Algorithmic Marketing Ai For Marketing Operations r1 3g

I LYA K AT S O V
INTRODUCTION TO
ALGORITHMIC MARKETING
Introduction to Algorithmic Marketing
by Ilya Katsov
Visit the book’s website at https://algorithmicweb.wordpress.com

for additional resources.
ISBN 978-0-692-98904-3
Copyright © 2018 Ilya Katsov

All rights reserved. The digital version of the book is available for
download on algorithmicweb.wordpress.com and can be shared only
through a link to the landing page. No part of this book may be
modified, reposted, emailed, or made available for download on other
sites or channels, without the prior written permission of the author.
v
“At a time when power is shifting to consumers, while brands and retailers
are grasping for fleeting moments of attention, everyone is competing on data
and the ability to leverage it at scale to target, acquire, and retain customers.
This book is a manual for doing just that. Both marketing practitioners and
technology providers will find this book very useful in guiding them through
the marketing value chain and how to fully digitize it. A comprehensive and
indispensable reference for anyone undertaking the transformational journey
towards algorithmic marketing.”
—Ali Bouhouch, CTO, Sephora Americas
“This book is a live portrait of digital transformation in marketing. It shows

how data science becomes an essential part of every marketing activity. The
book details how data-driven approaches and smart algorithms result in deep
automation of traditionally labor-intensive marketing tasks. Decision-making
is getting not only better but much faster, and this is crucial in our ever-
accelerating competitive environment. It is a must-read for both data scientists
and marketing officers–even better if they read it together.”
—Andrey Sebrant, Director of Strategic Marketing, Yandex
“This books delivers a complete end-to-end blueprint on how to fully digitize

your company’s marketing operations. Starting from a conceptual architecture
for the future of digital marketing, it then delves into detailed analysis of best
practices in each individual area of marketing operations. The book gives the
executives, middle managers, and data scientists in your organization a set of
concrete, actionable, and incremental recommendations on how to build better
insights and decisions, starting today, one step at a time.”
—Victoria Livschitz, founder and CTO, Grid Dynamics

vi
“This book provides a much-needed collection of recipes for marketing prac-

titioners on how to use advanced methods of machine learning and data science
to understand customer behavior, personalize product offerings, optimize the
incentives, and control the engagement – thus creating a new generation of
data-driven analytic platform for marketing systems.”
—Kira Makagon, Chief Innovation Officer, RingCentral;

serial entrepreneur, founder of RedAril and Octane
“While virtually every business manager today grasps the conceptual

importance of data analytics and machine learning, the challenge of imple-
menting actual competitive solutions rooted in data science remains quite
daunting. The scarcity of data scientist talent, combined with the difficulty
of adapting academic models, generic open-source software and algorithms
to industry-specific contexts are among the difficulties confronting digital
marketers around the world. This book by Ilya Katsov draws from the deep
domain expertise he developed at Grid Dynamics in delivering innovative,
yet practical digital marketing solutions to large organizations and helping
them successfully compete, remain relevant, and adapt in the new age of data
analytics.”
—Eric Benhamou,
Founder and General Partner, Benhamou Global Ventures;
former CEO and Chairman of 3Com and Palm
CONTENTS
1 introduction 1
1.1 The Subject of Algorithmic Marketing . . . . . . . . . . . 2
1.2 The Definition of Algorithmic Marketing . . . . . . . . . . 4
1.3 Historical Backgrounds and Context . . . . . . . . . . . . 5
1.3.1 Online Advertising: Services and Exchanges . . . . . 5
1.3.2 Airlines: Revenue Management . . . . . . . . . . . . . 7
1.3.3 Marketing Science . . . . . . . . . . . . . . . . . . . . . 10
1.4 Programmatic Services . . . . . . . . . . . . . . . . . . . . 11
1.5 Who Should Read This Book? . . . . . . . . . . . . . . . . 15
1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2 review of predictive modeling 19
2.1 Descriptive, Predictive, and Prescriptive Analytics . . . . 19
2.2 Economic Optimization . . . . . . . . . . . . . . . . . . . . 20
2.3 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Supervised Learning . . . . . . . . . . . . . . . . . . . . . . 25
2.4.1 Parametric and Nonparametric Models . . . . . . . . 26
2.4.2 Maximum Likelihood Estimation . . . . . . . . . . . . 27
2.4.3 Linear Models . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.3.1 Linear Regression . . . . . . . . . . . . . . . . . . 30
2.4.3.2 Logistic Regression and Binary Classification . . 31
2.4.3.3 Logistic Regression and Multinomial Classifica-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.3.4 Naive Bayes Classifier . . . . . . . . . . . . . . . . 34
2.4.4 Nonlinear Models . . . . . . . . . . . . . . . . . . . . . 36
2.4.4.1 Feature Mapping and Kernel Methods . . . . . . 36
2.4.4.2 Adaptive Basis and Decision Trees . . . . . . . . 40
2.5 Representation Learning . . . . . . . . . . . . . . . . . . . 41
2.5.1 Principal Component Analysis . . . . . . . . . . . . . 42
2.5.1.1 Decorrelation . . . . . . . . . . . . . . . . . . . . . 42
2.5.1.2 Dimensionality Reduction . . . . . . . . . . . . . 46
2.5.2 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.6 More Specialized Models . . . . . . . . . . . . . . . . . . . 52
2.6.1 Consumer Choice Theory . . . . . . . . . . . . . . . . 52
2.6.1.1 Multinomial Logit Model . . . . . . . . . . . . . . 54
2.6.1.2 Estimation of the Multinomial Logit Model . . . 57
2.6.2 Survival Analysis . . . . . . . . . . . . . . . . . . . . . 58
2.6.2.1 Survival Function . . . . . . . . . . . . . . . . . . 60
2.6.2.2 Hazard Function . . . . . . . . . . . . . . . . . . . 62
2.6.2.3 Survival Analysis Regression . . . . . . . . . . . . 64
vii
viii contents
2.6.3 Auction Theory . . . . . . . . . . . . . . . . . . . . . . 68

2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3 promotions and advertisements 75
3.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2 Business Objectives . . . . . . . . . . . . . . . . . . . . . . 80
3.2.1 Manufacturers and Retailers . . . . . . . . . . . . . . . 80
3.2.2 Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.2.3 Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.3 Targeting Pipeline . . . . . . . . . . . . . . . . . . . . . . . 85
3.4 Response Modeling and Measurement . . . . . . . . . . . 88
3.4.1 Response Modeling Framework . . . . . . . . . . . . . 89
3.4.2 Response Measurement . . . . . . . . . . . . . . . . . 93
3.5 Building Blocks: Targeting and LTV Models . . . . . . . . 94
3.5.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . 95
3.5.2 Tiered Modeling . . . . . . . . . . . . . . . . . . . . . . 96
3.5.3 RFM Modeling . . . . . . . . . . . . . . . . . . . . . . 98
3.5.4 Propensity Modeling . . . . . . . . . . . . . . . . . . . 99
3.5.4.1 Look-alike Modeling . . . . . . . . . . . . . . . . 100
3.5.4.2 Response and Uplift Modeling . . . . . . . . . . . 104
3.5.5 Segmentation and Persona-based Modeling . . . . . . 107
3.5.6 Targeting by using Survival Analysis . . . . . . . . . . 109
3.5.7 Lifetime Value Modeling . . . . . . . . . . . . . . . . . 113
3.5.7.1 Descriptive Analysis . . . . . . . . . . . . . . . . . 113
3.5.7.2 Markov Chain Models . . . . . . . . . . . . . . . . 116
3.5.7.3 Regression Models . . . . . . . . . . . . . . . . . . 119
3.6 Designing and Running Campaigns . . . . . . . . . . . . . 120
3.6.1 Customer Journeys . . . . . . . . . . . . . . . . . . . . 121
3.6.2 Product Promotion Campaigns . . . . . . . . . . . . . 123
3.6.2.1 Targeting Process . . . . . . . . . . . . . . . . . . 123
3.6.2.2 Budgeting and Capping . . . . . . . . . . . . . . . 126
3.6.3 Multistage Promotion Campaigns . . . . . . . . . . . 131
3.6.4 Retention Campaigns . . . . . . . . . . . . . . . . . . . 133
3.6.5 Replenishment Campaigns . . . . . . . . . . . . . . . 136
3.7 Resource Allocation . . . . . . . . . . . . . . . . . . . . . . 137
3.7.1 Allocation by Channel . . . . . . . . . . . . . . . . . . 138
3.7.2 Allocation by Objective . . . . . . . . . . . . . . . . . . 142
3.8 Online Advertisements . . . . . . . . . . . . . . . . . . . . 144
3.8.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . 144
3.8.2 Objectives and Attribution . . . . . . . . . . . . . . . . 146
3.8.3 Targeting for the CPA-LT Model . . . . . . . . . . . . 149
3.8.3.1 Brand Proximity . . . . . . . . . . . . . . . . . . . 151
3.8.3.2 Ad Response Modeling . . . . . . . . . . . . . . . 152
3.8.3.3 Inventory Quality and Bidding . . . . . . . . . . 152
contents ix
3.8.4 Multi-Touch Attribution . . . . . . . . . . . . . . . . . 154

3.9 Measuring the Effectiveness . . . . . . . . . . . . . . . . . 157
3.9.1 Randomized Experiments . . . . . . . . . . . . . . . . 157
3.9.1.1 Conversion Rate . . . . . . . . . . . . . . . . . . . 157
3.9.1.2 Uplift . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.9.2 Observational Studies . . . . . . . . . . . . . . . . . . 163
3.9.2.1 Model Specification . . . . . . . . . . . . . . . . . 166
3.9.2.2 Simulation . . . . . . . . . . . . . . . . . . . . . . 169
3.10 Architecture of Targeting Systems . . . . . . . . . . . . . . 172
3.10.1 Targeting Server . . . . . . . . . . . . . . . . . . . . . . 172
3.10.2 Data Management Platform . . . . . . . . . . . . . . . 174
3.10.3 Analytics Platform . . . . . . . . . . . . . . . . . . . . 174
3.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
4 search 179
4.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . 180
4.2.1 Relevance Metrics . . . . . . . . . . . . . . . . . . . . . 185
4.2.2 Merchandising Controls . . . . . . . . . . . . . . . . . 190
4.2.3 Service Quality Metrics . . . . . . . . . . . . . . . . . . 191
4.3 Building Blocks: Matching and Ranking . . . . . . . . . . 193
4.3.1 Token Matching . . . . . . . . . . . . . . . . . . . . . . 194
4.3.2 Boolean Search and Phrase Search . . . . . . . . . . . 195
4.3.3 Normalization and Stemming . . . . . . . . . . . . . . 196
4.3.4 Ranking and the Vector Space Model . . . . . . . . . 198
4.3.5 TFIDF Scoring Model . . . . . . . . . . . . . . . . . . 201
4.3.6 Scoring with n-grams . . . . . . . . . . . . . . . . . . . 205
4.4 Mixing Relevance Signals . . . . . . . . . . . . . . . . . . . 206
4.4.1 Searching Multiple Fields . . . . . . . . . . . . . . . . 207
4.4.2 Signal Engineering and Equalization . . . . . . . . . . 209
4.4.2.1 One Strong Signal . . . . . . . . . . . . . . . . . . 209
4.4.2.2 Strong Average Signal . . . . . . . . . . . . . . . . 212
4.4.2.3 Fragmented Features and Signals . . . . . . . . . 213
4.4.3 Designing a Signal Mixing Pipeline . . . . . . . . . . 215
4.5 Semantic Analysis . . . . . . . . . . . . . . . . . . . . . . . 218
4.5.1 Synonyms and Hierarchies . . . . . . . . . . . . . . . 219
4.5.2 Word Embedding . . . . . . . . . . . . . . . . . . . . . 222
4.5.3 Latent Semantic Analysis . . . . . . . . . . . . . . . . 224
4.5.4 Probabilistic Topic Modeling . . . . . . . . . . . . . . 231
4.5.5 Probabilistic Latent Semantic Analysis . . . . . . . . . 232
4.5.5.1 Latent Variable Model . . . . . . . . . . . . . . . . 233
4.5.5.2 Matrix Factorization . . . . . . . . . . . . . . . . . 236
4.5.5.3 pLSA Properties . . . . . . . . . . . . . . . . . . . 237
4.5.6 Latent Dirichlet Allocation . . . . . . . . . . . . . . . . 237
x contents
4.5.7 Word2Vec Model . . . . . . . . . . . . . . . . . . . . . 240

4.6 Search Methods for Merchandising . . . . . . . . . . . . . 247
4.6.1 Combinatorial Phrase Search . . . . . . . . . . . . . . 249
4.6.2 Controlled Precision Reduction . . . . . . . . . . . . . 253
4.6.3 Nested Entities and Dynamic Grouping . . . . . . . . 255
4.7 Relevance Tuning . . . . . . . . . . . . . . . . . . . . . . . . 258
4.7.1 Learning to Rank . . . . . . . . . . . . . . . . . . . . . 260
4.7.2 Learning to Rank from Implicit Feedback . . . . . . . 265
4.8 Architecture of Merchandising Search Services . . . . . . 269
4.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
5 recommendations 273
5.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . 275
5.1.1 Properties of Customer Ratings . . . . . . . . . . . . . 277
5.3 Quality Evaluation . . . . . . . . . . . . . . . . . . . . . . . 280
5.3.1 Prediction Accuracy . . . . . . . . . . . . . . . . . . . 282
5.3.2 Ranking Accuracy . . . . . . . . . . . . . . . . . . . . . 284
5.3.3 Novelty . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
5.3.4 Serendipity . . . . . . . . . . . . . . . . . . . . . . . . . 286
5.3.5 Diversity . . . . . . . . . . . . . . . . . . . . . . . . . . 286
5.3.6 Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . 287
5.3.7 The Role of Experimentation . . . . . . . . . . . . . . 288
5.4 Overview of Recommendation Methods . . . . . . . . . . 288
5.5 Content-based Filtering . . . . . . . . . . . . . . . . . . . . 291
5.5.1 Nearest Neighbor Approach . . . . . . . . . . . . . . . 294
5.5.2 Naive Bayes Classifier . . . . . . . . . . . . . . . . . . 296
5.5.3 Feature Engineering for Content Filtering . . . . . . . 301
5.6 Introduction to Collaborative Filtering . . . . . . . . . . . 303
5.6.1 Baseline Estimates . . . . . . . . . . . . . . . . . . . . 305
5.7 Neighborhood-based Collaborative Filtering . . . . . . . . 309
5.7.1 User-based Collaborative Filtering . . . . . . . . . . . 310
5.7.2 Item-based Collaborative Filtering . . . . . . . . . . . 316
5.7.3 Comparison of User-based and Item-based Methods 317
5.7.4 Neighborhood Methods as a Regression Problem . . 318
5.7.4.1 Item-based Regression . . . . . . . . . . . . . . . 319
5.7.4.2 User-based Regression . . . . . . . . . . . . . . . 322
5.7.4.3 Fusing Item-based and User-based Models . . . 323
5.8 Model-based Collaborative Filtering . . . . . . . . . . . . . 324
5.8.1 Adapting Regression Models to Rating Prediction . . 325
5.8.2 Naive Bayes Collaborative Filtering . . . . . . . . . . . 327
5.8.3 Latent Factor Models . . . . . . . . . . . . . . . . . . . 331
5.8.3.1 Unconstrained Factorization . . . . . . . . . . . . 335
5.8.3.2 Constrained Factorization . . . . . . . . . . . . . 339
contents xi
5.8.3.3 Advanced Latent Factor Models . . . . . . . . . . 342

5.9 Hybrid Methods . . . . . . . . . . . . . . . . . . . . . . . . 346
5.9.1 Switching . . . . . . . . . . . . . . . . . . . . . . . . . . 346
5.9.2 Blending . . . . . . . . . . . . . . . . . . . . . . . . . . 347
5.9.2.1 Blending with Incremental Model Training . . . 349
5.9.2.2 Blending with Residual Training . . . . . . . . . . 350
5.9.2.3 Feature-weighted Blending . . . . . . . . . . . . . 351
5.9.3 Feature Augmentation . . . . . . . . . . . . . . . . . . 354
5.9.4 Presentation Options for Hybrid Recommendations . 356
5.10 Contextual Recommendations . . . . . . . . . . . . . . . . 357
5.10.1 Multidimensional Framework . . . . . . . . . . . . . . 358
5.10.2 Context-Aware Recommendation Techniques . . . . . 360
5.10.3 Time-Aware Recommendation Models . . . . . . . . . 364
5.10.3.1 Baseline Estimates with Temporal Dynamics . . 364
5.10.3.2 Neighborhood Model with Time Decay . . . . . . 366
5.10.3.3 Latent Factor Model with Temporal Dynamics . 367
5.11 Non-Personalized Recommendations . . . . . . . . . . . . 368
5.11.1 Types of Non-Personalized Recommendations . . . . 368
5.11.2 Recommendations by Using Association Rules . . . . 370
5.12 Multiple Objective Optimization . . . . . . . . . . . . . . . 373
5.13 Architecture of Recommender Systems . . . . . . . . . . . 377
5.14 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
6 pricing and assortment 383
6.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . 384
6.2 The Impact of Pricing . . . . . . . . . . . . . . . . . . . . . 386
6.3 Price and Value . . . . . . . . . . . . . . . . . . . . . . . . . 388
6.3.1 Price Boundaries . . . . . . . . . . . . . . . . . . . . . 388
6.3.2 Perceived Value . . . . . . . . . . . . . . . . . . . . . . 390
6.4 Price and Demand . . . . . . . . . . . . . . . . . . . . . . . 393
6.4.1 Linear Demand Curve . . . . . . . . . . . . . . . . . . 394
6.4.2 Constant-Elasticity Demand Curve . . . . . . . . . . . 395
6.4.3 Logit Demand Curve . . . . . . . . . . . . . . . . . . . 397
6.5 Basic Price Structures . . . . . . . . . . . . . . . . . . . . . 399
6.5.1 Unit Price . . . . . . . . . . . . . . . . . . . . . . . . . 399
6.5.2 Market Segmentation . . . . . . . . . . . . . . . . . . . 401
6.5.3 Multipart Pricing . . . . . . . . . . . . . . . . . . . . . 406
6.5.4 Bundling . . . . . . . . . . . . . . . . . . . . . . . . . . 410
6.6 Demand Prediction . . . . . . . . . . . . . . . . . . . . . . 414
6.6.1 Demand Model for Assortment Optimization . . . . 416
6.6.2 Demand Model for Seasonal Sales . . . . . . . . . . . 419
6.6.2.1 Demand Data Preparation . . . . . . . . . . . . . 419
6.6.2.2 Model Specification . . . . . . . . . . . . . . . . . 420
6.6.3 Demand Prediction with Stockouts . . . . . . . . . . . 421
xii contents
6.7 Price Optimization . . . . . . . . . . . . . . . . . . . . . . . 424

