha, J. 1999 Latent class models for collaborative ltering. In Proc. 16th
Int. Joint Conf. on Articial Intelligence, pp. 688693.
Hofmann, T., Puzicha, J. and Jordan, M. I. 1999 Learning from dyadic data. In Advances in
Neural Information Processing Systems 11: Proc. 1998 Conf. (ed. M. S. Kearns, S. A. Solla
and D. Cohen), pp. 466472. Cambridge, MA: MIT Press.
266 REFERENCES

Huberman, B. A. and Adamic, L. A. 1999 Growth dynamics of the World Wide Web. Nature
401, 131.
Huberman, B. A., Pirolli, P. L. T., Pitkow, J. E. and Lukose, R. M. 1998 Strong regularities in
World Wide Web surng. Science 280, 9597.
Hunter, J. and Shotland, R. 1974 Treating data collected by the small world method as a Markov
process. Social Forces 52, 321.
ISO 1986 Information Processing, Text and Ofce Systems, Standard Generalized Markup
Language (SGML), ISO 8879, 1st edn. Geneva, Switzerland: International Organization for
Standardization.
Itti, L. and Koch, C. 2001 Computational modelling of visual attention. Nature Rev. Neurosci.
2, 194203.
Jaakkola, T. S. and Jordan, I. 1997 Recursive algorithms for approximating probabilities in
graphical models. In Advances in Neural Information Processing Systems (ed. M. C. Mozer,
M. I. Jordan and T. Petsche), vol. 9, pp. 487493. Cambridge, MA: MIT Press.
Jaeger, M. 1997 Relational Bayesian networks In Proc. 13th Conf. on Uncertainty in Articial
Intelligence (UAI-97) (ed. D. Geiger and P. P. Shenoy), pp. 266273. San Francisco, CA:
Morgan Kaufmann.
Janiszewski, C. 1998 The inuence of display characteristics on visual exploratory behavior.
J. Consumer Res. 25, 290301.
Jansen, B. J., Spink, A., Bateman, J. and Saracevic, T. 1998 Real-life information retrieval: a
study of user queries on the Web. SIGIR Forum 32, 517.
Jaynes, E. T. 1986 Bayesian methods: general background. In Maximum Entropy and Bayesian
Methods in Statistics (ed. J. H. Justice), pp. 125. Cambridge University Press.
Jaynes, E. T. 2003 Probability Theory: The Logic of Science. Cambridge University Press.
Jensen, F. V. 1996 An Introduction to Bayesian Networks. Springer.
Jensen, F. V., Lauritzen, S. L. and Olesen, K. G. 1990 Bayesian updating in causal probabilistic
networks by local computations. Comput. Statist. Q. 4, 269282.
Jeong, H., Tomber, B., Albert, R., Oltvai, Z. and Barabsi, A.-L. 2000 The large-scale organi-
zation of metabolic networks. Nature 407, 651654.
Joachims, T. 1997 A probabilistic analysis of the Rocchio algorithm with TFIDF for text
categorization. In Proc. 14th Int. Conf. on Machine Learning, pp. 143151. San Francisco,
CA: Morgan Kaufmann.
Joachims, T. 1998 Text categorization with support vector machines: learning with many rele-
vant features. In Proc. 10th Eur. Conf. on Machine Learning, pp. 137142. Springer.
Joachims, T. 1999a Making large-scale SVM learning practical In Advances in Kernel Methods:
Support Vector Learning (ed. B. Schlkopf, C. J. C. Burges and A. J. Smola), pp. 169184.
Cambridge, MA; MIT Press.
Joachims, T. 1999b Transductive inference for text classication using support vector machines.
In Proc. 16th Int. Conf. on Machine Learning (ICML), pp. 200209. San Francisco, CA:
Morgan Kaufmann.
Joachims, T. 2002 Learning to Classify Text using Support Vector Machines. Kluwer.
Jordan, M. I. (ed.) 1999 Learning in Graphical Models. Cambridge, MA: MIT Press.
Jordan, M. I., Ghahramani, Z. and Saul, L. K. 1997 Hidden Markov decision trees. In Advances
in Neural Information Processing Systems (ed. M. C. Mozer, M. I. Jordan and T. Petsche),
vol. 9, pp. 501507. Cambridge, MA: MIT Press.
Jumarie, G. 1990 Relative information. Springer.
Kask, K. and Dechter, R. 1999 Branch and bound with mini-bucket heuristics. In Proc. Int.
Joint Conf. on Articial Intelligence (IJCAI99), pp. 426433.
Kessler, M. 1963 Bibliographic coupling between scientic papers. Am. Documentat. 14, 10
25.
REFERENCES 267

Killworth, P. and Bernard, H. 1978 Reverse small world experiment. Social Networks 1, 159.
Kira, K. and Rendell, L. A. 1992 A practical approach to feature selection. In Proc. 9th Int.
Conf. on Machine Learning, pp. 249256. San Francisco, CA: Morgan Kaufmann.
Kittler, J. 1986 Feature selection and extraction. In Handbook of Pattern Recognition and Image
Processing (ed. T. Y. Young and K. S. Fu), ch. 3. Academic.
Kleinberg, J. 1998 Authoritative sources in a hyperlinked environment. In Proc. 9th Ann. ACM
SIAM Symp. on Discrete Algorithms, pp. 668677. New York: ACM Press. (A preliminary
version of this paper appeared as IBM Research Report RJ 10076, May 1997.)
Kleinberg, J. 1999 Hubs, authorities, and communities. ACM Comput. Surv. 31, 5.
Kleinberg, J. 2000a Navigation in a small world. Nature 406, 845.
Kleinberg, J. 2000b The small-world phenomenon: an algorithmic perspective. In Proc. 32nd
ACM Symp. on the Theory of Computing.
Kleinberg, J. 2001 Small-world phenomena and the dynamic of information. Advances in
Neural Information Processing Systems (NIPS), vol. 14. Cambridge, MA: MIT Press.
Kleinberg, J. and Lawrence, S. 2001 The structure of the Web. Science 294, 18491850.
Kleinberg, J., Kumar, R., Raghavan, P., Rajagopalan, S. and Tomkins, A. 1999 The Web as
a graph: measurements, models, and methods. In Proc. Int. Conf. on Combinatorics and
Computing. Lecture notes in Computer Science, vol. 1627. Springer.
Kohavi, R. and John, G. 1997 Wrappers for feature subset selection. Artif. Intel. 97, 273324.
Koller, D. and Sahami, M. 1997 Hierarchically classifying documents using very few words.
In Proc. 14th Int. Conf. on Machine Learning (ICML-97), pp. 170178. San Francisco, CA:
Morgan Kaufmann.
Koller, D. and Sahami, N. 1996 Toward optimal feature selection. In Proc. 13th Int. Conf. on
Machine Learning, pp. 284292.
Korte, C. and Milgram, S. 1978 Acquaintance networks between racial groups: application of
the small world method. J. Pers. Social Psych. 15, 101.
Koster, M. 1995 Robots in the Web: threat or treat? ConneXions 9(4).
Krishnamurthy, B., Mogul, J. C. and Kristol, D. M. 1999 Key differences between HTTP/1.0
and HTTP/1.1. In Proc. 8th Int. World-Wide Web Conf. Elsevier.
Kruger, A., Giles, C. L., Coetzee, F., Glover, E. J., Flake, G. W., Lawrence, S. and Omlin,
C. W. 2000 DEADLINER: building a new niche search engine. In Proc. 2000 ACMCIKM
International Conf. on Information and Knowledge Management (CIKM-00) (ed. A. Agah,
J. Callan and E. Rundensteiner), pp. 272281. New York: ACM Press.
Kullback, S. 1968 Information theory and statistics. New York: Dover.
Kumar, S. R., Raghavan, P., Rajagopalan, S. and Tomkins, A. 1999a Extracting large-scale
knowledge bases from the Web. Proc. 25th VLDB Conf. VLDB J., pp. 639650.
Kumar, S. R., Raghavan, P., Rajagopalan, S. and Tomkins, A. 1999b Trawling the Web for
emerging cyber communities. In Proc. 8th World Wide Web Conf. Comp. Networks 31,
1116.
Kumar, S. R., Raghavan, P., Rajagopalan, S., Sivakumar, D., Tomkins, A. and Upfal, E. 2000
Stochastic models for the Web graph. In Proc. 41st IEEE Ann. Symp. on the Foundations of
Computer Science, pp. 5765.
Kushmerick, N., Weld, D. S. and Doorenbos, R. B. 1997 Wrapper induction for information
extraction. In Proc. Int. Joint Conf. on Articial Intelligence (IJCAI), pp. 729737.
Lafferty, J., McCallum, A. and Pereira, F. 2001 Conditional random elds: probabilistic models
for segmenting and labeling sequence data. In Proc. 18th Int. Conf. on Machine Learning,
pp. 282289. San Francisco, CA: Morgan Kaufmann.
268 REFERENCES

Lam, W. and Ho, C.Y. 1998 Using a generalized instance set for automatic text categorization. In
Proc. SIGIR-98, 21st ACM Int. Conf. on Research and Development in Information Retrieval
(ed. W. B. Croft, A. Moffat, C. J. van Rijsbergen, R. Wilkinson and J. Zobel), pp. 8189.
New York: ACM Press.
Lang, K. 1995 Newsweeder: Learning to lter news Proc. 12th Int. Conf. on Machine Learning
(ed. A. Prieditis and S. J. Russell), pp. 331339. San Francisco, CA: Morgan Kaufmann.
Langley, P. 1994 Selection of relevant features in machine learning. In Proc. AAAI Fall Symp.
on Relevance, pp. 140144.
Lau, T. and Horvitz, E. 1999 Patterns of search: analyzing and modeling Web query renement.
In Proc. 7th Int. Conf. on User Modeling, pp. 119128. Springer.
Lauritzen, S. L. 1996 Graphical Models. Oxford University Press.
Lauritzen, S. L. and Spiegelhalter, D. J. 1988 Local computations with probabilities on graphical
structures and their application to expert systems. J. R. Statist. Soc. B 50, 157224.
Lawrence, S. 2001 Online or invisible? Nature 411, 521.
Lawrence, S. and Giles, C. L. 1998a Context and page analysis for improved Web search. IEEE
Internet Computing 2, 3846.
Lawrence, S. and Giles, C. L. 1998b Searching the World Wide Web. Science 280, 98100.
Lawrence, S. and Giles, C. L. 1999a Acccessibility of information on the Web. Nature 400,
107109.
Lawrence, S., Giles, C. L. and Bollacker, K. 1999 Digital libraries and autonomous citation
indexing. IEEE Computer 32, 6771.
Leek, T. R. 1997 Information extraction using hidden Markov models. Masters thesis, Uni-
versity of California, San Diego.
Lempel, R. and Moran, S. 2001 SALSA: the stochastic approach for link-structure analysis.
ACM Trans. Informat. Syst. 19, 131160.
Letsche, T. A. and Berry, M. W. 1997 Large-scale information retrieval with latent semantic
indexing. Information Sciences 100, 105137.
Lewis, D. D. 1992An evaluation of phrasal and clustered representations on a text categorization
task. In Proc. 15th Ann. Int. ACM SIGIR Conf. on Research and Development in Information
Retrieval, pp. 3750. New York: ACM Press.
Lewis, D. D. 1997 Reuters-21578 text categorization test collection. (Documentation and data
available at http://www.daviddlewis.com/resources/testcollections/reuters21578/.)
Lewis, D. D. 1998 Naive Bayes at forty: the independence assumption in information retrieval.
In Proc. 10th Eur. Conf. on Machine Learning, pp. 415. Springer.
Lewis, D. D. and Catlett, J. 1994 Heterogeneous uncertainty sampling for supervised learning.
In Proc. ICML-94, 11th Int. Conf. on Machine Learning (ed. W. W. Cohen and H. Hirsh),
pp. 148156. San Francisco, CA: Morgan Kaufmann.
Lewis, D. D. and Gale, W. A. 1994 A sequential algorithm for training text classiers. In Proc.
17th Ann. Int. ACM SIGIR Conf. on Research and Development in Information Retrieval,
pp. 312. Springer.
Lewis, D. D. and Ringuette, M. 1994 Comparison of two learning algorithms for text categoriza-
tion. In Proc. 3rd Ann. Symp. on Document Analysis and Information Retreval, pp. 8193.
Li, S., Montgomery, A., Srinivasan, K. and Liechty, J. L. 2002 Predicting online purchase
conversion using Web path analysis. Graduate School of Industrial Administration, Carnegie
Mellon University, Pittsburgh, PA. (Available at http://www.andrew.cmu.edu/alm3/papers/
purchase%20conversion.pdf.)
Li, W. 1992 Random texts exhibit Zipfs-law-like word frequency dsitribution. IEEE Trans.
Inform. Theory 38, 18421845.
REFERENCES 269