6.7.1 Price Differentiation . . . . . . . . . . . . . . . . . . . 425
6.7.1.1 Differentiation with Demand Shifting . . . . . . . 428
6.7.1.2 Differentiation with Constrained Supply . . . . . 430
6.7.2 Dynamic Pricing . . . . . . . . . . . . . . . . . . . . . 433
6.7.2.1 Markdowns and Clearance Sales . . . . . . . . . . 436
6.7.2.2 Markdown Price Optimization . . . . . . . . . . . 439
6.7.2.3 Price Optimization for Competing Products . . . 441
6.7.3 Personalized Discounts . . . . . . . . . . . . . . . . . . 444
6.8 Resource Allocation . . . . . . . . . . . . . . . . . . . . . . 446
6.8.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . 447
6.8.2 Allocation with Two Classes . . . . . . . . . . . . . . . 450
6.8.3 Allocation with Multiple Classes . . . . . . . . . . . . 453
6.8.4 Heuristics for Multiple Classes . . . . . . . . . . . . . 455
6.8.4.1 EMSRa . . . . . . . . . . . . . . . . . . . . . . . . . 456
6.8.4.2 EMSRb . . . . . . . . . . . . . . . . . . . . . . . . 456
6.9 Assortment Optimization . . . . . . . . . . . . . . . . . . . 457
6.9.1 Store-Layout Optimization . . . . . . . . . . . . . . . . 458
6.9.2 Category Management . . . . . . . . . . . . . . . . . . 460
6.10 Architecture of Price Management Systems . . . . . . . . 466
6.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
a appendix: dirichlet distribution 471
index 475
bibliography 481
ACKNOWLEDGEMENTS
This book would not have been possible without the support and
help of many people. I am very grateful to my colleagues and friends,
Ali Bouhouch, Max Martynov, David Naylor, Penelope Conlon, Sergey
Tryuber, Denys Kopiychenko, and Vadim Kozyrkov, who reviewed the
content of the book and offered their feedback. Special thanks go to
Konstantin Perikov, who provided a lot of insightful suggestions about
search services and also helped with some of the examples.
I am indebted to Igor Yagovoy, Victoria Livschitz, Leonard Livschitz,
and Ezra Berger for supporting this project and helping with the pub-
lishing. Last, but certainly not least, thanks to Kathryn Wright, my
editor, who helped me shape the manuscript into the final product.
xiii
1
INTRODUCTION
In 1888, Vincent van Gogh, a little-known Dutch artist at that time,

wrote to his brother, Theo, that the painter of the future would be a colorist
the like of which has never yet been seen. Leaving the artistic aspect aside,
the very way in which Van Gogh poses the question about the painters
of the future and answers it is striking and admirable. Van Gogh, no
doubt, was right in anticipating that the artists of the upcoming century
would develop skills that have never yet been seen and questioning the
ways of doing art. What if we ask the same question about marketing
practitioners who live in the era of digital media and abundance of
data? Who will be the marketers of the future? Will they be artists of
client communications? Statisticians the like of whom have never yet
been seen? Software engineers who create marketing systems? Experts
in economic modeling?
The history of marketing can be viewed as the evolution of princi-
ples, techniques, and best practices for a certain kind of business op-
timization. It has always been recognized that this optimization prob-
lem can be approached in a scientific way and that rigorous mathe-
matical methods can be applied to a wide range of marketing appli-
cations. Adopters of such methods, however, traditionally struggled
with challenges related to incompleteness of data, complexity of real-
life marketing settings, inflexibility of business processes, and software
limitations. The challenges were especially overwhelming in areas that
required far-reaching strategic decisions, where human judgment was
often the only viable solution for practical applications.
The advancement of digital marketing channels changed the game
and created an environment that requires millions of micro-decisions
to be made, which simply cannot be done efficiently without intelli-
gent marketing software and algorithms. Targeted sales promotions,
dynamic pricing in brick-and-mortar and online stores, e-Commerce
search and recommendation services, online advertising – all of these
applications require advanced methods of economic modeling, data
1
2 introduction
science, and software engineering to realize the potential of the digi-

tal environment. For instance, this potential cannot be fully achieved
without tailoring personalized experiences for millions of individual
customers, which, in turn, requires millions of unique decisions to be
made. Moreover, ubiquitous digital media and mobile devices have
empowered customers to go through the entire marketing funnel from
research to purchase in seconds, anywhere and anytime, and this per-
vasive micro-moment behavior requires marketing decisions to be taken
in microseconds as well. This environment introduces the challenge
of building marketing systems that make decisions and act at an un-
precedented level of autonomy, scale, and depth of analysis. In certain
cases, not only can individual decisions and analytics be done in a
data-driven way, but entire business processes can be planned, exe-
cuted, measured, and optimized by an automatic software system.
Although the problem of marketing automation can be studied from
different perspectives, including economics, management, statistics,
and engineering, the creators of such systems have to line up all of
these pieces into a cohesive set of methods that can be efficiently
implemented in software to achieve the business objectives. Leading
a modern marketing technology project is not unlike conducting
an orchestra of diverse instruments and making them operate in
concert. We take exactly this perspective on marketing throughout
this book and put together the results of the vast experience gained
by developers of marketing systems in retail, online advertising, and
other industries over the last few decades, as well as the guiding
theoretical principles. It should be noted that we deliberately focus
on the results reported by industry practitioners that have proved to
be useful in business solutions, rather then theoretical and academic
studies. Fortunately, the number of methods, models, and architec-
tures published by such practitioners is high and sometimes provides
a great level of detail. Some of these publications are mainly focused
on the technology and implementation aspects, whereas others dive
deep into mathematical modeling, optimization, and econometrics. In
practice, both aspects are important for the successful creation and
operation of a marketing system. Many of these results are also based
on or related to models developed in scientific marketing by academic
researchers.
1.1 the subject of algorithmic marketing
One of the traditional definitions of marketing describes it as the activ-

ity of defining products and services offered by a company and com-
municating them to existing or potential customers. This activity can
1.1 the subject of algorithmic marketing 3
be broken down into several streams that are typically described as

variations of the following categories [McCarthy, 1960]:
• Product – analysis of marketing opportunities, planning of prod-
uct lines and product features, assortment planning.
• Promotion – all methods of communication between the com-

pany and its customers: advertisements, recommendations, cus-
tomer care, and others.
• Price – pricing strategies, including posted prices, price dis-

counts, and price changes over time.
• Place – historically, this refers to the process of making a prod-

uct or service available to the end user through various distri-
bution channels. More recent interpretations emphasize the role
of product discovery and convenience to buy, with the argument
that distribution is becoming less relevant with the rise of digital
marketing channels [Lauterborn, 1990].
This categorization is well known as the marketing mix. The mix can
be viewed as a set of variables that a marketing manager or marketing
software can control to influence the position of products and brands in
the market. Each component of the marketing mix represents a broad
area that can be viewed and studied from different perspectives. The
subject of algorithmic marketing can be better understood by distin-
guishing the following two aspects of marketing activities: strategy and
process. We use the term strategy to label long-term top-level business
decisions that define the value proposition of the company and set the
overall direction for its marketing processes. For example, a retailer
has to define its target market, customer services, and product lines
as parts of the business strategy. The process is an implementation of
the strategy that focuses on tactical decisions that support continuous
functioning of the company. Continuing the example with a retailer,
high-level pricing and promotional strategies require numerous deci-
sions about how to select consumers for promotional campaigns or
how prices for individual products should change over time.
Although the scope of neither strategy nor tactical processes can be
rigorously defined, and there is no clear boundary between these two
counterparts, we can argue that the strategy side is more focused on ex-
ploration, analysis, and planning involving human judgment, whereas
the process side is more focused on execution, micro-decisioning, and,
most importantly, automation. This makes the process side of market-
ing especially attractive for our study, although both strategy and pro-
cess can be described from the viewpoint of data science and clearly
4 introduction
benefit from data-driven methods. The short summary is that the sub-
ject of algorithmic marketing mainly concerns the processes that can
be found in the four areas of the marketing mix and the automation
of these processes by using data-driven techniques and econometric
methods.
1.2 the definition of algorithmic marketing
We define algorithmic marketing as a marketing process that is auto-

mated to such a degree that it can be steered by setting a business
objective in a marketing software system. This implies that the market-
ing system should be intelligent and knowledgeable enough to under-
stand a high-level objective, such as the acquisition of new customers
or revenue maximization, to plan and execute a sequence of business
actions, such as an advertisement campaign or price adjustment, with
the aim of achieving the objective, and to learn from the results to
correct and optimize the actions if needed. This basic principle is illus-
trated in Figure 1.1. In this book, we also use the term programmatic
to refer to highly automated marketing software systems and services,
and the terms algorithmic and programmatic are used interchangeably
in most contexts.
Figure 1.1: A conceptual view of the algorithmic marketing ecosystem.
Although it would be ideal for a programmatic system to be perfectly

automated and autonomous, we do not consider this a principal goal
or design requirement. On the contrary, a programmatic system is typ-
ically maintained by many people, including data scientists, engineers,
and analysts, who develop and adjust models and algorithms to im-
prove the system’s efficiency and capabilities. It can also consume the
outputs of strategic analysis and planning done elsewhere with non-
programmatic methods and, possibly, in connection with some other
problems. However, the system’s ability to understand the business
objective and work through the entire process from the objective to
measurable results is essential. Again, it is important to keep in mind
the limitations and perils of automation in marketing. In many real-life
1.3 historical backgrounds and context 5
applications, it is more appropriate to view programmatic systems as

intelligent tools that enable marketers to efficiently achieve what they
want, rather than as their replacements.
1.3 historical backgrounds and context
There is no sharp boundary between algorithmic and non-algorithmic

marketing. In some senses, it is even invalid to draw such a boundary
because algorithmic systems are just a method of answering old mar-
keting questions, not posing new ones. However, it is evident that the
level of marketing automation differs sharply across industries, which
indicates that some environments are more favorable in that regard
than others. Conversely, successful acquisition of advanced algorithmic
methods can drastically transform an industry and create increasingly
better conditions for further development. Analysis of such favorable
environments is a natural starting point for understanding algorith-
mic marketing. Let us briefly review several business cases that laid
the foundation for algorithmic marketing in search of the patterns and
characteristics that enabled the systematic approach.
1.3.1 Online Advertising: Services and Exchanges
The history of internet advertising can be traced back to May 3, 1978,

when the first spam email was sent to 400 users of the computer net-
work ARPANET, deployed at that time in just four locations: the Uni-
versity of Utah, UCLA, UC Santa Barbara, and Stanford Research. Fif-
teen years later, by 1993, when ARPANET had developed into the In-
ternet and the spread of the Web enabled multimedia websites, the
market of banner ads appeared. This new market originally relied on
direct selling of banner slots offered by website publishers to advertis-
ers, but this approach started to lose its efficiency very quickly when
there was a surge in the number of websites. It became operationally
difficult or even impossible for advertisers to run ad campaigns and
manage budgets across thousands of publishers. On the other hand,
publishers needed a robust and centralized way to sell their inventory
at scale.
The challenge was taken up by ad networks that acted as brokers
between publishers and advertisers. DoubleClick, launched in 1996, of-
fered a platform that enabled advertisers to run ad campaigns across a
wide network of websites, dynamically customize a campaign accord-
ing to its performance, and measure the return on investment. This
created a perfect environment for automatic decision making because
6 introduction
the measurements and adjustments could be done dynamically. How-

ever, it was not really programmatic at that time.
The online search engines, meanwhile, were also struggling to im-
prove their advertising capabilities. Advertisers paid for the number of
times their ad was demonstrated by the search engine – the cost per
thousand impressions (CPM) model – similarly to banner ads. This
approach was inflexible from the pricing perspective, causing certain
revenue losses for search engines, and also from the targeting perspec-
tive because irrelevant ads were not penalized in any way. The break-
through happened in 1998 when the GoTo.com search engine intro-
duced an automated auction model with two innovative features:
• Advertisers could bid how much they would be willing to pay to
appear at the top of the results for specific search queries.
• Advertisers paid per click, not per impression.

The per-pay-click (PPC) model improved both revenues and ad rele-
vancy because advertisers who were willing to pay for top ad spots for
specific search queries generally offered more relevant and better re-
sources. This model was adopted by Google in 2002 with one principal
improvement: the ad was selected based on Google’s expected revenue,
not the bid amount. Google measured the click-through rate for each
ad as a ratio between clicks and impressions, and the expected revenue
was estimated as
revenue bid price click-through rate
This was a programmatic self-learning technique that optimized the

business objective, both in terms of revenue and relevancy, because
click-through rates tend to be low for irrelevant ads, so even high-
budget advertisers were not able to clog up the bandwidth.
The trajectories of ad networks and search engines converged in
2007–2009 with adoption of the auction model across the board. Adver-
tisers and publishers became connected by ad exchanges that accepted
real-time bids for individual ad impressions, and a new era of real-time
bidding, commonly abbreviated as RTB, thus began. The advent of RTB
exchanges gave impetus to programmatic tools for advertisers – data-
management platforms (DMPs) and demand-side platforms (DSPs) –
that provided the ability to collect data about the behavior of Internet
users and make bids on RTB exchanges depending on the estimated
propensity of a given user to respond. The success of RTB was impres-
sive: the share of inventory sold by DoubleClick (acquired by Google
by that time) through RTB rose from 8% in January 2010 to 68% in May
2011 [Google Inc., 2011].
Reflecting on the history of RTB, we can conclude that one of the

most prominent achievements of programmatic advertising is a frame-
work that enables owners of consumer bases, originally the publishers
of web content, to provide personalized marketing services to parties
who are limited in their ability to interact with consumers, originally
the advertisers of products and services. The infrastructure that sits
in between the publishers and advertisers is typically provided by an
independent party and includes the following:
• Advertising services that enable advertisers to run advertising cam-
paigns using the publisher’s resources. These services are typi-
cally used to connect multiple advertisers with multiple publish-
ers and resemble a marketplace where resources are sold and
bought, often on a bidding basis.
• Data services that collect and store information about consumers,

taking it from publishers, advertisers, and third parties. Advertis-
ing services take advantage of this data to run ad campaigns and
make real-time automatic decisions on ads to be delivered.
Later on, this pattern started to spread across other industries. Other
types of consumer-base owners, such as retailers and mobile opera-
tors, were also looking for an efficient way to commercialize their data
and relationships with consumers, and other types of service users,
such as banks, product manufacturers, and insurance companies, were
willing to know more about their customers and have more channels
for communication with them. For instance, a manufacturer of con-
sumer packaged goods can use a retailer’s channels, such as stores and
eCommerce websites, to offer personalized discounts to consumers to
promote new products and increase their market share.
Consequently, advertising services and data services started to trans-
form into the more generic model illustrated in Figure 1.2, which rep-
resents a multipurpose marketplace of services and data that connects
actors from different industries. The range of services offered by such
a marketplace can go far beyond advertising, covering areas like credit
scores and insurance premiums. The heterogeneity of this environ-
ment, where one constantly deals with someone else’s data, often in
real-time, leads to overwhelming complexity of data flows and oper-
ational decisions, and programmatic methods are probably the only
way to tackle it.
1.3.2 Airlines: Revenue Management
Online advertising and data marketplaces are perhaps the most famous
and successful cases in the history of programmatic marketing, but
8 introduction
Figure 1.2: The marketplace model for data and services.
definitely not the only one. Online advertising, to a degree, was a com-
pletely new environment that provided unprecedented capabilities for
solving unprecedented challenges. Programmatic methods, however,
can be successfully applied in more traditional settings. Let us con-
sider a prominent case that, by coincidence, also began to unfold in
1978, the same year that the first spam email was sent.
The federal Civil Aeronautics Board had regulated all US interstate
air transport since 1938, setting schedules, routes, and fares based on
standard prices and profitability targets for the airlines. The Airline
Deregulation Act of 1978 eliminated many controls and enabled air-
lines to freely change prices and routes. It opened the door for low-cost
carriers, who pioneered simpler operational models, no-frills services,
and reduced labor costs. One of the most prominent examples was
PeopleExpress, which started in 1981 and offered fares about 70 per-
cent lower than the major airlines.
The low-cost carriers attracted new categories of travelers who had
rarely traveled by air before: college students visiting their families,
leisure travelers getting away for a few days, and many others. In 1984,
PeopleExpress reported revenue of $1 billion and its profits hit $60 mil-
lion [Talluri and Van Ryzin, 2004]. The advent of low-fare airlines was
a growing threat to the major carriers, who had almost no chance to
win the price war. Moreover, the established airlines could not afford to
lose their high-revenue business travelers in pursuit of the low-revenue
market.
The solution was found by American Airlines. First, they recognized
that unsold seats could be used to compete on price with the low-cost
carriers, because the marginal cost of such seats was close to zero any-
way. The problem, however, was how to prevent the business travelers
from purchasing tickets at discounted prices. The basic solution was
to introduce certain constraints on the discounted offers; for example,
the tickets had to be purchased at least three weeks in advance and
were non-refundable. The challenging part was that the surplus of seats
varied substantially across flights and the optimal allocation could be
achieved only by using dynamic optimization. In 1985, after a few years
of development, American Airlines released a system called Dynamic
Inventory Allocation and Maintenance Optimizer (DINAMO) to man-
age their prices across the board. PeopleExpress also used some simple
price-management strategies to differentiate peak and off-peak fares,
but their information technology system was much simpler than DI-
NAMO and was not able to match its efficiency. PeopleExpress started
to lose money at a rate of $50 million a month, headed straight to
bankruptcy, and eventually ceased to exist as a carrier in 1987 after
acquisition by Continental Airlines [Vasigh et al., 2013]. American Air-
lines, however, not only won the competition with PeopleExpress but
also increased its revenue by 14.5% and profits by 47.8% the year after
DINAMO was launched.
The case of American Airlines was the first major success of a rev-
enue management practice. By the early 1990s, the approach had been
adopted by other industries with perishable inventories and customers
booking the service in advance: hospitality, car rentals, and even tele-
vision ad sales. The success of revenue management in the airline in-
dustry is clearly related to the specific properties of the inventory and
demand in this domain:
• The demand varies significantly across customers, flights, and
time: the purchasing capacity of business travelers can be much
higher than that of discretionary travelers, peak flights can be
much more loaded than off-peak, etc.
• The supply, that is, the available seats, is not flexible. The airline
produces seats in large chunks by scheduling flights, and once
the flight is scheduled, the number of seats cannot be changed.
Unsold seats cannot be removed, so the airlines’ profit completely
depends on its ability to manage the demand and sell efficiently.
We can conclude from the above that revenue management can be
considered as a counterpart of supply-chain management or, alterna-
tively, a demand-chain management solution that struggles with in-
flexible production and adapts it to the demands of the market (and,
conversely, manipulates the demand to align it with the supply), as
illustrated in Figure 1.3.
10 introduction
Figure 1.3: Revenue management as a counterpart of supply-chain manage-