Lieberman, H. 1995 Letizia: an agent that assists Web browsing. In Proc. 14th Int. Joint Conf. on
Articial Intelligence (IJCAI-95) (ed. C. S. Mellish), pp. 924929. San Mateo, CA: Morgan
Kaufmann.
Little, R. J. A. and Rubin, D. B. 1987 Statistical Analysis with Missing Data. John Wiley &
Sons, Ltd/Inc.
Liu, H. and Motoda, H. 1998 Feature Selection for Knowledge Discovery and Data Mining.
Kluwer Academic.
Lovins, J. B. 1968 Development of a stemming algorithm. Mech. Transl. Comput. Linguistics
11, 2231.
McCallum, A. and Nigam, K. 1998 A comparison of event models for naive Bayes text classi-
cation. AAAI/ICML-98 Workshop on Learning for Text Categorization, pp. 4148. Menlo
Park, CA: AAAI Press.
McCallum A., Freitag, D. and Pereira, F. 2000a Maximum entropy Markov models for informa-
tion extraction and segmentation. In Proc. 17th Int. Conf. on Machine Learning, pp. 591598.
San Francisco, CA: Morgan Kaufmann.
McCallum, A., Nigam, K. and Ungar, L. H. 2000b Efcient clustering of high-dimensional
data sets with application to reference matching. In Proc. 6th ACM SIGKDD Int. Conf. on
Knowledge Discovery and Data Mining, pp. 169178. New York: ACM Press.
McCallum, A. K., Nigam, K., Rennie, J. and Seymore, K. 2000c Automating the construction
of Internet portals with machine learning. Information Retrieval 3, 127163.
McCann, K., Hastings, A. and Huxel, G. R. 1998 Weak trophic interactions and the balance of
nature. Nature 395, 794798.
McClelland, J. L. and Rumelhart, D. E. 1986 Parallel Distributed Processing: Explorations in
the Microstructure of Cognition. Cambridge, MA: MIT Press.
McEliece, R. J. 1977 The Theory of Information and Coding. Reading, MA: Addison-Wesley.
McEliece, R. J. and Yildirim, M. 2002 Belief propagation on partially ordered sets. In Math-
ematical Systems Theory in Biology, Communications, and Finance (ed. D. Gilliam and J.
Rosenthal). Institute for Mathematics and its Applications, University of Minnesota.
McEliece, R. J., MacKay, D. J. C. and Cheng, J. F. 1997 Turbo decoding as an instance of
Pearls belief propagation algorithm. IEEE J. Select. Areas Commun. 16, 140152.
MacKay, D. J. C. and Peto, L. C. B. 1995a A hierarchical Dirichlet language model. Natural
Language Engng 1, 119.
McLachlan, G. and Peel, D. 2000 Finite Mixture Models. John Wiley & Sons, Ltd/Inc.
Mahmoud, H. M. and Smythe, R. T. 1995 A survey of recursive trees. Theory Prob. Math.
Statist. 51, 127.
Manber, U. and Myers, G. 1990 Sufx arrays: a new method for on-line string searches. In
Proc. 1st Ann. ACMSIAM Symp. on Discrete Algorithms, pp. 319327. Philadelphia, PA:
Society for Industrial and Applied Mathematics.
Mandelbrot, B. 1977 Fractals: Form, Chance, and Dimension. New York: Freeman.
Marchiori, M. 1997 The quest for correct information on the Web: hyper search engines. In
Proc. 6th Int. World-Wide Web Conf., Santa Clara, CA. Comp. Networks 29, 12251235.
Mark, E. F. 1988 Searching for information in a hypertext medical handbook. Commun ACM
31, 880886.
Maron, M. E. 1961 Automatic indexing: an experimental inquiry. J. ACM 8, 404417.
Maslov, S. and Sneppen, K. 2002 Specicity and stability in topology of protein networks.
Science 296, 910913.
Melnik, S., Raghavan, S., Yang, B. and Garcia-Molina, H. 2001 Building a distributed full-text
index for the Web. ACM Trans. Informat. Syst. 19, 217241.
Mena, J. 1999 Data Mining your Website. Boston, MA: Digital Press.
270 REFERENCES

Menczer, F. 1997 ARACHNID: adaptive retrieval agents choosing heuristic neighborhoods for
information discovery. In Proc. 14th Int. Conf. on Machine Learning, pp. 227235. San
Francisco, CA: Morgan Kaufmann.
Menczer, F. and Belew, R. K. 2000 Adaptive retrieval agents: internalizing local context and
scaling up to the Web. Machine Learning 39, 203242.
Mihail. M. and Papadimitriou, C. H. 2002 On the eigenvalue power law. In Randomization and
Approximation Techniques, Proc. 6th Int. Workshop, RANDOM 2002, Cambridge, MA, USA,
1315 September 2002 (ed. J. D. P. Rolim and S. P. Vadhan). Lecture Notes in Computer
Science, vol. 2483, pp. 254262. Springer.
Milgram, S. 1967 The small world problem. Psychology Today 1, 61.
Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N., Chklovskii, D. and Alon, U. 2002 Network
motifs: simple building blocks of complex networks. Science 298, 824827.
Mitchell, T. 1997 Machine Learning. McGraw-Hill.
Mitzenmacher, M. 2002 A brief history of generative models for power law and lognormal
distributions. Technical Report, Harvard University, Cambridge, MA.
Moffat, A. and Zobel, J. 1996 Self-indexing inverted les for fast text retrieval. ACM Trans.
Informat. Syst. 14, 349379.
Montgomery, A. L. 2001 Applying quantitative marketing techniques to the Internet. Interfaces
30, 90108.
Mooney, R. J. and Roy, L. 2000 Content-based book recommending using learning for text
categorization. In Proc. 5th ACM Conf. on Digital Libraries, pp. 195204. New York: ACM
Press.
Mori, S., Suen, C. andYamamoto, K. 1992 Historical review of OCR research and development.
Proc. IEEE 80, 10291058.
Moura, E. S., Navarro, G. and Ziviani, N. 1997 Indexing compressed text. In Proc. 4th South
American Workshop on String Processing (ed. R. Baeza-Yates), International Informatics
Series, pp. 95111. Ottawa: Carleton University Press.
Najork, M. and Wiener, J. 2001 Breadth-rst search crawling yields high-quality pages. In
Proc. 10th Int. World Wide Web Conf., pp. 114118. Elsevier.
Neal, R. M. 1992 Connectionist learning of belief networks. Artif. Intel. 56, 71113.
Neville-Manning, C. and Reed, T. 1996 A PostScript to plain text converter. Technical report.
(Available at http://www.nzdl.org/html/prescript.html.)
Newman, M. E. J., Moore, C. and Watts, D. J. 2000 Mean-eld solution of the small-world
network model. Phys. Rev. Lett. 84, 32013204.
Ng, A. Y. and Jordan, M. I. 2002 On discriminative vs generative classiers: a comparison of
logistic regression and Naive Bayes. Advances in Neural Information Processing Systems
14. Proc. 2001 Neural Information Processing Systems (NIPS) Conference. MIT Press.
Ng, A. Y., Zheng, A. X. and Jordan, M. I. 2001 Stable algorithms for link analysis. In Proc.
24th Ann. Int. ACM SIGIR Conf. on Research and Development in Information Retrieval,
pp. 258266. New York: ACM Press.
Nigam, K. and Ghani, R. 2000 Analyzing the effectiveness and applicability of co-training. In
Proc. 2000 ACMCIKM Int. Conf. on Information and Knowledge Management (CIKM-00)
(ed. A. Agah, J. Callan and E. Rundensteiner), pp. 8693. New York: ACM Press.
Nigam, K., McCallum A., Thrun, S. and Mitchell, T. 2000 Text classication from labeled and
unlabeled documents using EM. Machine Learning 39, 103134.
Nothdurft, H. 2000 Salience from feature contrast: additivity across dimensions. Vision Res.
40, 11831201.
Olshausen, B. A., Anderson, C. H. and Essen, D. C. V. 1993 A neurobiological model of
visual attention and invariant pattern recognition based on dynamic routing of information.
J. Neurosci. 13, 47004719.
REFERENCES 271

Oltvai, Z. N. and Barabsi, A.-L. 2002 Lifes complexity pyramid. Science 298, 763764.
ONeill, E. T., McClain P. D. and Lavoie, B. F. 1997 A methodology for sampling the World
Wide Web Annual Review of OCLC Research. (Available at http://www.oclc.org/research/
publications/arr/ 1997/oneill/o%27neillar98%0213.htm.)
Page, L., Brin, S., Motwani, R. and Winograd, T. 1998 The PageRank citation ranking: bring-
ing order to the Web. Technical report, Stanford University. (Available at http://www-
db.stanford.edu/backrub/pageranksub.ps.)
Paine, R. T. 1992 Food-web analysis through eld measurements of per capita interaction
strength. Nature 355, 7375.
Pandurangan, G., Raghavan, P. and Upfal, E. 2002 Using PageRank to characterize Web struc-
ture. In Proc. 8th Ann. Int. Computing and Combinatorics Conf. (COCOON). Lecture Notes
in Computer Science, vol. 2387, p. 330. Springer.
Papineni, K. 2001 Why inverse document frequency? Proc. North American Association for
Computational Linguistics, pp. 2532.
Passerini, A., Pontil, M. and Frasconi, P. 2002 From margins to probabilities in multiclass
learning problems. In Proc. 15th Eur. Conf. on Articial Intelligence (ed. F. van Harmelen).
Frontiers in Articial Intelligence and Applications Series. Amsterdam: IOS Press.
Pazzani, M. 1996 Searching for dependencies in Bayesian classiers. In Proc. 5th Int. Workshop
on Articial Intelligence and Statistics, pp. 239248. Springer.
Pearl, J. 1988 Probabilistic Reasoning in Intelligent Systems. San Mateo, CA: Morgan Kauf-
mann.
Pennock, D. M., Flake, G. W., Lawrence, S., Glover, E. J. and Giles, C. L. 2002 Winners
dont take all: characterizing the competition for links on the Web. Proc. Natl Acad. Sci. 99,
52075211.
Perline, R. 1996 Zipfs law, the central limit theorem, and the random division of the unit
interval. Phys. Rev. E 54, 220223.
Pew Internet Project Report 2002 Search engines. (Available at http://www.pewinternet.org/
reports/toc.asp?Report=64.)
Phadke, A. G. and Thorp, J. S. 1988 Computer Relaying for Power Systems. John Wiley &
Sons, Ltd/Inc.
Philips, T. K., Towsley, D. F. and Wolf, J. K. 1990 On the diameter of a class of random graphs.
IEEE Trans. Inform. Theory 36, 285288.
Pimm, S. L., Lawton, J. H. and Cohen, J. E. 1991 Food web patterns and their consequences.
Nature 350, 669674.
Pittel, B. 1994 Note on the heights of random recursive trees and random m-ary search trees.
Random Struct. Algorithms 5, 337347.
Platt, J. 1999 Fast training of support vector machines using sequential minimal optimization.
In Advances in Kernel Methods Support Vector Learning (ed. B. Schlkopf, C. J. C. Burges
and A. J. Smola,), pp. 185208. Cambridge, MA: MIT Press.
Popescul, A., Ungar, L. H., Pennock, D. M. and Lawrence, S. 2001 Probabilistic models for
unied collaborative and content-based recommendation in sparse-data environments. In
Proc. 17th Int. Conf. on Uncertainty in Articial Intelligence, pp. 437444. San Francisco,
CA: Morgan Kaufmann.
Porter, M. 1980 An algorithm for sufx stripping. Program 14, 130137.
Quinlan, J. R. 1986 Induction of decision trees. Machine Learning 1, 81106.
Quinlan, J. R. 1990 Learning logical denitions from relations. Machine Learning 5, 239266.
Raei, D. and Mendelzon, A. 2000 What is this page known for? Computing Web page repu-
tations. In Proc. 9th World Wide Web Conf.
Raggett, D., Hors, A. L. and Jacobs, I. (eds) 1999 HTML 4.01 Specication. W3 Consortium
Recommendation. (Available at http://www.w3.org/TR/html4/.)
272 REFERENCES

Raskinis, I. M. G. and Ganascia, J. 1996 Text categorization: a symbolic approach. In Proc.

5th Ann. Symp. on Document Analysis and Information Retrieval. New York: ACM Press.
Redner, S. 1998 How popular is your paper? An empirical study of the citation distribution.
Eur. Phys. J. B 4, 131134.
Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P. and Riedl, J. 1994 GroupLens: an open
architecture for collaborative ltering of netnews. In Proc. 9th ACM Conf. on Computer-
Supported Cooperative Work, pp. 175186. New York: ACM Press.
Ripeanu, M., Foster, I. and Iamnitchi, A. 2002 Mapping the Gnutella network: properties of
large-scale peer-to-peer systems and implications for system design. IEEE Internet Comput.
J. 6, 99100.
Roberts, M. J. and Mahesh, S. M. 1999 Hotmail. Technical report, Harvard University, Cam-
bridge, MA. Case 899-185, Harvard Business School Publishing.
Robertson, S. E. 1977 The probability ranking principle in IR. J. Doc. 33, 294304. (Also
reprinted in Jones and Willett (1997), pp. 281286.)
Robertson, S. E. and Sprck Jones, K. 1976 Relevance weighting of search terms. J. Am. Soc.
Informat. Sci. 27, 129146.
Robertson, S. E. and Walker, S. 1994 Some simple effective approximations to the 2-Poisson
model for probabilistic weighted retrieval. In Proc. 17th Ann. Int. ACM SIGIR Conf. on
Research and Development in Information Retrieval, pp. 232241. Springer.
Rosenblatt, F. 1958 The perceptron: a probabilistic model for information storage and organi-
zation in the brain. Psychol. Rev. 65, 386408.
Ross, S. M. 2002 Probability Models for Computer Science. San Diego, CA: Adademic Press.
Russell, S. and Norvig, P. 1995 Articial Intelligence: A Modern Approach. Prentice Hall.
Sahami, M., Dumais, S., Heckerman, D. and Horvitz, E. 1998 A Bayesian approach to ltering
junk e-mail. AAAI-98 Workshop on Learning for Text Categorization, pp. 5562.
Salton, G. 1971 The SMART Retrieval System: Experiments in Automatic Document Processing.
Englewood Cliffs, NJ: Prentice Hall.
Salton, G. and McGill, M. J. 1983 Introduction to modern information retrieval. McGraw-Hill.
Salton, G., Fox, E. A. and Wu, H. 1983 Extended Boolean information retrieval. Commun.
ACM 26, 10221036.
Sarukkai, R. R. 2000 Link prediction and path analysis using Markov chains. Comp. Networks
33, 377386.
Sarwar, B. M., Karypis, G., Konstan, J. A. and Riedl, J. T. 2000 Analysis of recommender
algorithms for e-commerce. In Proc. 2nd ACM Conf. on Electronic Commerce, pp. 158
167. New York: ACM Press.
Saul, L. and Pereira, F. 1997 Aggregate and mixed-order Markov models for statistical language
processing. In Proc. 2nd Conf. on Empirical Methods in Natural Language Processing (ed.
C. Cardie and R. Weischedel), pp. 8189. Somerset, NJ: Association for Computational
Linguistics.
Saul, L. K. and Jordan, M. I. 1996 Exploiting tractable substructures in intractable networks.
In Advances in Neural Information Processing Systems (ed. D. S. Touretzky, M. C. Mozer
and M. E Hasselmo), vol. 8, pp. 486492.Cambridge, MA: MIT Press.
Savage, L. J. 1972 The foundations of statistics. New York: Dover.
Schafer, J. B., Konstan, J. A. and Riedl, J. 2001 E-commerce recommendation applications. J.
Data Mining Knowl. Discovery 5, 115153.
Schapire, R. E. and Freund,Y. 2000 Boostexter: a boosting-based system for text categorization.
Machine Learning 39, 135168.
Schoelkopf, B. and Smola, A. 2002 Learning with Kernels. Cambridge, MA: MIT Press.
Sebastiani, F. 2002 Machine learning in automated text categorization. ACM Comput. Surv. 34,
147.
REFERENCES 273