ment.
This friction between supply and demand can be found not only
in air transportation but in other industries as well. Hotels and car
rentals are clearly the closest examples, but advertising, retail, and oth-
ers also demonstrate features that indicate the applicability of algorith-
mic methods for demand-chain management.
1.3.3 Marketing Science
The case studies of online advertising and airline ticketing give a gen-
eral idea of how algorithmic methods advanced in the industry. In-
dustrial adoption was backed by rapid developments in marketing sci-
ence and, conversely, the development of scientific marketing methods
was boosted and propelled by industrial needs. Marketing as a dis-
cipline emerged in the early 1900s, and, for the first five decades of
its existence, it was mainly focused on the descriptive analysis of pro-
duction and distribution processes; that is, it collected the facts about
the flow of goods from producers to consumers. The idea that mar-
keting decisions could be supported by mathematical modeling and
optimization methods began to gain currency in the 1960s, a fact that
can be attributed to several factors. First, marketing science was influ-
enced by advances in operations research, a discipline that deals with
decision-making and efficiency problems in military and business ap-
plications by using statistical analysis and mathematical optimization.
Operations research, in turn, arose during World War II in the context
of military operation planning and resource optimization. Second, the
advancement of mathematical methods in marketing can be attributed
to technological changes and the adoption of the first mainframes in
organizations, which made it possible to collect more data and im-
plement data analysis and optimization algorithms. Finally, marketing
practitioners began to feel that the old ways of selling were wearing
thin and that marketing needed to be redefined as a mix of ingredi-
ents that can be controlled and optimized; this is how the concept of
the marketing mix appeared in 1960. Marketing science boomed in the
sixties and seventies of the twentieth century when numerous quan-
titative models for pricing, distribution, and product planning were
1.4 programmatic services 11
developed using advanced probabilistic and optimization techniques.

Some of these methods were readily adopted in practice, as in the case
of revenue management in the airline and hotel industries, but many
others had limited practical applicability, so the overall level of market-
ing automation remained low in many industries [Wierenga, 2010].
The advancement of digital channels has changed the situation dras-
tically. Digital media created the need and the opportunity to make
millions of micro-decisions at the level of individual customers and
provide totally new services, such as product search or real-time mo-
bile notifications. This has created challenges that often go beyond eco-
nomic modeling and optimization, which is the main theme in tradi-
tional marketing science, and require the use of advanced software
engineering and data analysis methods that were not originally con-
nected to marketing. In modern retail, for instance, a substantial part
of the revenue can be generated by search and recommendations ser-
vices that internally rely on text analysis methods, rather than eco-
nomic models. Some of these methods came from areas as distant from
marketing as biology and genome studies. To summarize, we can say
that traditional economic modeling, data science, software engineering,
and conventional marketing practices are all important for the creation
of programmatic systems.
1.4 programmatic services
The marketing mix model defines four factors that can be controlled
by a company to influence consumer purchase decisions: product, pro-
motion, price, and place. As such, this categorization is very broad
and provides little guidance on how exactly programmatic marketing
systems should be built. So far, we have learned that a programmatic
system can be viewed as a provider of one or more functional services
that implement certain business processes, such as price or promotion
management. Consequently, we can make our problem statement more
specific by defining a set of services, each of which implements a cer-
tain function and has its own inputs (objectives) and outputs (actions).
There are different options for how the marketing mix can be broken
down into functional services depending on the industry and busi-
ness model of a particular company. We choose to define six major
functional services that are relevant for a wide range of business-to-
consumer (B2C) verticals: promotions, advertisements, search, recom-
mendations, pricing, and assortment. These six services are the main
subject of this book, and we will spend the later chapters discussing
how to design and build them. The applications and design principles
are very different across the services, but there are many relationships
12 introduction
between them as well. Let us briefly review these relationships and

some common design guidelines that will be used in the rest of this
book to streamline the detailed discussions of individual services.
The relationships between the six services we have defined, as well as
their connections to the marketing mix, can be established as follows:
• The primary purpose of promotion and advertisement services is
to match customers with offerings and convey the right message
to them. This typically requires customers to be found who can
be incentivized to do actions that contribute toward the desired
business objective. The capability for identifying the right cus-
tomers and offerings is the keystone of this service group. From
the marketing mix perspective, these services directly address
the Promotion domain of the mix and are also related to the Price
domain through the costs and profits associated with promotion
and advertising campaigns.
• The search and recommendation services solve the problem of

finding the right products for a given customer, which naturally
complements the previous service group. The principal goal of
these services is to enable and simplify product discovery, which
is related to the Place and Promotion domains of the market-
ing mix. This group of services requires an understanding of the
purchasing intent of the customer, expressed either explicitly or
implicitly, and the ability to find offerings that match it well.
• The goal of pricing and assortment services is to determine and

optimize the set of offerings and their properties, including price.
These services often rely on a capability to predict the demand
as a function of assortment, prices, and other parameters, which
enables what-if analysis and optimization of different options.
This group mainly covers the Price and Product domains of the
marketing mix.
This three-group classification, illustrated in Figure 1.4, is convenient
because it reflects the similarities in the objectives and design princi-
ples shared across the services. We structure the rest of this book on
the basis of these three groups to address both fundamental capabili-
ties, such as identification of the right customers for an offering, and
individual services.
The next question we can pose is what these services have in com-
mon from the design and implementation standpoints. Although the
design principles and implementation methods are very different for
the different services, the programmatic approach introduces common
guidelines that can be explicitly or implicitly followed by all services.
1.4 programmatic services 13
Figure 1.4: Programmatic services.
We can use these guidelines to define the basic terminology and compo-
nents that can later be elaborated within the scope of the corresponding
domains.
The concept of the programmatic method accentuates the objective-
driven design approach, so we can attempt to define a common frame-
work by starting with the notion of the business objective. In order to
understand and execute the objective, a programmatic service is likely
to include a certain set of functional components that can have different
designs for different domains (see Figure 1.5):
• Since any automatic decision making is driven by data, the
decision-making pipeline starts with data collection. Examples
of input data for most marketing applications include customers’
personal and behavioral profiles, inventory data, and sales
records.
• The raw input data often has to be transformed into well-defined

features that can be fed into analysis and decision-making algo-
rithms. The reason is that programmatic services often rely on
some measure of similarity between entities like products or cus-
tomers to learn the patterns and make decisions, which requires
the entities to be represented as comparable sets of attributes.
Search services, for example, often rely on some similarity mea-
sure between a user query and the products to find the most
relevant offerings, which requires both products and queries to
be converted into well-defined comparable representations. The
engineering of these attributes, referenced as features in many
14 introduction
contexts, plays a critically important role in programmatic sys-

tems.
Figure 1.5: A high-level design framework for programmatic services.
• The most important step in the programmatic pipeline is to eval-

uate how well different action strategies fit the business objective.
This generally requires the development of one or more models
that consume a candidate solution and produce signals that indi-
cate the level of fitness. For example, a promotion service can rely
on a model that scores the customer according to their propensity
to buy a certain product, a pricing service can score different
pricing options according to expected profits, and a search ser-
vice can score products by their relevance for a query.
• The signals generated by models carry information about the

quality of different decisions that a marketer can make. A busi-
ness action, however, often requires the combination of many sig-
nals and intermediate decisions into the final action plan. For
example, a marketer can send promotions to the most valuable
clients, which requires individual customers to be evaluated, but
the final mailing list has to be constrained to the campaign bud-
get. Thus, programmatic services typically contain an optimiza-
tion or signal-mixing component that makes the final decisions.
• A programmatic service communicates with the outside world

through marketing channels and integrates with other services.
These channels determine a set of possible business actions that a
1.5 who should read this book? 15
service can take and action parameters that a service can control,
such as price levels, discount values, email messages, or prod-
uct order in search result lists. These controls are used by the
programmatic service to execute its decisions, so the decisions
should eventually be expressed as parameters for available con-
trols.
• Finally, the feedback collected from the execution channels can be

routed back to the models and optimization procedures to learn
from the results and adjust the decision-making logic. The mea-
surement stage is a mandatory part of all marketing services, and
many marketing methods rely on a measurement-driven trial
and error approach to optimization.
In practice, a marketer often wants to achieve multiple strategic goals
and can leverage multiple services to do it. The programmatic infras-
tructure can support this by providing capabilities for market opportu-
nities analysis and global resource allocation that help to elaborate the
objectives and parameters for individual services, as well as consoli-
date the measurements. This planning functionality and the execution
pipelines of individual services make up the complete programmatic
ecosystem. The framework we have defined is very abstract at this
point, but we will gradually elaborate design methods for all stages
of the pipeline in the subsequent chapters.
1.5 who should read this book?
This book is for everyone who wants to learn how to build advanced
marketing software systems. It will be useful for a variety of marketing
and software practitioners, but it was written with two target audiences
in mind. The first target audience is implementers of marketing soft-
ware, product managers, and software engineers who want to learn
about the features and techniques that can be used in marketing soft-
ware products and also learn about the economic foundations for these
techniques. The other target audience is marketing strategists and tech-
nology leaders who are looking for guidance on how marketing organi-
zations and marketing services can benefit from machine learning and
Big Data and how modern enterprises can leverage advanced decision
automation methods.
It is assumed that readers have an introductory background in statis-
tics, calculus, and programming. Although most methods described in
the book use relatively basic math, this book may not be suitable for
readers who are interested only in the business aspects of marketing:
this book is about marketing automation; it is not a traditional marketing
16 introduction
textbook. Overall, if you are comfortable with the following expression,

you are probably good to go:
»8
E rXs x fX pxqdx
8
This book is divided into six chapters. This first chapter, Introduc-
tion, outlines the main concepts and principles of algorithmic mar-
keting and also discusses several inspiring case studies that illustrate
the prerequisites and benefits of the algorithmic approach. The sec-
ond chapter, Review of Predictive Modeling, focuses on the mathemat-
ical foundations of algorithmic marketing. The last four chapters cover
four different domains of marketing – advertisements and promotions,
search services, recommendations, and pricing – one domain at a time.
These four main chapters follow the same algorithmic methodology
and, thus, have the same structure: we start each chapter with a descrip-
tion of the environment to understand the constraints and variables
that we can optimize, we then discuss business objectives to better de-
fine the optimization problem, and we finally work through decision
automation methods for different tasks and scenarios that appear in
the domain. As each of the four main chapters addresses its own do-
main, the chapters are relatively independent, and readers can thus
use this book as a reference to read selectively about the areas that are
most relevant for their needs or read sequentially from cover to cover.
Readers should feel free to skip parts that are not relevant to their
background or intended use. For example, readers familiar with prob-
ability, statistics, and machine learning can just quickly scan through
the second chapter or skip it completely. Readers who are primarily in-
terested in the business applications and capabilities of the algorithmic
approach can focus on sections that describe environments, business
objectives, and optimization problems, and they can skip the mathe-
matical details and numerical examples. On the other hand, readers
who implement marketing systems should most likely examine the al-
gorithms, numerical examples, and implementation details.
1.6 summary
• A programmatic approach to marketing focuses on creating highly

automated marketing systems and processes that can be steered by
business objectives. Programmatic methods can be applied to all
of the domains of the marketing mix, namely, product, promotion,
price, and place.
• Programmatic marketing can be viewed as services that provide in-

sights into the market state and enable one to interact with the mar-
1.6 summary 17
ket. These services can be used internally by the company that owns
the customer base or can be sold to a third party. Programmatic com-
ponents should be self-contained, well-packaged, and able to imple-
ment a reasonably high level of abstraction to be sold as high-value
services.
• A programmatic service often acts as a dynamic controller of

demand and, consequently, a counterpart of supply-chain man-
agement. The efficiency of programmatic methods increases with
variability of demand across consumers or over time and decreases
with flexibility of production. Consequently, programmatic methods
are most rewarding in areas with high diversity of consumer tastes
and incomes and/or areas with inflexible production that creates
a surplus of goods or services with low marginal cost if demand
drops.
• The most important examples of programmatic services include

promotions, advertisements, search, recommendations, pricing, and
assortment. The design principles differ across the services, but a
few functional capabilities and logical components are widely used
across all service types. Examples of such components are scoring
models that assess the fitness of candidate solutions for a selected
business objective, optimization models that analyze and mix the
scores to make the final decisions, and controls that are used to turn
decisions into actions.
2
REVIEW OF PREDICTIVE MODELING
Algorithmic marketing, by definition, cannot exist without a methodol-

ogy for the evaluation of possible business actions and corresponding
outcomes based on the available data. In this chapter, we will review
the basic methods of machine learning and economic modeling that
enable predictive analysis and provide the building blocks that will be
used in the rest of the book. Our primary goal here is to describe the
main capabilities and limitations of predictive modeling, rather than
to provide a comprehensive study of machine learning algorithms. We
will describe only a few methods that are routinely used in market-
ing applications and provide mathematical details and examples only
to illustrate the capabilities, limitations, and relationships with other
methods. We will not go into the practical aspects of modeling, such as
preparation of the input data and evaluation of the resulting models,
in depth in this chapter; these details are left until later in the book.
2.1 descriptive, predictive, and prescriptive analytics
Before we proceed to predictive modeling, a brief note regarding the

terminology used in the marketing domain. In business applications,
data analytics methods are often categorized at a high level into three
distinct types: descriptive, predictive, and prescriptive. Descriptive an-
alytics refers to methods for data summarization, data quality assess-
ment, and finding correlations. Examples of descriptive analytics are
management reports providing aggregated sales data or market basket
analyses yielding information about the products frequently bought to-
gether. Descriptive analytics does not aim to explain how the observed
results can be influenced or optimized. Predictive analytics focuses on
estimation of the likelihood of a potential outcome by using data that
are observed or known prior to the outcome. Common examples of
predictive analytics are the forecasting of demand and the propensity
scoring used to determine the likelihood of customers responding to
19
20 review of predictive modeling
a promotion. Note that “predictive” does not necessarily mean a fore-

cast of the future – it is used in the sense that the predictive model
can estimate some output variable given some change in an input vari-
able. Finally, prescriptive analytics refers to modeling of the dependency
between decisions and future outcomes for optimal decision making.
One of the basic examples of prescriptive analysis is price optimiza-
tion where the profit is modeled as a function of the price, so that one
can estimate how many dollars of profit would be generated by ev-
ery dollar of price discount and determine the profit-optimal discount
value.
In the marketing domain, data-related activities and processes are
commonly viewed through the lenses of these three types of analytics,
and all three types are important. In programmatic applications, where
automated decision making is key, the main focus is on prescriptive an-
alytics, which, in turn, uses predictive models as building blocks. Con-
sequently, it should be kept in mind that programmatic applications
only use a subset of marketing analytics methods.
2.2 economic optimization
Marketing is an activity that aims to achieve certain business objectives

by executing business actions of a certain type. The first step we need
to take as we begin to discuss the algorithmic approach is to translate
this business language into more formal models that describe the ob-
jective we are trying to achieve, the space of possible actions, and the
constraints we should meet. Most marketing problems translated into
this econometric language are naturally optimization problems that ex-
press business metrics, like revenue, as a function of possible actions,
such as marketing campaigns or assortment adjustments, and require
the optimal action to be found from among the various possible strate-
gies.
The economic model is also a function of data in the sense that it
uses properties and parameters estimated from past experience. For
instance, consider a retailer planning a marketing mailing campaign.
The space of possible actions can be defined as a set of send/no-send
decisions with regard to individual customers, and the revenue of the
campaign depends both on actions, that is, who will receive the incen-
tive and who will not, and data, such as expected revenue from a given
customer and mailing costs. This approach can be expressed in a more
formal way as follows:
sopt argmax G ps, Dq (2.1)

s PS
2.2 economic optimization 21
in which D is the data available for analysis, S is the space of actions

and decisions, G is an economic model that maps actions and data to
an economic outcome, and sopt is the optimal strategy. The design of
model G completely depends on the application. We will discuss the
construction of different models in the next chapters in the context of
particular marketing problems, but there are several basic considera-
tions that should be taken into account in any model design.
First, we need to define the business objective and express it as a nu-
merical metric that can be a subject of optimization. In many cases, it
is reasonable to model and optimize the profit, but, as we will see later,
other objectives can be set for particular applications. The design of
the objective can be especially challenging if the objective represents a
trade-off between the enterprise’s profit and the usefulness to the con-
sumer, for example, a search in an online catalog that should provide
results both relevant to the query entered by the consumer and aligned
with the merchandising goals and rules.
Second, we should account for available data or address the data
collection problem. The role of data mining in problem 2.1 is crucial
because model G, which links the actions with the outcomes, can be
complex and may have to be determined from data by means of re-
gression analysis or other data mining techniques. In some cases, the
model cannot be completely specified, either because of high complex-
ity (e. g., user behavior cannot be precisely predicted) or because it is
impossible to extrapolate the existing data to the case in point (e. g., the
action is to introduce a completely new product or service). In either
case, the data should be treated as a business asset that might require
investment, and there is a trade-off between the cost of data acquisi-
tion and the value delivered by the acquired data. For example, the
real-life testing of multiple models in parallel is a simple way to trade
economic efficiency (running different models simultaneously is obvi-
ously suboptimal) for acquisition of more data points and, sometimes,
lower model complexity.
Third, the model can be created at different levels of granularity. We
can express the same objective by using different models depending on
the space of possible actions, available data, and our knowledge of the
business constraints. One of the key considerations in model design is
the level of data aggregation. Classic economic models often operate
with a small number of high-level aggregates, such as the overall de-
mand. This makes the models simple from the computational and data
collection perspectives but limits their ability to model complex depen-
dencies. Algorithmic methods assume the availability of a powerful
data processing infrastructure and high-resolution data, which enables
more granular modeling. The difference between these two approaches
can be illustrated by the following simplistic example [Kleinberg et al.,