Sen, R. and Hansen, M. H. 2003 Predicting a Web users next request based on log data. J.
Computat. Graph. Stat. (In the press.)
Seneta, E. 1981 Nonnegative Matrices and Markov Chains. Springer.
Shachter, R. D. 1988 Probabilistic inference and inuence diagrams. Oper. Res. 36, 589604.
Shachter, R. D., Anderson, S. K. and Szolovits, P. 1994 Global conditioning for probabilistic
inference in belief networks. In Proc. Conf. on Uncertainty in AI, pp. 514522. San Francisco,
CA: Morgan Kaufmann.
Shahabi, C., Banaei-Kashani, F. and Faruque, J. 2001 A framework for efcient and anonymous
Web usage mining based on client-side tracking. In Proceedings of WEBKDD 2001. Lecture
Notes in Articial Intelligence, vol. 2356, pp. 113144. Springer.
Shannon, C. E. 1948a A mathematical theory of communication. Bell Syst. Tech. J. 27, 379423.
Shannon, C. E. 1948b A mathematical theory of communication. Bell Syst. Tech. J. 27, 623656.
Shardanand, U. and Maes, P. 1995 Social information ltering: algorithms for automating
word of mouth. In Proc. Conf. on Human Factors in Computing Systems, pp. 210217.
Shore, J. E. and Johnson, R. W. 1980 Axiomatic derivation of the principle of maximum entropy
and the principle of minimum cross-entropy. IEEE Trans. Inform. Theory 26, 2637.
Silverstein, C., Henzinger, M., Marais, H. and Moricz, M. 1998 Analysis of a very large
AltaVista query log. Technical Note 1998-14, Digital System Research Center, Palo Alto,
CA.
Slonim, N. and Tishby, N. 2000 Document clustering using word clusters via the information
bottleneck method. In Proc. 23rd Int. Conf. on Research and Development in Information
Retrieval, pp. 208215. New York: ACM Press.
Slonim, N., Friedman, N. and Tishby, N. 2002 Unsupervised document classication using
sequential information maximization. In Proc. 25th Int. Conf. on Research and Development
in Information Retrieval, pp. 208215. New York: ACM Press.
Small, H. 1973 Co-citation in the scientic literature: a new measure of the relationship between
two documents. J. Am. Soc. Inf. Sci. 24, 265269.
Smyth, P., Heckerman, D. and Jordan, M. I. 1997 Probabilistic independence networks for
hidden Markov probability models. Neural Comp. 9, 227267.
Soderland, S. 1999 Learning information extraction rules for semi-structured and free text.
Machine Learning 34, 233272.
Sperberg-McQueen, C. and Burnard, L. (eds) 2002 TEI P4: Guidelines for Electronic
Text Encoding and Interchange. Text Encoding Initiative Consortium. (Available at
http://www.tei-c.org/.)
Spink, A., Jansen, B. J., Wolfram, D. and Saracevic, T. 2002 From e-sex to e-commerce: Web
search changes. IEEE Computer 35, 107109.
Stallman, R. 1997 The right to read. Commun. ACM 40, 8587.
Sutton, R. S. and Barto, A. G. 1998 Reinforcement Learning: An Introduction. Cambridge,
MA: MIT Press.
Tan, P. and Kumar, V. 2002 Discovery of Web robot sessions based on their navigational
patterns. Data Mining Knowl. Discov. 6, 935.
Tantrum, J., Murua, A. and Stuetzle, W. 2002 Hierarchical model-based clustering of large
datasets through fractionation and refractionation. In Proc. 8th ACM SIGKDD Int. Conf. on
Knowledge Discovery and Data Mining. New York: ACM Press.
Taskar, B., Abbeel, P. and Koller, D. 2002 Discriminative probabilistic models for relational
data. In Proc. 18th Conf. on Uncertainty in Articial Intelligence. San Francisco, CA: Morgan
Kaufmann.
Tauscher, L. and Greenberg, S. 1997 Revisitation patterns in World Wide Web navigation. In
Proc. Conf. on Human Factors in Computing Systems CHI97, pp. 97137. New York: ACM
Press.
274 REFERENCES