1998]. Suppose that a retailer sells a product with the margin m, and
qu is the monthly amount of this product purchased by customer u.
The monthly revenue can therefore be described as
¸
G qu m (2.2)
u
The retailer wants to run a promotion campaign that boosts sales by

the factor of k, and the cost of each promotion is c. The retailer can
control both k and c by choosing a more or less aggressive promotional
strategy. Consequently, the optimization problem to be solved can be
defined as
¸
max k qu m c (2.3)
s
u
in which s is the promotional strategy determined by the pair of pa-

rameters k and c. We can see that the definition in terms of individual
consumers is redundant, and the problem can be redefined in terms of
aggregates, namely the total demand
¸
Q qu (2.4)
u
and the total budget of the campaign C. Thus, the problem is defined
as
max kQmC (2.5)

s
Let us now assume that the retailer wants to create two different
consumer segments and assign the strategy si pki , ci q to one of
these segments and the strategy sj pkj , cj q to the other segment.
With the assumption that the strategies are selected from domain S,
the optimization problem will be
¸
max maxt qu ki m ci , qu kj m cj u
si , sj PS u
(2.6)
This expression is not linear with regards to the k and c parameters,

so we cannot easily redefine it in terms of aggregates. Consequently,
more sophisticated data mining techniques might need to be used. This
trade-off between aggregates and high-resolution data is a common
pattern that arises in many problems because of nonlinear dependen-
cies between the variables involved in the optimization.
Finally, it should be noted that optimization problem 2.1 as a whole
is somewhat dependent on time because of environmental changes
(new products appear on the market, competitors make their moves,
2.3 machine learning 23
etc.) and the enterprise’s own actions. One possible approach for han-
dling this dependency is to use a stateless model, treating it as a math-
ematical function, but allow for time-dependent arguments to account
for memory effects. For example, a demand prediction model can fore-
cast the demand for the next month by taking the discount levels for
the last week, last two weeks, last three weeks, etc. as arguments.
2.3 machine learning
We stated in the previous section that the optimization objective can

be defined as a function of the data and marketing strategy, Gps, Dq.
Our next step is to provide a more formal definition of data that will
help to bridge the gap between the economic model and data mining
methods.
The first thing we note is that the economic modeling process is
concerned only with certain metrics of the company or consumer that
are directly related to the modeled objective. Examples of such metrics
are the demand for a certain product or the propensity of a consumer
to respond to a promotion. In most cases, the marketing strategy and
actions do not unconditionally determine these metrics but only influ-
ence them. A price discount, for instance, can increase the demand for
a certain product, but this may not be the case if a similar discount
is simultaneously offered by a competitor. Consequently, we are inter-
ested in finding the functional dependency between the controlled or
uncontrolled factors and the metrics of interest. In probabilistic terms,
this can be expressed as the conditional distribution
ppy | xq (2.7)
in which x is the vector of factors and y is the metrics. In the example

with discounts, x can include variables like the price of a given product,
prices of related products, and competitor prices, and y is the demand
measured in units sold. Each marketing strategy s then corresponds
to a particular combination of factors, that is, some value of vector x,
which we denote as xpsq. With the assumption that the distribution
is known, the economic optimization problem can be rewritten in the
terms introduced above as
max Gpppy | xpsqqq (2.8)
s
The data come into play if the conditional distribution ppy | xq is com-
plex and should be learned from the data as opposed to being manu-
ally defined. Consequently, we are interested in the data that contain
pairs px, yq that are drawn from the true but unknown distribution
pdata py | xq. We will refer to these pairs as samples or data points. The
input data are often observed (collected) in a form that is not suitable
or optimal for modeling, so vector x and metric y are typically con-
structed from the raw data by using cleansing and normalizing trans-
formations. We will refer to the elements of such a prepared vector x
as features or independent variables and to y as the response label or de-
pendent variable. With the assumption that we have n samples and m
features, all feature vectors can be represented as an n m matrix, X,
called the design matrix, and all response variables as an n-dimensional
column-vector y. Each row of the design matrix X is a feature vector x
and each element of y is a response label y. All data points can then be
represented as an n pm 1q matrix, D, with the following structure:

x1 y1

x2 y2
D rX | ys

.. .. (2.9)

. .
xn yn
Our goal then is to create a statistical model that approximates the
true distribution pdata py | xq with the distribution pmodel py | xq learned
from the data, so the economic model will actually be evaluated by
using the following approximation:
max G ppmodel py | xpsqqq (2.10)

s
In many practical applications, we do not need to specify the distri-

bution but rather need to estimate the most likely value of y based on
x, that is, to learn the function
p ypxq
y (2.11)
in which the left-hand side is the estimated value of the response vari-
able. Estimation of the most likely value of y is easier than estimation
of the entire distribution and, as we will see shortly, it can be done with-
out accurate estimation of the actual probability values. The economic
model can then be evaluated as
max G pypxpsqqq (2.12)
s
In a general case, model G can use multiple data models derived

from one or more data sets. This approach divides the original mod-
eling problem into the following smaller tasks that can be explored
separately:
• The distribution ppy | xq should be estimated from the data. This
is a standard machine learning problem that falls into the class
of supervised learning tasks.
2.4 supervised learning 25
• In certain cases, appropriate x and y values are not explicitly

present in the available data. However, it may be possible to find
a transformation that maps the original data to a new represen-
tation that is more suitable for modeling. This task is known as
feature engineering. In some cases, feature engineering can be done
semimanually by using relatively basic methods. For example, it
may be found that taking a logarithm of the input values im-
proves the accuracy of the model. In other cases, one needs to
use more advanced machine learning methods to find a proper
representation. We refer to this problem as representation learning.
• Finally, the economic model that estimates the business outcomes

from the distribution should be defined. This is an economic
problem, rather than a machine learning one, so we will discuss it
in detail in the subsequent chapters dedicated to particular mar-
keting problems. However, there exist a number of standard mod-
els for fundamental problems, such as prediction of consumer
choice behavior, that we will touch on later in this chapter.
We proceed with an overview of each of these three areas to describe
the toolkit that is available for developers of applicable programmatic
systems and to define a vocabulary of the building blocks that we will
use throughout the book. Herein, we will focus on the conceptual prob-
lems and solutions and do not dive into the algorithmic and implemen-
tation details, which can be found in textbooks on machine learning,
such as [Bishop, 2006], [Murphy, 2012], and [Zaki and Meira, 2014], to
name a few.
2.4 supervised learning
Previously, we have seen that the modeling task can be partially re-
duced to learning of the distribution ppy | xq based on the available
samples x and y. The function p, which maps an m-dimensional vector
of features to the probability values, can be interpreted as a probability
density function for continuous y or a probability function for discrete
y. In many applications, we do not need to learn the entire distribu-
tion but only a function that predicts the most likely y response based
on the input x. The task of learning such distributions or functions is
known as supervised learning because the data contain the response
variables that “guide” the learning process. This problem comes in two
types. If the response variable is categorical, that is, y belongs to some
finite set of classes, the problem is known as classification. If the response
variable is continuous, the problem is known as regression.
In this section, we will first discuss how the problem of estimating

the distributions and predictive models can be approached and the
relationship between the true distribution pdata py | xq and the learned
distribution pmodel py | xq, and we will then consider a few examples of
how the actual models can be built.
2.4.1 Parametric and Nonparametric Models
One of the main considerations in predictive model design is the choice

between parametric and nonparametric approaches. The parametric
approach makes the assumption that the data distribution has a cer-
tain functional form specified by a fixed number of parameters, so the
problem of distribution estimation can be redefined as the problem of
model fitting, that is, the selection of model parameters θ so that the
distribution model
pmodel py | x, θq (2.13)
can be optimally fitted with the data. The optimality condition, of

course, needs to be formally specified as well. The nonparametric ap-
proach assumes that the number of parameters can grow with the
amount of training data, and, in some methods, each data point can
be viewed as a parameter. One of the most commonly used nonpara-
metric methods is the k nearest neighbor (kNN) algorithm. The idea is
to estimate the response variable for the feature vector x based on the
response variables of the data samples that are the nearest neighbors
of x in the feature space. In a case of classification, the probability that
the response variable y belongs to the class c can then be estimated as
¸
Prpy c | x, kq Ip y i cq
1
(2.14)
k
i P Nk pxq
in which k is the algorithm parameter that defines the neighborhood
size, Nk pxq is the neighborhood of k nearest data points in the training
data set, and I is the indicator function equal to 1 if its argument is
true and to 0 otherwise. The neighborhood of the input vector x can
be determined by using any vector distance metric, such as Euclidean
distance. The classification decision can then be made by choosing the
most probable class as follows:
p argmax Prpy c | x, kq
y (2.15)
c
This process is illustrated in Figure 2.1. A regression model can be

defined in a similar way by estimating the response as the average
values of response variables in the neighborhood.
Figure 2.1: Classification of two-dimensional points by using the kNN algo-

rithm. (a) Training data set with two classes of points. Each point

in the space is classified based on k 4 nearest neighbors. (b) Class
probabilities estimated as functions of the features according to equa-
tion 2.14.
The nearest neighbor algorithm is one of the most basic supervised

learning methods; however, it can work quite well in many settings. For
example, it is widely used in recommendation algorithms, as we will
discuss in the appropriate chapter. The shortcoming of nonparamet-
ric methods is that the higher the dimensionality of data, the sparser
the space becomes, and we then have to look at neighbors that are so
far away from the given point that they do not really predict the de-
pendency between inputs and outputs in the required region. In other
words, the model uses only local observations and cannot generalize
the patterns observed through the data set. This problem can be allevi-
ated by using parametric models, which provide less flexibility because
of the limited number of parameters but which learn these parameters
globally.
2.4.2 Maximum Likelihood Estimation
Model fitting is an optimization problem in its own right, so we need to

specify an objective function that will be optimized. Let us assume that
a set of n samples is drawn independently from the data distribution
pdata px, yq. Each data example pxi , yi q can be interpreted as the input
xi that results in the output yi . The objective function can then be
defined as the probability of the observed response data given that the
probability density specified by the parameter vector θ is known:
Lpθq pmodel py | X, θq (2.16)
This function is referred to as the likelihood function or, simply, the like-
lihood. It is the probability of observing the training data under the
assumption that these data are drawn from the distribution specified
by the model with parameters θ. The logarithm of the likelihood func-
tion, often more convenient for analysis and calculations, is known as
the log-likelihood LLpθq. Our goal is to find the parameter vector that
maximizes the likelihood of the estimation:
θML argmax log pmodel py | X, θq (2.17)
θ
By assuming that the examples are independent and identically dis-

tributed, we can split the likelihood probability into a product of n
probabilities estimated for individual samples:
ņ
LLpθq log pmodel py | X, θq log pmodel pyi | xi , θq (2.18)

i 1
We can divide the equation by n because the argmax operator is

indifferent to rescaling. θML is then expressed in terms of mathematical
expectation over samples:
LLpθq Ex,ypdata r log pmodel py | x, θq s (2.19)
Now we can show that the maximum likelihood principle leads us to

minimization of the divergence between the data generation distribu-
tion and its estimation. The standard measure of divergence between
two distributions is the Kullback–Leibler divergence, commonly abbre-
viated as KL divergence, which is defined as

pdata py | xq
KLppdata , pmodel q Ex,ypdata
pmodel py | x, θq
log (2.20)
As pdata does not depend on θ and cannot be the subject of the opti-
mization, we only have to minimize the second term in order to min-
imize the divergence, which is equivalent to maximization of the log-
likelihood defined in equation 2.19:
argmin KLppdata , pmodel q argmin LLpθq (2.21)

θ θ
The maximum likelihood can thus be viewed as optimization of the

model parameters such that the model distribution matches the empir-
ical distribution.
2.4.3 Linear Models
The maximum likelihood principle provides a common framework that

can be used to derive applicable algorithms for the creation of data
models. We will now describe how this principle can be used to build
several basic but very useful predictive models. We will first look at
the regression problem and derive a model that predicts the response
y as a continuous linear function of the input x. Next, we will turn
to the classification problem and discuss several models that predict
a categorical response by fitting a hyperplane that splits the feature
space into areas, each of which corresponds to some response class.
Figure 2.2: (a) Example of linear regression with one-dimensional feature space.
(b) Example of classification with a linear decision boundary and
two-dimensional feature space. Training points are shown as circles.
All of the models that we will consider are linear. The regression
model is linear in the sense that the dependency between the features
and response is modeled as a linear function, as shown in Figure 2.2.
If the observed dependency is not actually linear, the model may not
be accurately fitted to the data. The classification models are linear
in the sense that the boundary between the classes is modeled as a
hyperplane, so the data are linearly inseparable; that is, if the groups
of points cannot be accurately separated by a hyperplane, the models
may not be properly fitted to the data.
2.4.3.1 Linear Regression

The goal of a regression model is to map a continuous input x to a
continuous output y. Linear regression models map this as a linear
function of the input:
ypxq wT x (2.22)
in which w is the vector of model parameters to be learned. Thus, the

estimation error will be
y ypxq y wT x (2.23)
With the assumption that the error has a normal distribution, the dis-
tribution of the estimate produced by the model is also described by a
normal distribution, that is, the Gaussian with the mean wT x and the
variance σ2 :

ppy | x, wq N y | wT x, σ2

1 2
(2.24)
1
2σ12 yi wT xi
2
exp
2πσ2
By inserting this probability distribution into the definition of the log-
likelihood defined in equation 2.18 and doing some algebra, we get the
following log-likelihood expression that has to be maximized:
ņ
LLpwq log ppy | x, wq

i 1
ņ
log N y | wT xi , σ2 (2.25)

i 1
2σ12 SSEpwq n2 logp2πσ2 q

in which SSE is the sum of squared errors, defined as
ņ 2
SSEpwq yi wT xi (2.26)

i 1
It is an intuitive result – maximization of the likelihood is equivalent

to minimization of the estimation error. By assuming that the variance
is fixed to some constant parameter σ2 (in the general case, it can also
be estimated), we can discard the second term in log-likelihood equa-
tion 2.25, so maximization of the log-likelihood can be performed with
respect to the SSE only. We first rewrite the likelihood expression in a

more compact vector form:
LLpwq py XwqT py Xwq

1
2 (2.27)
wT pXT yq 21 wT pXT Xqw 21 yT y
in which
ņ ņ
XT X xi xTi and XT y xi yi (2.28)

i 1
i 1
and we then take the gradient with respect to w
∇w LLpwq XT y XT Xw (2.29)
By equating this gradient to zero and solving the equation for w, we

obtain an estimation for the ML-optimal w in a closed form:
wML pXT Xq1 XT y (2.30)
Linear regression is the most basic method of predictive modeling, but

the way we have derived it here is a good illustration of how model
fitting can be performed based on the maximum likelihood principle.
We will show later that methods that are able to capture nonlinear
dependencies can be derived from linear regression.
2.4.3.2 Logistic Regression and Binary Classification

The second example we consider is how the maximum likelihood prin-
ciple can be used to build a binary classification model, that is, a model
that maps the input x to one of two possible classes y P t0, 1u. We fol-
low the same path that we used for linear regression and first define
the form of the model. The goal is to find a linear decision boundary
(hyperplane) that separates two classes at the point where
Prpy 0 | xq Prpy 1 | xq (2.31)
We can rewrite this equation in a logarithmic form as follows:

Prpy 0 | xq
log
Prpy 1 | xq
0 (2.32)
As we are looking for a linear boundary, the boundary hyperplane can

be described by a linear function of the input x with the coefficient
vector w:
Prpy 0 | xq
log
Prpy 1 | xq
wT x (2.33)
This means that we assign the input to class 0 if wT x is positive and to

class 1 if wT x is negative. Equation 2.33 is equivalent to
Prpy 1 | xq gpwT xq
1
1 exppwT xq
(2.34)
exppwT xq
Prpy 0 | xq 1 g pw xq
T
1 exppwT xq
in which g is known as a logistic function. This model is called a lo-
gistic regression. Note that this is a classification model, despite the
confusing name. Next, we have to calculate the log-likelihood for this
distribution:
ņ
LLpwq log ppyi | xi q

i 1
ņ 1yi
log gpwT xi qyi 1 gpwT xi q (2.35)

i 1
ņ
yi log gpwT xi q p1 yi q log 1 gpwT xi q

i 1
We can calculate the gradient of this expression, but, unfortunately,

we cannot obtain the optimal solution for w in closed form by equat-
ing this gradient to zero and solving the equation with respect to w.
So numerical methods, such as gradient descent, should be used to
maximize log-likelihood equation 2.35 and estimate the optimal weight
wML .
Logistic regression models the class probabilities by using a logis-
tic function defined by equations 2.34. This function represents an S-
curve, also known as a sigmoid curve, with the steepness controlled by
parameter w. Examples of logistic functions are shown in Figure 2.3
where two curves are fitted on the data. Note that the decision bound-
ary between the classes, that is, the intersection of the two surfaces, is
a straight line. This illustration uses the same data set as our previous
example with the kNN classifier in Figure 2.1, so it is interesting to com-
pare the probability surfaces. The surfaces of the kNN classifier have a
more complex shape than those of the logistic regression because the
kNN algorithm is a nonparametric method. The surfaces fitted by lo-
gistic regression have a much more simple shape determined by the
logistic curve.
Logistic regression is one of the simplest classification methods that
is unable to capture nonlinear boundaries between the classes. It is
possible, however, to extend it and other linear methods in such a way
that nonlinear decision boundaries can be captured, as we discuss later.
Figure 2.3: Classification of two-dimensional points by using logistic regression.

(a) Training data set with two classes of points. (b) Class probabilities
estimated as functions of the features according to equation 2.34.
2.4.3.3 Logistic Regression and Multinomial Classification

Logistic regression can be straightforwardly extended for cases with
multiple classes. As we have multiple classes, we cannot use the conve-
nient relationship
Prpy 0 | xq 1 Prpy 1 | xq (2.36)
used in the binary case and one linear decision boundary is not suf-
ficient. Instead, we can estimate the probability of each class c sep-
arately by using a dedicated coefficient vector wc . This means that
equation 2.33 can be rewritten as follows:
log Prpy 0 | xq wT0 x log Z

log Prpy 1 | xq wT1 x log Z
.. (2.37)
.
log Prpy c | xq wTc x log Z
in which log Z is a normalizing term that ensures that the resulting

distribution over y is, in fact, a probability distribution that sums to
one. The role of Z becomes more clear when equation 2.37 is written
in the exponential form:

Prpy c | xq
1
exp wTc x (2.38)
Z
The normalizing factor thus scales the probability distribution. We

can estimate Z from the fact that all class probabilities must sum up to
one
¸
Pr py c | xq 1 (2.39)
c
and, consequently, the normalizing factor can be found by substitut-

ing 2.38 into 2.39:
¸
Z exp wTc x (2.40)
c
By substituting this result back into equation 2.38, we obtain a for-

mula for the class probability estimation:

exp wTc x
Prpy c | xq ° T
(2.41)
i exp wi x
This probability can be used to define the log-likelihood, and the co-
efficient vectors wi can then be estimated by using numerical methods
such as gradient descent.
It is important to note that resultant equation 2.41 can be interpreted
as a generic method for mapping a vector of real values to a vector of
class probabilities, regardless of the underlying model that produces
these values. To see this, let us assume that some model, not necessar-
ily a linear one, produces a vector of values v, where each value can
be interpreted as a relative weight of the corresponding class. These
weights are not necessarily normalized as probabilities: values v can be
out of the r0, 1s range or may not add up to one. To map these weights
to normalized class probabilities, let us define a generic function, com-
monly known as the softmax function, that repeats equation 2.41 with
a vector of real values as a parameter:
exp pvi q
softmaxpi, vq ° (2.42)
j exp vj
The normalized class probabilities can then be acquired by passing the

class weights v through the softmax function. We will use this property
in the later chapters to construct some predictive models.
2.4.3.4 Naive Bayes Classifier

The last model that we consider in this section is the Naive Bayes classi-
fier. This method is widely used for text classification, so it will be very
helpful later in building search and recommendation services. Recall
that a classification problem with multiple classes can be defined as
p argmax Prpy c | xq
y (2.43)
c
By applying the Bayes rule to the conditional probability of class c, we

get the following equation:
Prpx | y cqPrpy cq
Prpy c | xq
Prpxq
(2.44)
The probability of the feature vector in the denominator is the same for
all classes, so it can be discarded for a classification problem:
p argmax Prpy cqPrpx | y cq

y (2.45)
c
The key assumption made by the Naive Bayes classifier is that each
feature xi is conditionally independent of every other feature, given
the class c. This means that the probability of observing the feature
vector x, given the class c, can be factorized as
¹
m
Prpx | y cq Prpxi | y cq (2.46)

i 1
in which m is the length of the feature vector. This assumption, called

the conditional independence assumption, is rarely true in practical appli-
cations, but the Naive Bayes algorithm performs reasonably well in a
wide range of such cases. For example, this classifier is successfully
used for text classification when each feature is a word, although the
words in the text are not independent. This can be explained by the fact
that the Naive Bayes algorithm can still be correct if the features are
dependent, but the dependencies have a certain structure and cancel
each other out [Zhang, 2004]. By using the independence assumption,
the classification problem can be rewritten as
¹
m
p argmax Prpy cq
y Prpxi | y cq (2.47)
c
i 1
The parameters of this model are the values Prpy cq and

Prpxi | y cq. One possible approach to fit the model is to treat
these values simply as unknown variables, rather than probabilities,
and to maximize the log-likelihood that corresponds to equation 2.47.
However, it can be easily shown that this leads to exactly the same
results as if we treat the parameters as empirical probabilities. In other
words, the maximum likelihood estimate of y p in expression 2.47 can
be obtained if Prpy cq is estimated as the frequency of class c in
the training data set and Prpxi | y cq is estimated as the frequency
of samples that belong to class c and have the feature value xi . This
makes it easy to fit the Naive Bayes model in practice.
In the general case, the Naive Bayes classifier is not linear. However,
it is linear under certain assumptions that are accurate in many appli-
cations, so it is often described as linear. For example, let us consider a
case where the distribution Prpx | y cq is assumed to be multinomial.
This is a valid assumption, for example, for text classification when
each element of the feature vector is a word counter. The probability
of a feature vector follows a multinomial distribution with parameter
vector qc
¹
m
Prpx | y cq 9 qx
ci
i
(2.48)
i 1
in which qci is the probability of the feature value xi occurring in class
c. We can rewrite this expression in the vector form as follows:
log Prpx | y cq xT log qc constant (2.49)
We can now show that the decision boundary between the classes
is linear by considering the ratio of class probabilities, similarly to the
method we used for logistic regression. By assuming only two classes
y P t0, 1u for the sake of simplicity, we can write
Prpy 1 | xq
log
Prpy 0 | xq
log Prpy 1 | xq log Prpy 0 | xq
(2.50)
x plog q1 log q0 q
T
log Prpy 1q log Prpy 0q
This is a linear function of x, which means that the decision boundary

between the classes is linear.
2.4.4 Nonlinear Models
Linear methods can be an appropriate solution for many marketing

applications and the power of these methods should not be underes-
timated, but they can perform poorly on data sets with nonlinear de-
pendencies. We need to develop methods that are able to model more
complex distributions. This problem can be approached from different
perspectives, and we discuss here two major families of methods that
are often used in practice. We will discuss a few more methods, such as
neural networks, later in the book in the context of specific application
areas.
2.4.4.1 Feature Mapping and Kernel Methods

The linearity or nonlinearity of the regression or classification problems
that we are trying to solve reflects the nature of the process that we are
modeling, but we should also realize that data representation can con-
tribute to this. For example, two linearly dependent values would not
be linearly dependent if one of them was measured on a logarithmic
scale. The opposite is also true – data sets that are not tractable for
linear methods can be become tractable if mapped to a different space.
Consider the example in Figure 2.4: the one-dimensional data set on
the left-hand side consists of two classes that are not linearly separa-
ble, but the two-dimensional data set constructed from the first one by
using the mapping pxq Ñ px, x2 q is linearly separable.
Figure 2.4: Feature mapping by using a quadratic function. (a) Original data
points. (b) Mapped data points.
This transformation of the original feature space into another space,

typically of higher dimensionality, is referred to as feature mapping. It
is intuitively clear that the addition of more dimensions, specified as
nonlinear functions of one or several existing features, provides more
flexibility for a regression or classification algorithm that we are trying
to improve. However, we need a method to define the mapping func-
tion φpxq that produces a new feature vector of higher dimensionality.
The first thing we have to note is that many regression and classi-
fication methods can be expressed in terms of distances between the
input vector and the training vectors. For example, we have shown that
coefficients of linear regression can be calculated as
w pXT Xq1 XT y (2.51)
We can multiply this expression by the identity matrix I, which can be
expressed as
I XT XpXT Xq1 (2.52)
and, by doing some algebra, we get
w XT XpXT Xq1 pXT Xq1 XT y

ņ
(2.53)
XT a ai xi

i 1
in which the vector a is defined as
a XpXT Xq2 XT y (2.54)
This means that we can estimate the response variable by using only
dot products between the input x and training vectors xi :
ņ
ypxq wT x ai xT xi (2.55)

i 1
This is a very important result because we can now avoid the explicit
feature mapping and replace it with a kernel function that encapsulates
computation of the dot product in the mapped space:
kpx, zq φpxqT φpzq (2.56)
in which x and z are two feature vectors. Equation 2.55 can then be
rewritten purely in terms of the kernel function:
ņ
ypxq ai kpx, xi q (2.57)

i 1
In other words, we simply modify the distance function between the

feature vectors, as opposed to using the dot product in expression 2.55.
Thus, the input vectors do not need to be mapped, but the algorithm
should be modified to use the kernel function instead of the dot prod-
uct. The relationship between the kernel function and the mapping
function can be illustrated by using a quadratic kernel:
kpx, zq pxT zq2 (2.58)
If the original feature space is two dimensional, the kernel expands

it into a three-dimensional space that contains both derivatives of the
individual features and cross-products of features that can make the
data set tractable for linear methods:
kpx, zq pxT zq2 px1 z1 x2 z2 q2
x21 z21 x22 z22 2x1 x2 z1 z2
? ? (2.59)
px21 , 2x1 x2 , x22 qpz21 , 2z1 z2 , z22 q
φpxqT φpzq
in which
! ? )
φpxq x21 , 2x1 x2 , x22 (2.60)
Note that the kernel is simply a distance function between the original
feature vectors, so the underlying expansion of dimensionality is to-
tally hidden; therefore, we can devise kernels that correspond to φpxq
with a very high or infinite number of dimensions but remain com-
putationally simple. This technique is known as the kernel trick, and a
number of machine learning methods can be extended this way. Selec-
tion of the right kernel can be a challenge, but there are a few kernel
functions that are known to be quite universal and are widely used
in practice. The choice of the kernel function also depends on the ap-
plication because it is essentially a measure of similarity between the
feature vectors – kernels that work well for consumer profiles might
not be the best choice for textual product descriptions, and so on.
Among the best known members of the kernel methods family are
support vector machines (SVMs). The basic SVM algorithms are linear
classification and regression methods, but they can be efficiently ker-
nalized to learn nonlinear dependencies. Consider the example of an
SVM classifier in Figure 2.5. It uses the same data that we used for
the nearest neighbor and logistic regression examples earlier, but its
decision boundary is clearly nonlinear.
Figure 2.5: Classification of two-dimensional points by using support vector ma-

chines with a nonlinear kernel. (a) Training data set with two classes
of points. (b) Class probabilities as a function of features.
2.4.4.2 Adaptive Basis and Decision Trees

The shortcoming of the kernel method is that the kernel function has
to be specified as a parameter instead of being learned. The kernel
function imposes limitations on the shape of the decision boundary,
and, although we can try several kernels or kernel parameters to find
a good fit in practice, we may sometimes be better off by choosing a
different approach. We can state the problem of learning a set of q basic
functions φi pxq so that the response can be predicted as their weighted
composition:
¸
q
ypxq wi φi pxq (2.61)

i 1
Provided that the basis functions are highly adaptive and nonlinear,
we can expect this solution to outperform linear methods in appropri-
ate tasks. One of the most widely used realizations of the adaptive
basis concept is classification and regression trees, a family of meth-
ods that generate the adaptive basis φi pxq by using a greedy heuristic
algorithm. Let us consider the most basic version of this solution.
A classification or regression tree is created by recursive splitting of
the feature space into two parts by using a linear decision boundary,
as illustrated in Figure 2.6. In each step of the recursion, the decision
boundary can be selected as follows:
• First, the candidate hyperplanes that can be chosen as the bound-
ary are enumerated. One possible way is to try all dimensions
(e. g., we can split by using either a horizontal or vertical line in
Figure 2.6) and, for each dimension, to try the coordinates of all
data points in the training set.
• The candidate boundary produces two regions, each of which

can be labeled with the most frequent class of examples in this
region or, in the case of regression, with the mean value of the
response variables of the examples in the region. This label is
then used as a predicted value for any data point that falls into
the region.
• The region label is used to score the quality of the candidate split
by the misclassification rate (the ratio between the number of in-
correctly and correctly classified training examples in the region)
or another metric. The boundary is selected as the candidate with
the highest score.
Once the boundary is selected, the algorithm is recursively applied
to the regions on both sides of the boundary. This algorithm produces
2.5 representation learning 41
Figure 2.6: Example of a classification tree. (a) Classification tree model. (b)
Training data and decision boundaries that correspond to the tree.
a set of labeled rectangular regions Ri that can be considered as an

adaptive basis. To see this, we can rewrite equation 2.61 in terms of the
regions as
¸
q
ypxq wi I px P Ri q (2.62)

i 1
in which wi is the region label and I is either 1 or 0, depending on

whether x falls into the corresponding region or not. Decision trees
and more complex derivatives of this approach, such as random forests,
provide a powerful and widely used regression and classification solu-
tion.
2.5 representation learning
The supervised learning methods we have just discussed help to de-

scribe the dependency between the input independent variables and
the response variable. The input variables, however, can have an entan-
gled structure and redundancies that complicate data exploration and
model training. We can attempt to find a different representation of
the data that better suits the modeling purposes by removing redun-
dancies and correlations, that is, by disentangling the original data.
Machine learning methods are typically categorized as supervised
and unsupervised. Supervised methods are concerned with dependen-
cies between the input variables and responses, that is, conditional
probability densities ppy | xq. The goal of unsupervised methods is

to learn the structure or patterns in the input data, that is, to model
the unconditional density ppxq. Unsupervised learning makes no as-
sumptions about response variables to guide the model training and
analyzes only the design matrix to find the dependencies between the
features or samples. A canonical example of unsupervised learning is
clustering, which can be defined as the grouping of data samples into
clusters such that the samples in one cluster are similar to each other
and samples from different clusters are dissimilar – this task does not
require response variables but relies on a sample similarity measure.
Unsupervised learning is widely used in marketing applications for
data exploration and analysis. Clustering of customer profiles and in-
terpretation of the obtained results, for example, is one of the most
important techniques in marketing analytics. In programmatic appli-
cations, however, we are more concerned with automation than with
exploration and interactive analysis. Representation learning is one ap-
plication of unsupervised learning methods that can be useful in this
context, so we are focusing here specifically on representation learning
aspects, rather than unsupervised learning in general.
2.5.1 Principal Component Analysis
Principal component analysis (PCA) is a powerful technique for find-

ing a condensed uncorrelated data representation. PCA is a mathe-
matical method that transforms the data into a new presentation that
guarantees certain properties and also produces artifacts that describe
the structure of the data. We may be interested in different properties
of the PCA transformation, depending on the application, so we will
discuss these properties sequentially in the next sections, although it
is important to keep in mind that all of them are based on the same
algorithm.
2.5.1.1 Decorrelation
In marketing applications, the data typically correspond to the ob-
served inputs, properties, and outputs of some real-world marketing
process. Examples of such processes include marketing campaigns, in-
teractions between customers and products, and the interplay of price
and demand, among many others. Each feature in the design matrix
can be viewed as a signal that carries the information about the pro-
cess. We do not have complete knowledge of the process and observe
only certain projections of the process on the feature dimensions that
are available in the input data, just as a physical object can be pho-
tographed by cameras from different perspectives. For example, one
does not observe consumer tastes and thoughts directly but captures
certain signals, like purchases, that partially reflect the tastes, thoughts,
and decisions. Representations obtained this way are likely to have
some redundancy, and dimensions are likely to be correlated, just as
images of the same object from different perspectives are redundant
and correlated. This idea is illustrated in Figure 2.7.
Figure 2.7: Different data dimensions can be correlated because they are projec-
tions of the same real-life process.
We can pose the problem of finding a new, potentially smaller, set of

features that are statistically independent and, consequently, provide
a less redundant and more structured data representation. PCA can
be applied to this problem under the assumption that the statistical
independence of features can be replaced by zero correlation, which
may or may not be true, depending on the data distribution. This is
a restrictive assumption because the feature values have to be jointly
normally distributed to ensure statistical independence with zero cor-
relation. If the distribution is different, the PCA may fail to achieve the
decorrelation goal (but can still be useful because of other properties).
Let us consider an n m design matrix X, in which n is the number of
samples and m is the number of features. We also assume that the data
are centered, that is, E rxs 0. If this is not the case, we can subtract
the mean from all data samples. The decorrelation problem can then
be defined in terms of the covariance matrix C:
CX Var rxs
1
XT X
n1
(2.63)
This is a square symmetric m m matrix in which the diagonal el-

ements are the variance of the corresponding features and the off-
diagonal elements are the covariance between the features. The fea-
tures are decorrelated if the covariance matrix is diagonal, that is, all
off-diagonal elements in C are zeros. If the covariance matrix is not
diagonal, the features are correlated, which makes it more difficult

to understand the distribution of x because it cannot be described in
terms of distributions of individual features. Our next step is to find
the transformation of X that produces a different n m design matrix
Z in which the covariance matrix is diagonal. In the PCA approach, we
make an assumption that this matrix can be obtained by using a linear
transformation
Z XT (2.64)
in which T is the transformation matrix. To construct this matrix, let

us first show how the design matrix can be factorized on the basis of
vectors that correspond to the directions of maximum variance in the
design matrix.
The first step of factorization is to find the directions of maximum

variance in the data. These directions can be thought of as the main
axes of the point cloud in a scatter plot, as illustrated in Figure 2.8. Each
of these directions can be defined as a vector, so we start by finding an
m-dimensional unit vector that satisfies the following condition:
v1 argmax kXvk2 , kvk 1 (2.65)

v
This vector corresponds to the axis with maximum variance. Next, we

find a second unit vector that is orthogonal to the first one to capture
the remaining variance:
v2 argmax kXvk2 , kvk 1 and v v1 0 (2.66)

v
We continue this process by requiring each vector to be orthogonal to

all previous ones, and, assuming that the design matrix has rank r, we
can create as many as r non-zero vectors v. By construction, each vec-
tor corresponds to the axis with maximum remaining variance in the
design matrix. These vectors are referred to as the principal components
of the design matrix.
Let us denote an m r matrix assembled from column vectors v as
V. As all of the unit vectors that we have constructed are orthogonal,
this matrix is column-orthonormal, that is
VT V I (2.67)
Unit vectors v capture the directions of the variance but not its magni-
tude. Let us calculate these values separately and denote them as
σi kXvi k (2.68)
Figure 2.8: Example of principal component analysis. (a) A data set of 500 nor-
mally distributed points and the corresponding principal compo-
nents. Features x1 and x2 are strongly correlated. (b) A decorrelated
representation obtained by using PCA. Features z1 and z2 are not
correlated.
Each principal component captures only the remaining variance, so the

value of σ is biggest for the first component and then decreases, in the
order
σ1¥ σ2 ¥ . . . ¥ σr (2.69)
Let us denote an r r diagonal matrix with the σ values on the main

diagonal as
Σ diagpσ1 , . . . , σr q (2.70)
At this point, we have an orthonormal basis of vectors V and the

corresponding scaling factors Σ. To complete the factorization of the
design matrix, we need a third factor that projects the design matrix
onto the principal component basis or, alternatively, mixes the basis
into the design matrix. By denoting this factor as U, we can define the
decomposition as
X UΣVT (2.71)
for which U can be obtained by solving equation 2.71 with respect to

this factor:
U XVΣ1 (2.72)
The decomposition defined by equation 2.71 is known as the singular

value decomposition (SVD). To summarize, the principal component,
that is, the columns of matrix V, can be interpreted as orthogonal prin-
cipal axes that are aligned with the directions of variance. The columns
of UΣ can be interpreted as coefficients that mix the principal vectors to
produce X. It can also be shown that matrix U is column-orthonormal:
UT U I (2.73)
Once the SVD of the design matrix is obtained, we can use the factors
to find the decorrelating linear transformation T. Let us consider the
product
Z XV (2.74)
and calculate its covariance matrix by using the fact that matrices V
and U are orthogonal:
CZ n 1 1 ZT Z
n 1 1 VT XT XV
(2.75)
n 1 1 VT VΣ2 VT V
n 1 1 Σ2
As Σ2 is diagonal, the representation Z is uncorrelated. This means
that the decorrelating transformation we were looking for is actually
given by matrix V. This transformation is effectively a rotation because
the matrix of principal components is orthonormal. Note that the as-
sumption about the linearity of the transformation is quite restrictive.
In the example provided in Figure 2.8, it works well because the data
set has an elliptical shape, which is the result of the normal distribution;
therefore, the correlation between features can be removed by simple
rotation. This may not be the case for data sets with more complex
shapes.
2.5.1.2 Dimensionality Reduction