Tedeschi, B. 2000 Easier to use sites would help e-tailers close more sales. New York Times,
12 June 2000.
Tishby, N., Pereira, F. and Bialek, W. 1999 The information bottleneck method. In Proc. 37th
Ann. Allerton Conf. on Communication, Control, and Computing (ed. B. Hajek and R. S.
Sreenivas), pp. 368377.
Titterington, D. M., Smith, A. F. M. and Makov, U. E. 1985 Statistical Analysis of Finite Mixture
Distributions. John Wiley & Sons, Ltd/Inc.
Travers, J. and Milgram, S. 1969 An experimental study of the smal world problem. Sociometry
32, 425.
Ungar, L. H. and Foster, D. P. 1998 Clustering methods for collaborative ltering. In Proc.
Workshop on Recommendation Systems at the 15th National Conf. on Articial Intelligence.
Menlo Park, CA: AAAI Press.
Vapnik, V. N. 1982 Estimation of Dependences Based on Empirical Data. Springer.
Vapnik, V. N. 1995 The Nature of Statistical Learning Theory. Springer.
Vapnik, V. N. 1998 Statistical Learning Theory. John Wiley & Sons, Ltd/Inc.
Viterbi, A. J. 1967 Error bounds for convolutional codes and an asymptotically optimum decod-
ing algorithm. IEEE Trans. Inform. Theory 13, 260269.
Walker, J. 2002 Links and power: the political economy of linking on the Web. In Proc. 13th
Conf. on Hypertext and Hypermedia, pp. 7273. New York: ACM Press.
Wall, L., Christiansen, T. and Schwartz RL. 1996 Programming Perl, 2nd edn. Cambridge,
MA: OReilly & Associates.
Wasserman, S. and Faust, K. 1994 Social Network Analysis. Cambridge University Press.
Watts, D. J. and Strogatz, S. H. 1998 Collective dynamics of small-world networks. Nature
393, 440442.
Watts, D. J., Dodds, P. S. and Newman, M. E. J. 2002 Identity and search in social networks.
Science 296, 13021305.
Weiss, S. M., Apte, C., Damerau, F. J., Johnson, D. E., Oles, F. J., Goetz, T. and Hampp, T.
1999 Maximizing text-mining performance. IEEE Intell. Syst. 14, 6369.
Weiss, Y. 2000 Correctness of local probability propagation in graphical models with loops.
Neural Comp. 12, 141.
White, H. 1970 Search parameters for the small world problem. Social Forces 49, 259.
Whittaker, J. 1990 Graphical Models in Applied Multivariate Statistics. John Wiley & Sons,
Ltd/Inc.
Wiener, E. D., Pedersen, J. O. and Weigend, A. S. 1995 A neural network approach to topic spot-
ting. In Proc. SDAIR-95, 4th Ann. Symp. on Document Analysis and Information Retrieval,
Las Vegas, NV, pp. 317332.
Witten, I. H., Moffat, A. and Bell, T. C. 1999 Managing Gigabytes: Compressing and Indexing
Documents and Images, 2nd edn. San Francisco, CA: Morgan Kaufmann.
Witten, I. H., Neville-Manning, C. and Cunningham, S. J. 1996 Building a digital library for
computer science research: technical issues. In Proc. Australasian Computer Science Conf.,
Melbourne, Australia.
Wolf, J., Squillante, M., Yu, P., Sethuraman, J. and Ozsen, L. 2002 Optimal crawling strategies
for Web search engines. In Proc. 11th Int. World Wide Web Conf., pp. 136147.
Xie, Y. and OHallaron, D. 2002 Locality in search engine queries and its implications for
caching. In Proc. IEEE Infocom 2002, pp. 12381247. Piscataway, NJ: IEEE Press.
Yang, Y. 1999 An evaluation of statistical approaches to text categorization. Information
Retrieval 1, 6990.
Yang, Y. and Liu, X. 1999 A re-examination of text categorization methods In Proc. SIGIR-99,
22nd ACM Int. Conf. on Research and Development in Information Retrieval (ed. M. A.
Hearst, F. Gey and R. Tong), pp. 4249. New York: ACM Press.
REFERENCES 275

Yedidia, J., Freeman, W. T. and Weiss, Y. 2000 Generalized belief propagation. Neural Comp.
12, 141.
York, J. 1992 Use of the Gibbs sampler in expert systems. Artif. Intell. 56, 115130.
Zamir, O. and Etzioni, O. 1998 Web document clustering: a feasibility demonstration. In Proc.
21st Int. Conf. on Research and Development in Information Retrieval (SIGIR), pp. 4654.
New York: ACM Press.
Zelikovitz, S. and Hirsh, H. 2001 Using LSI for text classication in the presence of background
text. In Proc. 10th Int. ASM Conf. on Information and Knowledge Management, pp. 113118.
New York: ACM Press.
Zhang, T. and Iyengar, V. S. 2002 Recommender systems using linear classiers. J. Machine
Learn. Res. 2, 313334.
Zhang, T. and Oles, F. J. 2000 A probability analysis on the value of unlabeled data for classica-
tion problems. In Proc. 17th Int. Conf. on Machine Learning, Stanford, CA, pp. 11911198.
Zhu, X., Yu, J. and Doyle, J. 2001 Heavy tails, generalized coding, and optimal Web layout. In
Proc. 2001 IEEE INFOCOM Conf., vol. 3, pp. 16171626. Piscataway, NJ: IEEE Press.
Zukerman, I., Albrecht, D. W. and Nicholson, A. E. 1999 Predicting users requests on the
WWW. In Proc. UM99: 7th Int. Conf. on User Modeling, pp. 275284. Springer.
This Page Intentionally Left Blank
Index

absolute URI 35 BackusNaur Form (BNF) 35

access protocol 29 bag of words representation 83, 95,
accuracy 86, 105 106
AdaBoost 108 base subgraph 130
agent log 41 Bayes decision rule for classication
aggregate Markov model 91 87
aging of documents 160, 1617 Bayes optimal separation 87
aliases 49 Bayes rule 3, 18
aliasing 163 Bayes theorem 2, 11
AltaVista 58, 202, 204, 205 Bayesian interpretation of probability
America Online (AOL) 202 2
American Registry for Internet Bayesian (belief) networks 1315,
Numbers (ARIN) 160 116
anchor based prediction 154 Bayesian parameter estimation for
anchors 32 Markov models 1879
Arachnid agents 157 belief propagation (inference) 16, 92
aspect model 91 Bellman optimality principle 156
attributes 31, 94 Bernoulli distribution 96
authority 35, 131 Bernoulli likelihood model 96
automated logging 172 Bernoulli model for the Naive Bayes
automated recommender systems classier 87
21224 best-rst search strategy 149
evaluating 2145 Beta distribution 7, 8, 240
model-based collaborative bibliographic coupling matrix 145
ltering 21823 binary independence retrieval (BIR)
nearest-neighbor collaborative 87
ltering 2158 Binomial distribution 22, 2378, 240
biological networks 54
bow-tie structure 589
Modeling the Internet and the Web
breadth-rst search (BFS) 47
P. Baldi, P. Frasconi and P. Smyth breakeven point 105
2003 P. Baldi, P. Frasconi and P. Smyth broad-scale networks 57
ISBN: 0-470-84906-1
browser 29

277
278 INDEX

browsing behavior, client-side studies connectivity 236

17983 power-law 535
19951997 1801 content-based ranking 826
2002 1813 context graphs 1545
probabilistic models 184201 continuous (discrete power-law)
video-based analysis of Web distributions 22, 23, 23840
usage 183 co-occurrences 112
bucket 789 cookies 1767
business networks 54 coordination 159
copy models 667
caching 1746 cosine coefcient 83
canonicalization 49 co-training 1145
canopies 119 covariance 237
Central Limit Theorem 238 covariates 199
centroids 20 coverage 44, 459
chaining effect 118 crawler (spider; Web robot) 44
Chi square distribution 240 crawling
CiteSeer 121, 1578 basic 469
classication 1720 distributed 15860
classication rules, induction of 93 focused 1528
vs graph visiting 49
classication trees 19
relevance prediction 1524
classier 17
selective 14952
client-side data 1779
cross-entropy 84, 243, 244, 246
handling massive Web server logs
crossover 160
179
current search action 207
cliquishness 62
curse of dimensionality 99
clustering 202
customer data on the Web 212
for documents 1179
co-citation 141 data logging at Web servers 174
co-citation matrix 145 DEADLINER 158
coding theory 15 Debian 89
collective classication 116 decision-theoretic surng model
common gateway interface (CGI) 49 1989
completeness 33 decision tree approximation 14
compression techniques 7980 decision trees 93, 102
ComScore 1778, 212 degree distribution 58
ComScore Media Metrix 229 degrees of belief 2
conditional entropy 247 dendrogram 118
conditional independence 220 dependency networks 219
conditioning 2 depth of scoring functions 150
connement 15960 descriptive markup 30
confusion matrix 104 dimension-reduction techniques 119
connectionism 127 directed acyclic graph (DAG) 236
INDEX 279

Dirichlet distribution 7, 188, 23940, Excite search engine 202, 203, 204,
250 205
Dirichlet prior 7, 12, 96, 97, 189 expectation 237
discrete distributions 2378 expectation maximization (EM) 5, 16
discrete power-law (continuous) algorithm 1013, 21
distributions 22, 23, 23840 Naive Bayes and 112
discriminative (decision-boundary) exploratory browsing state 232
approach to classication 18 exponential distribution 22, 239, 240
disordered lattice approach 63 exponential family 2401
distributed search algorithms 6870 extensibility 33
document clustering 11620 eXtensible Markup Language (XML)
algorithm 1179 30
background and examples 1167 extreme value distribution 241
related approaches 11920
document identier (DID) 7880 FASTUS system 120
document similarity 834 feature selection 1024
Document Type Denition (DTD) 30 feature space 102
frameset 32 Fermis model 257, 61
of HTML 31 lters 102
nite-state Markov models 10, 14
strict 32
rewall 160
transitional 32
rst-in rst-out (FIFO) policy 47
domain name service (DNS) 38, 49
Fish algorithm 157
domain name system 379
FOIL 1156
dominant eigenvalue 129
frameset DTD 32
dominant eigenvector 129
Freenet 73
dynamic coordination 159
frequentist interpretation of
probability 2
e-commerce logs 212 freshness 167
eigengap 138
80/20 rule 20, 24 Gamma distribution 22, 239, 240
element name 31 Gaussian distribution 11, 22, 23,
element types 30 2389
email 11011 Gentoo 89
ltering 1718 geometric distribution 238240
product recommendations 2246 GET method 39, 43, 174
embarrassment level 169 Gibbs sampling 15
energy function 5 Gnutella 70, 73
entropy 243, 2445 good citizen search engines 173
error function 5 Google search engine 44, 53, 57, 225
error log 41 grace period 160
error rate 86 gradient descent 5
escape matrix 137 graph theory 2357
exchange 160 graph visiting vs crawling 49
280 INDEX

graphical models 1317 Informant, The 163

greedy agglomerative clustering 120 information bottleneck 120
greedy navigation algorithms 69 information extraction 1202
Gumbel distribution 241 information gain 1023
information retrieval (IR) 77
hard focusing 153 information theory 24351
hazard rate 161 informational goals 207, 208
HEAD method 41 Inquirus 46
Heaps law 245 intelligent crawling 157
Heaviside step function 164 Internet Corporation for Assigned
heterogeneity of behavior 200 Names and Numbers (ICANN) 38
hidden (invisible) Web 39 Internet Engineering Task Force 34,
hidden Markov models (HMMs) 14, 44
122 Internet Protocol (IP) 37
hierarchical clustering 117, 118 address space sampling 401
hierarchical model 200 addresses 37
Hypertext Induced Topic Selection Internet service provider (ISP) 175
(HITS) 1304 Internet tomography 74
stability of 1389 inverse document frequency (IDF) 84
Hotmail 2257 inverse Gaussian distribution 1989
hubness 131 inversion 78
hubsauthority bipartite graph inverted index 78
pattern 71 irreducible matrix 128
human-generated trafc 1734 ISO/OSI reference model 36
hyper information 127
hyperlinks, text analysis and 1146 Jensens inequality 246, 247
HyperText Markup Language junction tree 15
(HTML) 301 k nearest neighbours (k-NN) 93
body 32 classier 108
document structure 312 Kalman lter models 14
DTD of 31 KarhunenLove transform 242
header 32 KarushKuhnTucker condition 101
tokenization of 801 kernel approaches 19, 102
version information 32 Kleinbergs algorithm 1324
hypertext transfer protocol 389 K-means 12
methods in 39 algorithm 20, 21
clustering 117, 120
impact factor 1256 knowledge bases (KB) 106
incremental crawler 169 kth-order Markov model 190
indegree distribution 58 KullbackLeibler distance 84, 243,
index age of a document 167 244, 246
indexing 7780 KullbackLeibler divergence 84
inductive logic programming 93
inference 16, 92 Lagrange multiplier 11
INDEX 281

Lagrangian 6, 11 for page prediction 1846

Laplace smoothing 97 position-dependent 193
LASER 156 predicting page requests 18993
latent semantic analysis 8892 variable-order 193
latent semantic indexing (LSI) zero-order 9
8892, 112 markup 30
text documents and 89 maximum a posteriori (MAP) 16
probabilistic 8992 estimation 5, 6
lattice perturbation models 613 framework 12
law of total probability 2 optimization problem 8
learning directed graphical models parameter estimation 6
from data 1617 maximum-distance algorithm 118
learning effect 201 maximum likelihood 16, 187
lexical processing 806 equations 12
likelihood 3 estimation 5, 6, 1956
likelihood function 187 mean eld approximation 62
linear classiers 219 mean eld methods 15
link analysis 71, 12645 mean eld solution 62
early approaches 1268 mean posterior (MP) estimation 5, 9,
probabilistic 1402 16
limitations 1425 measures of performance 1046
stability 1389 Mercator 48
link farm 143 meta-items 217
links 32 metalanguage 30
Linux 89 meta-search engines 46
log les 414 metasymbols 31
log probability score 192 methods in HTTP 39
luminosity 126 microaveraging 106
minimum description length 104
machine learning 77, 93, 102 minimum-distance algorithm 118
macroaveraging 106 mixture coefcients 11
Markov blanket 1034 mixture models 1012, 14
Markov chain 7, 21, 578, 243 ModApte split 108, 112, 113
mixtures of 191 model-based clustering 21
Markov model 120 model-based collaborative ltering
aggregate 91 21823
Bayesian parameter estimation conditional distribution models
1879 2213
nite-state 10, 14 general concepts 2189
rst-order 1901 joint density modeling 21921
tting to observed page-request model-based combining of votes and
data 1867 content 2234
hidden (HMMs) 14, 122 Monte Carlo methods 15
kth-order 190 mortality force 161
282 INDEX

multinomial distribution 2378, 240 optimal separating hyperplane

multiple conditional model paradigm 99100
219 overall information 127
mutual information 243, 244, 2467 overlap analysis 45