The key property of PCA is that the principal vectors are sorted by
the variance magnitude. This property is important because one can
argue that high-variance dimensions in the data are typically more in-
formative and carry a stronger signal than low-variance dimensions.
For example, we can say that axes x1 and x2 in Figure 2.8 are equally
important and the one-dimensional representations obtained by pro-
jecting the data points on x1 or x2 are not good approximations of
the original two-dimensional representation. At the same time, dimen-

sions z1 and z2 of the decorrelated data set obtained by using the PCA
transformation are not equally important and dimension z2 can be dis-
carded with a relatively small loss of information.
This property can be used in several ways. The first application is
dimensionality reduction, in which the m-dimensional data set X needs
to be reduced to a k-dimensional data set and k m. In the case of
PCA, this reduction can be done by using a truncated matrix Vk that
includes only the first k principal component vectors and calculating a
new data representation as
Zk XVk (2.76)
Matrix Zk has only k columns, which correspond to the first principal

axes. This approach is often used for data visualization to project a
multidimensional data set to a two- or three-dimensional space that
can be shown on a plot.
The second important application is low-rank approximation of the
design matrix. Consider the SVD given by expression 2.71. Similarly to
dimensionality reduction, we can remove the least significant columns
of matrix V to create the truncated version Vk . The factors U and Σ
then also need to be cropped to remove the columns and rows that
correspond to the least significant components, as shown in Figure 2.9.
Let us denote these truncated versions as Uk and Σk , respectively. The
design matrix X can then be approximated from the product of these
truncated factors:
p
X k Uk Σk Vk (2.77)
Figure 2.9: Dimensionality reduction by removing axes with low variance.
p is of the same size as X, but, of course,

The reconstructed matrix X
has some approximation error because of the discarded dimensions.
This error is smallest if we discard only the least significant dimen-

sion and increases as we remove more columns from the SVD factors.
In fact, it can be shown that this method of matrix approximation is
optimal, in the sense that the reconstructed matrix has the smallest
possible approximation error, given the constraint that the ranks of fac-
tors cannot exceed k. In other words, expression 2.77 is a solution of
the following optimization problem:
min kX Ak
A (2.78)
subject to rank pAq ¤ k
Low-rank approximation is useful in marketing applications, espe-

cially in searches and recommendations, because it can help to deal
with sparse, noisy, and redundant data sets. This is often the case for
data sets that describe the interactions between two entities. For exam-
ple:
• The design matrix can capture the interactions between cus-
tomers and products. Each row of the matrix corresponds to a
customer, each column represents a product, and each element
is some interaction metric, such as the number of purchases. In
practice, this matrix is likely to be very sparse because each user
purchases only a small fraction of the available products. The
data is also highly correlated because many products are similar
to each other and many customers have similar shopping habits.
• In search applications, texts such as product descriptions are of-

ten modeled as vectors with each element corresponding to a
word; hence, the length of the vector is equal to the total num-
ber of distinct words in the vocabulary. A collection of texts can
then be represented as a matrix in which each row corresponds
to a text document and each column corresponds to a word. This
matrix can be sparse, especially for short texts, and redundant
because words with related semantic meaning frequently appear
together.
In the examples above, each element xij of the design matrix is some
measure of affinity between the entities, such as the affinity between a
customer and product or between a word and document. However, the
raw affinity values in the design matrix are often noisy and incomplete.
How can we create a smoother affinity model that predicts the affinity
for any pair of entities? One possible approach is to model the affinity
as a dot product of two vectors
p
xij pi qTj (2.79)
in which the first entity (e. g., the customer) is somehow mapped to
numerical vector p and the second entity (e. g., the product) to numer-
ical vector q. The length of vectors k is typically chosen to be much
smaller than the size of the design matrix. By rewriting this expression
in matrix form as
p P QT
X (2.80)
we can see that the vectors p and q, which minimize the average affin-
ity approximation error, can actually be obtained from the low-rank
approximation expression 2.77 as
P Uk Σ k
Q Vk
(2.81)
This result is very useful because it helps to convert sparse and re-
dundant representations of entities into compact and dense numerical
vectors. It is a very powerful and generic modeling technique, which
will be extensively used in the following chapters.
2.5.2 Clustering
Clustering is a process of grouping similar items together. Alterna-

tively, it can be thought of as splitting of the data set into clusters in
such a way that data points inside one cluster have high similarity and
points that belong to different clusters have low similarity.
In traditional marketing applications, the most important usage of
clustering is exploratory data analysis. Clustering can partition a data
set into a relatively small number of clusters, and each cluster can then
be described, interpreted, and studied as one entity. A canonical ex-
ample is customer segmentation, where a large number of customer
profiles can be divided into a few clusters or segments based on a
similarity metric that accounts for demographic properties, behavior,
and shopping habits. Each segment can then be described based on its
geometric center in the profile vector space (average profile) and the
spread of feature values, so a typical cluster description may look like
price-sensitive customers under 30 who are engaged primarily through digital
channels. Thus, segmentation allows a large data set to be summarized
into a few points that are tractable for manual analysis. Segmentation
projects are among the largest and most strategic projects in market-
ing analytics because a large part of the corporate marketing strategy
can be built around customer segments and their typical needs and
properties.
Programmatic applications, which are more focused on execution
than strategic analysis, often consume the outputs of segmentation as
additional features. For example, a customer profile vector that con-

tains values such as age, income, and average monthly spending can
be extended with segment labels that describe the person as a bargain-
hunter, brand-loyal fashionista, and so on. These additional features can
then be used in predictive modeling, like any other features, to improve
the prediction accuracy and interpretability of the results. From this
standpoint, clustering can be viewed as a feature-engineering method.
Clustering can also be applied to entities that are used as features
in modeling of other entities. In online advertising, for example, user
profiles typically include the URLs visited by the user in the past, so
the profile data may look like
user 1 : purl1 , url2 , url3 , . . .q

user 2 : purl4 , url5 , url6 , . . .q
This representation can be extremely sparse because the number of
distinct URLs is extremely high. Consequently, it can easily be the case
that user profiles have no or very few URLs in common, so a model
that uses this profile representation as a feature cannot accurately fit on
the data. Each URL, however, can also be associated with a vector of at-
tributes, such as the domain name and related web sites, and a cluster-
ing algorithm can then be applied to group URLs into categories. The
categories produced by the clustering algorithm can have some seman-
tic meaning. For example, one cluster of URLs may mainly correspond
to sport-related resources, whereas another cluster may correspond to
technology-related resources. The URLs in the user profiles can then
be mapped to the clusters, so the profiles can be expressed in terms of
behavioral features:
user 1 : psport, fashion, technology, . . .q
user 2 : pnews, fashion, sport, . . .q
With the assumption that the number of clusters is much smaller than
the number of URLs, this representation is much denser, and many
profiles have common cluster labels. This usage of clustering is pure
representation learning that aims to find a convenient set of features,
and it is substantially different from the exploratory analysis applica-
tions.
Clustering is an inherently challenging problem because learning
is unsupervised, and, consequently, the optimization objective cannot
be unambiguously defined. There exist several families of clustering
algorithms that take different perspectives on this issue. One of the ap-
proaches is to consider clustering as a model fitting problem: because
our goal is to group similar items together, we can assume that each
group is produced by some random but unknown process, and we can
fit a mixture of distributions that is likely (in the sense of the maximum
likelihood principle) to generate the observed data set. This approach
is illustrated in Figure 2.10, where the clusters are determined by fitting
the mixture of three normal distributions on the data.
Figure 2.10: Example of clustering by using the mixture modeling approach. (a)
Input data set. (b) Three clusters found by fitting a mixture of three
normal distributions.
This model can be specified as follows:

Ķ
ppxq wk N px | µk ; Σk q (2.82)

k 1
in which K is the number of clusters, wk are the mixing weights, and

µk and Σk are the means and covariance matrices of the distributions
in the mixture, respectively (note that the normal distribution is just
one possible option and one can use a mixture of other distributions as
well). Given that the parameters of the distributions are estimated, we
can easily assign each data point to the corresponding cluster based on
the probability densities at this point. The opposite is also true – given
that the cluster assignments are known, we can easily estimate the pa-
rameters of each distribution in the mixture. The challenge, however, is
that we do not know either the cluster mappings or the parameters of
the distributions; we only know the raw data points. This leads to com-
plex likelihood functions that are much more difficult to compute than
the likelihoods we have discussed in the previous sections. However,
there exist a number of iterative methods that can find an approximate
solution. The most widely used include the expectation-maximization
(EM) and K-means algorithms.
2.6 more specialized models
Standard supervised and unsupervised learning methods address

the most typical modeling needs that arise in marketing applications.
Many marketing concerns can be translated into such standard
predictive modeling problems in a relatively straightforward way.
Other marketing problems, however, may need more specialized data
analysis techniques or complex economic models that bridge the
business objective and basic primitives of predictive modeling. Some
of these methods were originally developed in economics, others
in game theory, biology, and the social sciences. In this section, we
describe several specialized models and methods that extend the
toolkit of standard machine learning techniques.
2.6.1 Consumer Choice Theory
The understanding and prediction of consumer choice is one of the

most fundamental problems in marketing, as well as economics in gen-
eral, because many important questions related to product design, as-
sortment planning, and distribution cannot be answered if the demand
is not well understood. In this section, we consider the discrete choice
problem, that is, the situation when a decision-maker faces a choice
among a set of alternatives. For example, a consumer decides which
of several competing products to buy, whether to cancel their subscrip-
tion to a certain service or not, and so on. We can assume that the
decision-maker compares the options in a consistent way (if option k
is preferable to option m and option m is preferable to n, then k is
preferable to n) and chooses the most preferable one, so it is valid to
introduce a virtual numerical metric that is proportional to the utility
of a given option to the decision-maker.
Let us denote the utility that decision-maker n obtains by choos-
ing option j from the alternatives p1, . . . , Jq as Ynj . The decision-maker
chooses option Ynj among the other options if Ynj ¡ Yni for all cases
in which i j. The utility of the same alternative j is not necessarily
equal for all decision-makers because of differences in tastes, income,
and other properties between individuals.
A model of consumer choice can be created by using the known
properties of the individual and the alternatives. However, each
decision-maker is likely to have additional considerations that influ-
ence their choice that cannot be observed by the model creator. More
formally, we can state that utility Ynj is a function of known factors
xnj and unobserved factors hnj :
Ynj Y pxnj , hnj q (2.83)

2.6 more specialized models 53
The hidden factors hnj are known to the decision-maker but not to
the model creator, so the utility model Vnj V pxnj q approximates the
true utility Ynj with some error εnj that can be considered as a random
variable:
Ynj Vnj εnj (2.84)
This approach to utility analysis is known as the random utility model.

Definition 2.84 allows us to express the probability of choosing alterna-
tive j by decision-maker n as

Pni Pr Yni¡ Ynj , @j i
Pr Vni εni ¡ Vnj @j i
(2.85)
εnj ,
Let us denote the random vector of errors as
εn pεn1 , . . . , εnJ q (2.86)
By assuming that the distribution of εn is known, we can evalu-

ate the choice probability by integrating over the probability density
ppεn q:
»

Pnj I Vni εni ¡ Vnj εnj , @j i ppεn q dεn (2.87)
ε
in which I is the indicator function that equals 1 when the argument is

true and 0 otherwise. We do not need to make any specific assumptions
regarding model V to evaluate expression 2.87 – we are free to choose
any linear or nonlinear function of known factors xnj to estimate the
utility. However, for practicality, we have to make certain assumptions
regarding distribution ppεn q in order to make the evaluation of Pnj
tractable.
Different assumptions about residual errors εn lead to different
choice models with distinct strengths and limitations. The ultimate
goal is to find a computationally tractable formula that expresses Pnj
as a function of Vnj , which, in turn, is a function of the observed
properties xnj and some parameters w. For example, it can be a linear
model:
Vnj wT xnj (2.88)
As Pnj can typically be estimated from the known statistics for alter-
natives that were available in the past, the parameters w can also be
estimated and a prediction model for Pnj can be built. This model can
be used to evaluate new alternatives specified in terms of properties x,
and, consequently, economic metrics like demand for a new product
can be predicted. In the next section, we discuss one of the most sim-
ple yet powerful and practical models, multinomial logit, to show how
computationally tractable expressions for Pnj can be obtained. This
model will be used in the subsequent chapters as a component of more
complex models that are created for specific marketing problems.
2.6.1.1 Multinomial Logit Model

The multinomial logit model, commonly abbreviated as the MNL
model, can be derived from the random utility model by assuming
that the residual errors εnj are independent and follow a Gumbel
distribution. The assumption about the Gumbel distribution leads
to a convenient model and can also be considered as a practical ap-
proximation of the normal distribution [Train, 2003]. The assumption
about independent distributions is much more restrictive and leads
to limitations that will be discussed later in this section. In general,
the Gumbel distribution is used to describe the distribution of the
maximum or minimum value of a number of random samples drawn
from some underlying distribution. For example, if one generates
batches of random numbers by drawing them from the normal
distribution and then takes the maximum value in each batch, the
distribution of these maximums can be modeled by using the Gumbel
distribution. It is useful in applications that deal with extreme events,
such as earthquakes, manufacturing defects, and equipment failures.
A drug manufacturer that produces oral drug tablets in batches, for
instance, can use the Gumbel distribution to model the probability of
producing a batch where the level of active chemical components is
higher than the maximum acceptable level. The probability density for
the Gumbel distribution is defined as
ppεnj q eεnj exppeεnj q (2.89)
and the cumulative distribution is given by
Fpεnj q exppeεnj q (2.90)
To take advantage of the assumption that the residual errors follow the
Gumbel distribution, let us first rewrite the choice probability given by
equation 2.85 as follows:
Pni PrpVni εni ¡ Vnj εnj , @j iq
Prpεnj εni Vni Vnj , @j iq
(2.91)
By assuming for a moment that εni is given and by using the indepen-
dence of errors, we can state that
¹
Pni | εni Prpεnj εni Vni Vnj q (2.92)

j i
The terms on the right-hand side are effectively cumulative distribu-

tions of εnj , so by inserting the definition of the Gumbel distribution
given by equation 2.90, we get the following:
¹
Pni | εni exp epε ni
Vni Vnj q
(2.93)

j i
As εni is not actually given, we need to integrate over its probability

density to obtain a complete expression for Pni :
»
Pni pPni | εni q eε ni
exppeεni qdεni (2.94)
ε
A concise closed-form expression for the choice probability can be ob-

tained directly from the equation above by means of algebraic trans-
formations that we skip here for the sake of brevity; the result is a
canonical formula for the MNL model:
Vni
Pni °e eV nj
(2.95)
j
The model given by equation 2.95 has several important properties

and implications:
independence of irrelevant alternatives One of the most
important questions that should be answered by choice mod-
eling is how changes in the utility of one alternative influence
other alternatives. For example, a product manufacturer may
be interested in estimating the share of consumers that can be
drawn away from competitors by reducing the product price
or launching a new product. The MNL model implies that an
increase or decrease in the probability of one alternative will
evenly affect all other alternatives. To see this, consider the ratio
of any two probabilities:
L°
eeV eV V
Vni eVnk V
Pni
eeV
ni
L °k ni nj
(2.96)
Pnj Vnk
ke
nj nj
The ratio of probabilities depends only on the ratio of the

corresponding utilities, a property commonly referred to as
independence of irrelevant alternatives; if the utility Vni changes,
the pairwise ratios for all other pairs Pnp {Pnq remain constant.
This property of MNL is somewhat restrictive because products
within a group of interest are not always perfectly substitutable
and more complex substitution patterns can take place. This
limitation can be illustrated by the following paradox [Debreu,
1960].
Consider a transportation system where a consumer chooses be-

tween a car and a bus and where the initial choice probabilities
are equal:
Pcar Pbus 1/2 (2.97)
Suppose now that a second bus is introduced, and it is identical

to the first bus. Let’s say that the only difference is the color: the
first one is red and the second one is blue. The MNL model will
evenly redistribute the probabilities to produce
Pcar Pred bus Pblue bus 1/3 (2.98)
because the utilities are equal for both buses. A more realistic
assumption, however, is that the ratio
Pcar {pPred bus Pblue bus . . .q (2.99)
will remain constant no matter how many identical buses are of-
fered, which produces the probabilities Pcar 1/2 and Pred bus
Pblue bus 1/4.
completeness of the utility model The independence of the

residual errors εni implies that the utility model Vni should
capture all factors that influence the choice. If model Vni is not
complete, some systematic biases start to leak into the error
components and violate the independence assumption. For
example, we can build a utility model for a washing machine
by using the price p and energy consumption c as predictive
variables, so Vni w1 pi w2 si . However, the choice is likely
to depend on the consumer’s income g, which might not be
known, so the utility will actually be
w1 pi w2 si w3 gn εni Vni εni (2.100)