Naive Bayes classier 18, 947, 102, page request prediction 199201
105, 106 PageRank 1348, 151
Naive Bayes model 21, 93, 96, 119, models 678
220 power law of 578
nearest-neighbor collaborative stability of 139
ltering 2158 panel data 212
basic principles 215 parent based prediction 1534
computational issues and ParetoZipf distribution (Zipf) 22, 80
clustering 218 partitioning 160
curse of dimensionality 2178 path 35
pattern recognition 102
dening weights 2167
Pearson correlation coefcient 104
negative binomial distribution 240
peer-to-peer networks 70, 73
network value 227
perceptrons 19, 98
networks and recommendations
period 128
2248
Perl programming language 40, 81
diffusion model 2268
PerronFrobenius theorem 12830
email-based product
Pew Internet Project Report 44
recommendations 2246
PHITS 116, 142
neural networks 14, 19, 93
PLSA 116
news ltering 11011 Poisson distribution 22, 96, 194, 238,
news stories, classication of 10710 240
NewsWeeder 11011 polynomial time algorithms 71
next search action 207 polysemy 88
Nielsen 177, 212 popularity 126, 1501
noisy OR models 14 Portable Document Format (PDF) 81
nonnegative matrices 12830 tokenization 81
nonstationary Markov chain 193 posterior distribution 247
normal distribution 240 PostScript 81
tokenization of 81
one-shot coding strategy 102 power-law connectivity 535
one-step ahead prediction 191 power-law distribution 227, 52, 53
one vs all coding strategy 102 origin of 257
ontology 106 power-law size 53
Open Directory Project (ODP) 44 precision 86, 105
Open System Interconnect (OSI) 36 predicates 157
optical character recognition (OCR) preferential attachment models 636
81 PreScript 81
optimal resource allocation 168 primitive matrix 129
INDEX 283

principal component analysis (PCA) recursive query 38

242 reference models 367
prior probability 2 referrer log 41
priors 2, 3, 4 referrer page 174
distribution 247 reinforcement learning 1557
probabilistic approach to relational Bayesian networks 116
classication 18 relational learning 1156
probabilistic decision trees 2213 relative entropy (cross-entropy;
probabilistic link analysis 1402 KullbackLeibler distance) 84,
probabilistic Markov network model 243, 244, 246
116 relative URI 36
probabilistic model-based clustering relevance 2501
117, 119, 120 Rnyis entropies 250
probabilistic ranking principle (PRP) Rseaux IP Europens (RIPE) 160
868 resolution 38
probabilistic retrieval 868 resource 29
probability 13 resource identiers 29, 31, 336
procedural markup 30 retrieval and evaluation measures
Prolog 115 856
protocols 29, 3641 retrieval status value (RSV) 88
prototypes 20 rich-get-richer model 634
proxy server 1746 Riemanns zeta function 23, 195
Pstotext 81 robot-generated trafc 173
purchase-oriented state 232 robot program 41
Robots Exclusion Protocol 48
quadratic programming 101
robustness 713
quality of ranking 44
ROC curves (Receiver Operating
quasi-random graphs 74
query 35 Characteristic) 105
run-length modeling 1934
radial basis function (RBF) kernel
108 SALSA (Stochastic Approach for
random effects model 200 Link Structure Analysis) 1402
random graphs 63, 2367 SavvySearch 46
random recursive trees 65 scale-free networks 57
random sparse graphs 74 scale-free properties (80/20 rule) 20,
random tree 65 24
random uniform graph (random scaling property of power laws 53
exponential graph) 72 Search Engine Persuasion (SEP)
recall 86, 105 1423
recency 445, 160, 180 search engines 449
measures of 167 empirical studies of search
synchronization policies and behavior 2026
1679 models for search strategies 2078
284 INDEX

querying, modeling and spreading activation mechanism 225

understanding 2018 stability 1389
search query logs of HITS 1389
general characteristics of 2035 of PageRank 139
power-law characteristics 2056 Standard Generalized Markup
second-generation Web searching Language (SGML) 301
tools 126 static coordination 159
security 73 stationary distribution of the Markov
seeds 46 chain 130
self-scaling 24 status code 41
self-similarity 24 stemming 82
semantic linkage 154 stop words, removal of 82
semi-Markov model 193 strict DTD 32
separating hyperplane 98 strongly connected component (SCC)
server 29 58
server access log 41 structural risk minimization 97
server log les 41 subgraph patterns and communities
server-side data 1747 701
server transfer log 41 supervised learning 20
session length modeling 1948 with unlabeled data 1114
Shannons information 2478 support vector classiers 97102
Shark algorithm 157 support vector machines (SVMs) 19,
sigmoidal belief networks 14 93, 97102
simple Markov chain models 14 classier 106
simulated annealing 5 transductive 1134
single-link clustering 118 support vectors 101
single-scale networks 57 surprise 249
singular value decomposition (SVD) computing 24950
88, 89, 2178, 2412 learning and 250
six degrees of separation 55, 69 symbolic rule learning 104
slack variables 101 synonymy 88
sliding-window based classication
121 tags 31
small-world networks 557, 65 target anchor 32
classes 57 TCP/IP 35, 367
social networks 54 telecommunication networks 54
soft focusing 154 tempered EM (TEM) 92
source anchor 32 term frequencies (TF) 84
spam 178, 145, 224, 2256 terms 78
sparseness 52, 83 text categorization 77, 93114
specicity 105 text conation 82
spherical K-means algorithm 119 Text REtrieval Conference (TREC)
spider 44 86
spider program 41 textual information 127
INDEX 285

TFIDF 118 Vivisimo search engine 117, 203, 205

weight 84 vocabulary 78
Tightly Knit Community (TKC) vocabulary reduction 82
Effect 141 vulnerability 713
time interval 207
tokenization 801 Web data and measurement 1729
topic drift problem 134 Web documents 302
topic locality 153 Web dynamics 1609
top-level domain (TLD) 38 Web graphs
topological sort 127 applications 6873
transductive support vector machines Internet and 513
1134 reason for studying 52
transitional DTD 32 Web-KB system 1201
Transmission Control Protocol (TCP) Web pages
37 analysis for purchase prediction
transportation networks 54 22832
traps 49 classication of 1067
triangulated moral graph 15 growth 601
Web robot 44
unconditional marginal dependence Web server logs, identifying
220 individual users from 1767
uniform random graphs 73 Web user agents 29
Uniform Resource Identier (URI) WebFountain 169
32, 336 Weblogs 145
Uniform Resource Locator (URL) WebWatcher 157
34, 489 Weibull distribution 166, 240
Uniform Resource Name (URN) 34 weighted graph 74
unsupervised learning 20 within-site lock-in effect 201
user agent 38 word-of-mouth concept 2246
user search action 207 wrappers 102
www.amazon.com 179, 211
variance 237 www.google.com 179
variable-order Markov model 193
variational methods 15 XHTML 32
vector-space model 823
visibility 126 zero-order Markov model 9
visible Web 46 Zipf distribution 22, 80
Viterbi algorithm 122 Zipfs Law 22, 245