Yni
in which ε
ni w3 gn εni and represents the errors that are
not independent because of the random variable gn .
marginal choice probability The choice probability is a sig-

moid function of utility Vni , as shown in Figure 2.11. This means
that small changes in the utility cause a substantial increase or
decrease in the probability that the corresponding alternative
will be chosen only if the probability is close to 0.5, that is, if the
decision-maker is in a marginal state. If the probability of the
choice of a certain alternative is low or high, then even major
changes in the utility have limited impact on the probability.
𝑃𝑛𝑖
1.0
0.5
𝑉𝑛𝑖
0.0
Figure 2.11: Probability of choice as a function of the utility and its derivatives
at different points.
This S-shaped dependency between Vni and Pni suggests that

investment should be focused on the development of alterna-
tives with intermediate probability of selection by the majority
of decision-makers. For example, an online retailer that is im-
proving its order delivery network can expect the highest return
on investments by improving the service in the areas where it
already has average market share, whereas areas with very low
or very high shares are likely to be less responsive to improved
services.
2.6.1.2 Estimation of the Multinomial Logit Model

Let us now describe how the parameters of the utility model Vni can
be estimated from training data samples. Let us assume that we know
features xni that are included in the utility model for some subset of
decision-makers n 1, . . . , N and each alternative i 1, . . . , J, as well
as the actual choices made. Let yni P t0, 1u be the observed choice
of decision-maker n with respect to alternative i; this equals 1 if the
decision-maker has chosen this alternative and 0 otherwise. Under the
assumption of independent residual errors, we can express the prob-
ability of the choice that was actually made by the decision-maker as
¹
pPni qy ni
(2.101)
i
The probability that all decision-makers in the data set made their
choices as we actually observe, that is, the likelihood of the data set,
can then be expressed as
¹
N ¹
Lpwq pPni qy
ni
(2.102)
n 1 i
under the assumption that all decisions are independent. Consequently,

the log-likelihood will be
Ņ ¸
LLpwq yni logpPni q

n 1 i
(2.103)
Ņ ¸
eVni
yni log ° V

n 1 i ke
nk
in which Vni is a function of w and x, for example, a linear model

Vni wT xni . The log-likelihood described by equation 2.103 can fur-
ther be estimated by taking the gradient with regard to w and applying
numerical optimization methods.
2.6.2 Survival Analysis
Classification methods, even the most basic ones such as logistic re-
gressions, provide a powerful toolkit for estimating the probabilities
of consumer actions. For example, the response probability for a pro-
motional email can be estimated by building a model that uses cus-
tomer attributes, such as the number of purchases, as features and a
binary variable that indicates whether a customer responded to the pre-
vious promotional email as a response label. Although this approach
is widely used in practice, as we will discuss in detail in subsequent
chapters, it has a few shortcomings. First, in many marketing applica-
tions, it is more convenient and efficient to estimate the time until an
event, instead of the event probability. For example, it can be more use-
ful for a marketing system to estimate the time until the next purchase
or time until subscription cancellation, rather than the probabilities of
these events. Second, marketing data very often include records with
unknown or missed outcomes, which cannot be properly accounted
for in classification models. Going back to the example with the sub-
scription cancellation, it is often impossible to distinguish between cus-
tomers who have not defected and those who have not defected yet
because we build a predictive model at a certain point in time and
cannot wait indefinitely until the final outcomes for all customers are
observed. Consequently, we only know the outcomes for customers
who have defected and can certainly label them as negative samples;
the remaining records are incomplete but those customers will not nec-
essarily not defect in the future, so one can argue that labeling these
samples as positives or negatives is not really valid. This suggests that
it is not accurate to use a classification model with a binary outcome
variable determined on the basis of currently observed outcomes and
we might need a different statistical framework to deal with this sort

of problem.
A comprehensive framework for time-to-event modeling and incom-
plete data handling was originally developed for medical and biolog-
ical studies. The main focus of the studies was the survival of indi-
viduals after medical treatment, so the framework became known as
survival analysis. Let us describe the main methods of this framework
by starting with the basic terminology. The main goal of survival anal-
ysis is to predict the time to an event of interest and quantitatively
explain how this time depends on the properties of the treatment, in-
dividuals, and other independent variables. In marketing applications,
the treatment is typically an incentive or trigger, such as a promotion.
An event is typically a purchase, promotion redemption, subscription
cancellation, or any other customer action that a marketer might be in-
terested in influencing. Note that a positive outcome of the treatment
can be either acceleration or deceleration of the event, depending on
the application. Advertisements, for example, aim to stimulate earlier
purchases; meanwhile, retention offers aim to suppress subscription
cancellation events. By contrast, medical research typically measures
the time from diagnosis to death, so the standard terminology of sur-
vival analysis assumes that the event corresponds to some negative
outcome, which can be confusing when the opposite is the case.
As we discussed earlier, some events may be unknown, in the sense
that the outcomes have not been observed at the time of study. These
unknown outcomes can occur because the outcome is not known by
the time of the analysis (the customer has not yet purchased but may
still purchase in the future) or the customer record has been lost (for
example, because of a browser cookie expiration). The records with
unknown outcomes are referred to as censored records. By the time
of the analysis, we initially have a set of observations, each of which
has the time of the treatment and, optionally, the time of the event of
interest, as shown in Figure 2.12.
The time between the treatment and event is referred to as the sur-
vival time. We can transform the original observations to align all of the
treatment times, so the observed data for k individuals (customers) is
a sequence of pairs
pt1 , δ1 q , . . . , ptk , δk q , t1 ¤ . . . ¤ tk (2.104)
in which t stands for a timestamp of the event and δ is an indicator

that equals 1 if the observation is not censored and 0 otherwise. We
generally assume a continuous timescale, but two customers can have
the same event time, so we can summarize the input data as
pt1 , d1 q , . . . , ptn , dn q (2.105)

Figure 2.12: Data preparation for survival analysis. The filled circles correspond
to treatments. The crosses correspond to events. The empty circles
denote censored records.
in which n is the number of distinct event times; di is the total number

of observed events at time ti . We also assume that the events are non-
repeatable, that is, that an individual cannot experience more than one
event. This assumption is not literally true for many marketing events
such as purchases, but we can typically work around it by creating
separate models for first, second, and subsequent event occurrences,
as we will discuss in later chapters. At this point, we are interested
only in the distribution of the events and are not trying to explain
the dependency between the survival time and the properties of the
treatment or customers.
2.6.2.1 Survival Function

The distribution of survival times can be described in terms of the sur-
vival probability, also called the survival function, Sptq, which is defined
as the probability that an individual survives from the time origin to
time t. The survival function is a fundamental characteristic that de-
scribes the dynamics of a customer group. If the survival function falls
sharply, most customers are likely to experience the event relatively
soon. If the function falls slowly, most customers are likely to experi-
ence the event at a relatively distant point in the future.
Let us denote the survival time of a customer as T and its probability
density function as fptq. The cumulative distribution function of the
survival time, which corresponds to the probability of the event by

time t, will then be
»t
Fptq Pr pT ¤ tq fpτqdτ (2.106)
0
and the survival function can then be defined as
Sptq PrpT ¡ tq 1 Fptq (2.107)
The value of the survival function at time point t corresponds to the

fraction of customers who have not yet experienced the event at that
point. Note that the statistical properties of the survival time, such as
the mean, median, and confidence intervals, can be estimated based on
the cumulative distribution function. Consequently, these properties
can be estimated if an estimate for the survival function is available.
The survival function can be estimated from the observed data, by
taking into account both censored and uncensored records, based on
the assumption that events are independent from each other. In this
case, the cumulative survival probability can be obtained by multiply-
ing the probabilities for survival from one interval to the next. More
formally, the probability to survive to time t can be straightforwardly
estimated as
St nt n dt 1 ndt (2.108)
t t
in which nt is the number of individuals who have not yet experienced

the event at time t and dt is the number of individuals who have
experienced the event at time t. By multiplying the probabilities from
the origin time until time t, we can estimate the cumulative probability
to survive, that is, the survival function:

¹
p pt q
S 1
di
(2.109)
¤
i t
ni
This estimator is known as the Kaplan–Meier estimator, which can

be proved to be the maximum likelihood estimator [Kaplan and Meier,
1958]. The survival function equals 1 at time zero, and then each sam-
ple contributes to the estimate as we increase the time. Let us illustrate
the estimation of the survival function with a small numerical example.
example 2.1
Let us assume that we are analyzing a group of 14 customers after

each of them received a promotional email. All emails were sent at

different times and the time of the first purchase after an email has
been recorded. The observed data set looks like this:
t t2, 3, 3, 3, 4, 6, 7, 8, 12, 12, 14, 15, 20, 23u

(2.110)
δ t1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 0, 1, 1, 1u
in which the i-th element of set t is the observed event time for the i-th
customer measured in days since the email was sent. Set δ contains the
indicators for whether each observation is censored (0) or not censored
(1). For example, the first customer made a purchase on the second day
after the email, and the third customer did not make a purchase by the
time of the analysis although she got the email three days before the
analysis date. In this context, the probability to survive means the prob-
ability of not having made a purchase at a given time. By repeatedly
applying formula 2.109, we obtain the following sequence:
Sp0q 1 (all customers are “alive” at the beginning)

Sp2q 1 0.93
1
14

(2.111)
Sp3q Sp2q 1
2
0.79
13
...
This result corresponds to the stepwise survival curve plotted in Fig-

ure 2.13. The survival curve summarizes the dynamics of a customer
group, and curves for different groups can be compared. For exam-
ple, a survival curve for customers who were treated with a promotion
can be plotted together with a curve for those who were not, and the
efficiency of the promotion can thus be graphically assessed.
N
2.6.2.2 Hazard Function
The second important concept in survival analysis is the hazard function.
Whereas the survival function focuses on the probability of the event
not happening, that is, survival, the hazard function describes the risk
of the event. As we will see later, this perspective is convenient for an-
alyzing how different factors, such as treatment parameters, influence
the survival time.
The hazard function hptq is defined as the instantaneous hazard rate,
that is, the probability of the event in an infinitesimally small time
period between t and t dt, given that the individual has survived up
until time t:
Prpt T ¤t dt | T ¡ tq
hptq lim (2.112)
Ñ0
dt dt
Figure 2.13: An estimate of the survival function for the data set given in defini-
tion 2.110.
The hazard function can be linked to the survival function. To

see this, let us first decompose the conditional probability in defini-
tion 2.112 into two factors; note that one of them corresponds to the
survival function:
Prpt T ¤ t dtq
hptq lim
dt dt PrpT ¡ tq
Ñ0
Prpt T ¤ t dtq
dtlim
Ñ0 dt Sptq
(2.113)
Fpt dtq Fptq

dtlim
Ñ0 dt Sptq
Next, recall that the probability density function is defined as
Fpt dtq Fptq
fptq lim (2.114)
Ñ0
dt dt
By using this definition, as well as the definition of the survival func-
tion 2.107, we obtain the following result:
fp t q
h pt q
Sptq
1 fpFtqptq
dt log p1 Fptqq
d (2.115)
dt log pSptqq
d
By solving this equation with respect to Sptq, we can express the

survival function as a function of hptq:
Sptq exp pHptqq (2.116)
in which
»t
Hptq hpτqdτ log pSptqq (2.117)
0
is called the cumulative hazard function. This straightforward relation-

ship allows us to switch between the hazard and survival functions in
the analysis.
2.6.2.3 Survival Analysis Regression

The basic survival and hazard functions can be used to describe the
performance of a customer group or compare different groups to each
other. This is not enough for applications where we need to under-
stand and predict how survival and hazard are influenced by factors
like marketing actions and customer properties. This problem is simi-
lar to the classification and regression problems, in the sense that the
survival time has to be predicted as a function of the observed factors,
that is, the independent variables.
Let us assume that each individual is associated with a vector x that
consists of p independent variables, so that each individual is now
represented by three values:
t survival or censoring time

δ censoring indicator, taking the value 1 for observed
events and 0 for censored cases
x vector of features
The input data set contains observations for k individuals:
pt1 , δ1 , x1 q , . . . , ptk , δk , xk q , t1 ¤ . . . ¤ tk (2.118)
In marketing applications, the feature vector can include customer

demographic and behavioral properties, marketing communications to
that customer, and so on. The goal is to define and fit a model that ex-
presses the survival and hazard functions as a function of x. As both
Sptq and hptq are probabilities, we can construct different survival re-
gression models by assuming different probability distributions and
different functional dependencies between the features x and the pa-
rameters of the distribution.
Among the most commonly used choices for survival regression
models are proportional hazard models. This model family is based on
the assumption that a unit increase in observed factors is multiplicative
with respect to the hazard rate, that is
h pt | w, xq h0 ptq rpw, xq (2.119)

in which h0 ptq is the baseline hazard, r is the risk ratio that increases
or decreases the baseline hazard depending on the factors, and w is
the vector of model parameters. Note that the baseline hazard does
not depend on the individual, but the risk ratio does. In other words,
the risk ratio determines how the properties of an individual encoded
in the feature vector influence the hazard rate. The risk ratio cannot
be negative because the hazard rate is not negative, so it is typically
modeled as an exponential function:

h pt | w, xq h0 ptq exp wT x (2.120)
This model can be interpreted as a linear model for the log of the
risk ratio for an individual to the baseline:
h pt | xq
log rpw, xq log wT x
h 0 pt q
(2.121)
In regards to the baseline hazard h0 ptq, we have two choices: non-

parametric and parametric. The parametric approach assumes that the
hazard follows a certain probability distribution. In this case, we obtain
a fully parametric model that needs to be fitted to the data by finding
the optimal values of parameters w and the parameters of the distribu-
tion. The disadvantage of this approach is that the baseline hazard is
assumed to vary in a specific manner with time, so we need to be sure
that the distribution we choose matches the data. On the other hand,
the parametric approach smooths noisy data and provides a simple
model for the baseline hazard.
The second option is to use a nonparametric baseline hazard model
that can be estimated from the data by using the Kaplan–Meier esti-
mator or other methods. This leads to a semiparametric model for the
overall hazard where the parametric part is defined by expression 2.120
and the baseline hazard h0 ptq is the nonparametric part. This solution
is known as the Cox proportional hazard model [Cox, 1972]. The ad-
vantage of the Cox model, as we will see in a moment, is that it allows
us to estimate the hazard ratios without having to estimate the base-
line hazard function or making any assumptions about the structure
of the baseline hazard. This makes it very convenient for applications
that require only the risk factors to be estimated, not the absolute haz-
ard values. The downside of the Cox model is that the baseline hazard
needs to be estimated by using parametric methods. It is also impor-
tant to keep in mind that the Cox model belongs to the proportional
hazard models family and, consequently, is based on the proportional
hazard assumption, which may or may not hold for the observed data.
The Cox model is widely used in many domains including marketing,
and we will use it as the main vehicle for survival analysis in the next
chapter.
Our next step is to estimate the parameters of the Cox model from
the data. The standard approach to this problem is to derive the like-
lihood of the model and then find the parameters that maximize it.
The challenge, however, is that observations can be censored, which
requires us to specify how such records will be accounted for in the
likelihood. First, let us note that each observation contributes to the
likelihood. If the i-th observation is censored, it contributes the proba-
bility of survival up until ti :
Li pwq S pti | w, xq (2.122)
If the observation is not censored, it contributes the probability of the

event at ti , which is defined by using the survival time probability
density function:
Li pwq fpti q h pti | w, xq S pti | w, xq (2.123)
So the full likelihood can be expressed as follows:
¹
k
Li pwq h pti | w, xqδi S pti | w, xq (2.124)

i 1
We cannot maximize this by using numerical methods without spec-

ifying the form of the baseline hazard. However, it is possible to ap-
proximate the full likelihood with a different measure called the par-
tial likelihood. First, let us introduce the notion of the risk set at time t,
which is defined as the set of individuals who are at risk of the event
at time t, that is, those who have not yet experienced the event:
Rp t q ti : t i ¥ tu (2.125)
For simplicity, let us also assume that there are no event ties, that is,
all event times ti are distinct1 . In this case, the partial likelihood can be
defined by using the conditional probability that a particular person i
will fail at time ti , given the risk set at that time, and that exactly one
failure is going to happen [Cox, 1972, 1975]. This probability is given
by the area under the hazard curve for a small time interval dt, so the
likelihood contributed by individual i can be expressed as
hpti | w, xi qdt
Li pwq °
j P Rpti q hptj | w, xj qdt
(2.126)
1 The case with ties is more complex, but there exist a number of generalizations that account
for ties [Breslow, 1974; Efron, 1977]. In most marketing applications, we can avoid ties by
spreading the conflicting observations by a small margin.
By inserting the specification of Cox model 2.120 into this partial

likelihood, we can see that the baseline hazards cancel each other out
and we get

exp wT xi
Li pwq ° (2.127)
j P Rpti q exp w xj
T
Finally, the partial likelihood for the entire training data set is a prod-
uct of the individual partial likelihoods given by equation 2.127:
δi
¹
k
exp wT xi
Lpwq ° (2.128)
j P Rpti q exp w xj
T
i 1
This likelihood does not depend on the hazard function, so we can

fit it by using numerical methods with respect to the weights w. This,
in turn, enables us to estimate the risk ratios defined by equation 2.121.
The ability to estimate the risk ratios without estimating the hazard
function is one of the key advantages of the Cox model.
Thus far, we have been focused on the estimation of regression

weights. Our last step is to specify how the baseline hazard and
survival functions can be estimated. First, let us note that the expected
number of events at time ti can be approximated by the area under
the hazard function for a small interval between ti and ti dt:
¸
p
d i h0 pti q exp wT xj dt (2.129)
j P Rpti q
We can rewrite this relationship as
p
p pt qdt °
h
d i
(2.130)
0 i
j P Rpti q exp wT xj
and then approximate the cumulative hazard function as follows:

¸
p pt q
H p pt qdt
h
0 0 i (2.131)

i t
This result is known as the Breslow estimator [Breslow, 1972]. It en-

ables us to estimate the baseline survival function by inserting the esti-
mator into expression 2.116:
p ptq exppH
S p ptqq (2.132)
0 0
Finally, the complete survival function can be obtained directly

from the specification of the Cox model in equation 2.120 and
expression 2.116 as follows:
»t
Spt | xq exp h0 pτq exp wT x dτ
0
»t p
exp wT x q
exp h0 pτqdτ
(2.133)
0
S0 ptqexppw xq T
Survival functions for different values of features x can then be plot-

ted together and compared to each other to quantify the impact of
different features on the distribution of the survival time. We will con-
tinue to discuss this and other practical applications of survival analy-
sis in advertising and marketing communications in the next chapter.
2.6.3 Auction Theory
As we discussed in Chapter 1, the algorithmic approach facilitates the

development of marketing services that can be offered to clients by ex-
changes. An exchange or any other type of broker in between a service
provider and service client adds an extra layer of complexity because
both the provider and client should optimize their service buying and
selling strategies, in addition to pursuit of the primary marketing ob-
jectives.
The basic goal of a service exchange is to enable competition be-
tween buyers for a limited resource, such as advertisement placements.
The standard approach to this problem is an auction where each buyer
places a bid and the resource is auctioned off to the bidder with the
maximum bid. However, auction settings and rules can be set up dif-
ferently, so we need to spend some time discussing auction types.
First, we have to acknowledge that the bidders participate in the
auction because the auctioned resource has a certain value for each of
them and they are aiming to make profits by buying the item below
this valuation. Consequently, it is critically important for a bidder to
estimate the value of the item correctly, and we can classify all auction
settings by the following valuation types:
private value Each bidder evaluates the item independently from
other bidders and the estimate does not depend on other bids,
even if they are known.
interdependent value The actual value is not known to the bid-

ders, and, although each bidder has their own estimate of what
the value is, information about other bids can help to improve the
estimate. For example, a bidder who highly values the resource
might reduce their bid if other participants bid lower, because
this additional information can indicate the presence of negative
factors that are unknown to the given bidder but are somehow
recognized by the others.
common value This is a particular case of an interdependent value

auction where the actual value is the same for all bidders. Exam-
ples of common value auctions include selling natural resources
such as oil or timber, selling financial assets such as bonds, or sell-
ing a company. In all of these cases, the true value might not be
precisely known at the auction time and bidders must estimate
it based on the limited information that they have, but eventu-
ally the value becomes known (actual amount of recoverable oil,
long-term company performance, etc.) and it is the same for all
participants.
Although the value is often interdependent to some degree, a bid-
der’s ability to take advantage of knowing other bids depends on the
auctioning process. The four main types of auctions studied in theory
and used in practice are as follows:
open bid Every bidder observes the value of all other bids.
• Open ascending-price auction (English auction). The price
starts at a low level and increases. At any point of time, a
bidder can either stay or quit. The auction ends when only
one bidder remains, and the winner pays the final price.
• Open descending-price auction (Dutch auction). The price
starts at a high level and decreases. The auction ends when
any bidder accepts the current price.
sealed bid Bidders are unaware of what others have bid.

• First-price sealed-bid auction. All bidders submit their bids
simultaneously, so that no bidder knows the bid of the other
participants. The bidder with the highest bid wins and pays
the winning bid.
• Second-price sealed-bid auction (Vickrey auction). Similarly
to the first-price auction, all bidders submit their bids simul-
taneously and the bidder with the highest bid wins, but the
winner pays the second highest bid.
Although the optimization problem for the open-bid auctions might
seem dynamic, it is essentially static and equivalent to the sealed-bid
auctions. The Dutch auction ends right after the first bid, so bidders do
not receive any additional information during the process and can de-
cide on a bid in advance. Consequently, a Dutch auction is equivalent
to a first-price sealed-bid auction in the sense that,whatever strategy
the bidder chooses, it uses the same inputs and leads to the same win-
ners and prices, both for private and interdependent values. In an En-
glish auction with private values, the bidder can also evaluate the item
in advance. As the auction progresses and the price goes up, the bid-
der should always compare the current highest bid with the estimated
value and either make a new bid, calculated as the current highest plus
some small increment, or quit the auction if the price has gone above
his valuation. Hence, an English auction is equivalent to a second-price
sealed bid auction for private values, although this is not true for inter-
dependent values because the bidder can learn from the observed bids
in the case of an English auction.
We now study the Vickrey auction in detail to obtain a toolkit for
building optimization models that include auctions. We focus on the
Vickrey auction because it is convenient for analysis and widely used
in practical applications, although similar results can be obtained for
other auction types by using more advanced analysis methods.
First, we can prove that the optimal strategy for bidders is to bid
their true value. Consider Figure 2.14, in which the bidder evaluates
the item at a price v but makes a lower bid v δ. If the second highest
bid from another bidder is p, then the following three outcomes are
possible:
1. p ¡ v: the bidder loses; it does not matter if he bids v or v δ
2. p v δ: the bidder wins and pays price p; it does not matter if

he bids v or v δ
3. v δ p v: the bidder loses; a bid of v would have meant

winning and making a margin of v p
Figure 2.14: Vickrey auction – bidding below the true value.
So, a bid below the true value always gives the same or a worse result
as a bid of the true value. Figure 2.15 depicts the opposite situation,
when the bid is above the true value. We again have three possible
outcomes:
1. p ¡ v δ: the bidder loses; it does not matter if he bids v or v δ
2. p v: the bidder wins and pays the price p; it does not matter if
he bids v or v δ
3. v p v δ: the bidder wins and pays the price p with a loss of

p v, whereas a bid of v would mean losing the auction without
any financial loss
Figure 2.15: Vickrey auction – bidding above the true value.
We can conclude that bidding the true value is the optimal strategy.
This simple result implies that we need to focus on estimation of the
expected revenue for the bidder when discussing marketing settings
with sealed-bid auctions.
Our next step will be to take the seller’s perspective of the auction
and estimate the revenue for the seller. Let us assume that there are n
bidders participating in the auction and their bid prices V1 , . . . , Vn are
independent and identically distributed random values drawn from
some distribution Fpvq with the probability density fpvq. By recalling
that k-th order statistic Vpkq of a sample is equal to its k-th smallest
value, we can express the expected revenue as the mean of the second-
highest order statistic that corresponds to the second-highest bid:

revenue E Vpn1q (2.134)
Let us consider a slice of the probability density function for the

order statistics:

Pr v Vpkq v dv (2.135)
This is the probability that k 1 bids in the sample of n bids are smaller
than v, exactly one bid falls into the range rv, v dvs, and the remain-
ing n k bids are higher than v. These three conditions can be ex-
pressed by using the bid cumulative distribution Fpvq and probability
density fpvq, so we get the following expression for the order statistics
probability density:

f Vpkq dvlim
Ñ0
Pr v Vpkq v dv

k1
n
rFpvqsk1 pn k 1qfpvq r1 Fpvqsnk (2.136)
pk 1qn! fpvq rFpvqsk1 r1 Fpvqsnk

! pn k q !
We can simplify this expression with certain assumptions regarding
the bid distribution. For example, if bids are drawn from a uniform
distribution between zero and one, the expression reduces to

f Vpkq pk 1qn! vk1 p1 vqnk
! pn k q !
(2.137)
This is a beta distribution, so we can use a standard formula for its

mean to get the expected revenue of the auctioneer:
nn 11

E Vpn1q (2.138)
This result is in alignment with the intuitive expectation that an in-

crease in the number of bidders leads to an overall increase in the rev-
enue. We use these results in subsequent chapters, mainly to optimize
the bidding process for exchanges. It is worth noting, however, that
other marketing processes, including such important ones as the selec-
tion of the best offers on the market by a consumer, can be modeled as
auctions, which makes auction theory an important tool for building
programmatic solutions.
2.7 summary
• Many marketing problems can be expressed as optimization prob-

lems in which the business outcome is the subject of optimization
and the business actions are the variables.
• The dependency between the actions and business outcomes can

often be learned from historical data. This problem can be solved by
using supervised learning methods.
• The primary goal of supervised learning is to estimate the condi-

tional distribution of the response given the input. In many practical
applications, this problem can be reduced to finding the most proba-
ble outcome values. The main two types of supervised problems are
classification and regression.
2.7 summary 73
• The number of predictive model parameters can be fixed or can grow

with the size of the training data set. The former type of model
is referred to as a parametric model, and the latter is known as a
nonparametric model.
• Model fitting can be viewed as an optimization problem where the

model parameters need to be chosen to maximize the probability
that the observed data follows the model distribution.
• A wide range of supervised learning problems can be addressed by

linear models, that is, models that either express the dependency
between the input and output as a linear function or express the
boundary between the classes as a straight line. The most basic ex-
amples of linear models are linear regression and logistic regression.
• Nonlinear dependencies and decision boundaries can be captured

by using nonlinear models. Examples of nonlinear modeling meth-
ods include kernel methods, decision trees, and neural networks.
• Marketing data may have a redundant structure because different

features and metrics are projections of the same marketing process.
This structure can be nonoptimal for analysis and modeling, and a
better data representation can be found by removing correlations,
reducing data dimensionality, and clustering data points and enti-
ties. Some of these tasks can be solved with unsupervised learning
methods, such as principal component analysis and clustering.
• Some marketing tasks cannot be easily solved by using standard ma-

chine learning methods and require more specialized models and
techniques. Examples of such models include consumer choice mod-
els, survival analysis methods, and auction theory.
3
P R O M O T I O N S A N D A D V E RT I S E M E N T S
Every product or service has its target market, the group of consumers
at which the product or service is aimed. The distinction between the
target and non-target groups is often fuzzy and ambiguous because
consumers differ in their income, buying behavior, loyalty to a brand,
and many other properties. The diversity of customers is often so high
that an offering created for an average consumer, that is, for everyone,
does not really fit anyone’s needs. This makes it critically important
for businesses to identify the most relevant consumers and tailor their
offerings on the basis of consumer properties. This problem occurs in
virtually all marketing applications, and it plays an especially impor-
tant role in advertisements and promotions because the efficiency of
these services directly depends on the ability to identify the right audi-
ence and convey the right message.
The problem of finding the optimal match between consumers and
offerings can generally be viewed from two perspectives. First, it can be
stated as finding the right offerings for a given customer. This is a prod-
uct discovery problem, which we will discuss in the following chapters
dedicated to search and recommendations. The second perspective is
that of finding the right customers for a given offering. This problem is
known as targeting, and it is the main subject of this chapter. It should
be kept in mind, however, that we draw a line between product discov-
ery and targeting services mainly based on the principal applications
(interactive browsing versus advertising), and the methodologies used
to implement the services can sometimes be viewed from both per-
spectives. Consider customer base segmentation as an example. One
can argue that segmentation identifies the right groups of customers
first and the offerings and experiences are then tailored for each seg-
ment. However, it is also true that segmentation can be viewed as a
method for allocating different offerings and experiences to the most
appropriate customers.
75
76 promotions and advertisements
Although targeting is concerned with matching customers and of-

ferings, it should not be viewed simply as a set of methods for draw-
ing connecting lines between these two entities. Instead, it should be
viewed as a problem of customer experience optimization that is driven
by a mix of multiple business objectives and controls many different
marketing activities. The goal of a programmatic system is to unroll
these initial objectives into a detailed execution plan and specific rules
that can be used to manage interactions with the clients.
We will open this chapter with an overview of the retail promotion
environment that will help us to better understand the problem of tar-
geting. We will then describe a promotion targeting framework that
includes a more formal definition of business objectives, basic build-
ing blocks of behavioral modeling, and more complex constructs used
in marketing campaigns. After that, we will discuss the online adver-
tising environment and related targeting methods. Although the re-
tail and online advertising environments complement each other and
many targeting methods are universally applicable, we will study them
separately because of major structural differences and variations in ob-
jectives. Finally, we will discuss how the efficiency of targeting meth-
ods and marketing campaigns can be measured. Measurement plays
an extremely important role in all marketing applications, and the
framework that we develop will also be used in other programmatic
services, including search, recommendations, and pricing. In this chap-
ter, we generally avoid discussing price optimization, although it is an
important part of promotions. This topic will be covered in a dedicated
chapter later on.
3.1 environment
The first business environment we will consider is that of consumer

sales promotions, which are widely used in retail and brand relation-
ship management. The goal of promotions is to provide added value
or incentives to consumers in order to improve sales or build better
relationships. Promotions can be distributed by product manufactur-
ers, service providers, or retailers. In some market verticals , such as
consumer packaged goods (CPG), manufacturers and retailers often col-
laborate on promotional campaigns, with the manufacturer covering
the direct costs of the campaign and the retailer providing its physical
and digital channels to communicate the offerings to the audience. As
we discussed in Chapter 1, such interactions between promoters and
consumer-base owners can be favorable for the algorithmic approach,
so we will use this as our primary environment for further discussion.
However, most of the methods that we will develop are not limited to
3.1 environment 77
the CPG environment and can be applied in other domains such as

telecoms or insurance.
The model of the sales promotion environment is depicted in Fig-
ure 3.1. The main entities of this environment, assumptions, and termi-
nology can be described as follows:
• We use the term consumer to refer to any person who consumes
goods. A customer is a person who has purchased from a firm.
Finally, a prospect is a person who is not a customer yet but
who is known, in the sense that a firm can communicate with
them (for example, a person who has registered on the website
and provided an email). We also refer to consumers who interact
with online channels as users.
• Manufacturers (or brands) produce products that are associated

with product categories. We assume that each category has a rel-
atively narrow scope, such as low-fat cottage cheese, so products
within a category are considered by the consumer as interchange-
able. Consequently, several brands can compete within a category
for customers.
Figure 3.1: The sales promotion environment in the retail domain.
• A retailer purchases products from manufacturers and sells them

to consumers with an added markup. A retailer can also pro-
duce its own branded products, referred to as private-label prod-
ucts, to compete with other manufacturers in the corresponding
categories.
• We define a targeted promotion as an incentive that can be deliv-

ered to selected consumers through marketing channels. Promo-
tions can be offered with some condition (e. g. buy one, get one)
or without any condition, can provide a monetary value such as
a discount, or can just advertise a product or brand. Promotions
may or may not be redeemable, in the sense that a consumer might
need to submit evidence of the promotion (scan a bar code on a
printed coupon or enter a promotion code) to redeem its mone-
tary value with a purchase. We also use the word treatment as a
generic term that refers to promotions and other marketing com-
munications.
• A retailer owns marketing channels such physical stores or eCom-

merce websites that can be used to communicate promotions to
the consumers. The marketing channels of multiple retailers can
be combined into a promotion distribution network that can be oper-
ated by retailers or a third-party agency. For example, an agency
can install its coupon printers in stores that belong to multiple
retail chains.
It is critically important that the retailer or agency, as a marketing
channel owner, can track consumers at the individual level and
link together transactions made by the same consumer or house-
hold. This tracking is often based on loyalty IDs that are assigned
to the customers by using loyalty cards or online accounts, credit
card IDs, or other pieces of information that are available to a re-
tailer. This process, however, is often imperfect, and a significant
number of transactions can remain anonymous.
• Promotions can be distributed through the marketing channels

on behalf of both manufacturers and retailers. Distribution can
be done either in batch mode, when emails or printed catalogs
are sent to a large number of customers, or in real-time mode,
when promotions are generated in the scope of an individual
transaction, such as an in-store purchase or website visit.
• The main decisions that a targeting system needs to make with

respect to promotions are who are the right recipients for a pro-
motion, what are the right promotional properties, what is the
optimal time to offer it, and what is the right delivery channel.
• We assume that a retailer can identify consumers who have re-

ceived a promotion, consumers who have purchased a promoted
product, and, optionally, promotion redemption events. Note
that purchases and redemptions are completely different events
that should not be confused: consumers who have a promotion
3.1 environment 79
are not obligated to redeem it, and a product can typically be

purchased by any consumer although the purchase may be on
different conditions according to the granted promotions. Be-
yond these events, a targeting system can also access additional
or external consumer data, such as demographic records or
survey answers.
In the environment described above, the interactions with the con-
sumer are often structured as marketing campaigns, which is a conve-
nient unit of optimization. We define a targeted campaign as a market-
ing action constrained by a budget or duration that aims to achieve a
certain business objective by distributing targeted offerings to existing
customers or prospects. A targeted campaign generally includes the
following activities:
planning Planning of a campaign typically starts with setting busi-
ness objectives that have to be achieved. The main properties
of the campaign, such as budget, duration, or promotion types,
should also be determined at the planning stage and can be de-
rived from the objective.
execution The execution stage includes evaluation of potential re-

cipients and making decisions about the right offerings, right
messages, right communication time, and right delivery channels
.
measurement Measurement of the performance metrics is a criti-

cally important activity that can run in parallel with the execu-
tion to enable dynamic adjustments.
Let us note that the simple cycle depicted in Figure 3.1 does not
fully reflect all important aspects of promotion management. First, the
picture becomes much more complex for the management of multi-
ple campaigns or campaigns with a complex structure, which is, of
course, often the case in practice. Different marketing actions taken in
the scope of one or multiple campaigns can interact with each other,
which makes decisions and measurements more difficult. This means
that it may not be sufficient to track only the immediate events as-
sociated with an action; instead, the entire customer life cycle should
be taken into account. We discuss how a programmatic system can
handle such effects later in this chapter. The second important consid-
eration is that the customer’s perspective of the targeting process is
different from the perspective depicted in Figure 3.1. Each individual
customer life cycle can include multiple interactions with a retailer or
manufacturer and potentially involve multiple channels. The chain of
such interactions, called the customer journey, should deliver a consis-

tent experience across all touchpoints and throughout the life cycle.
This aspect is a key consideration in campaign design that we will dis-
cuss in Section 3.6.1.
3.2 business objectives
Each marketing campaign is associated with certain costs and certain

gains for each participant of the process, including the customers,
retailers, manufacturers, and agencies. Conceptually, each campaign
should have a positive return on investment (ROI), defined as the
difference between gains and costs. The return on investment can be
predicted before campaign execution or measured after the campaign
is fully or partially executed. Predictive models typically estimate
the ROI as a function of the campaign parameters, which enables
economic optimization of the campaign.
The challenge is that campaign gains typically have a complex struc-
ture, which includes both monetary and non-monetary components,
as well as immediate and long-term effects. These effects can be diffi-
cult to measure and even more difficult to predict. In this section, we
will discuss some basic considerations regarding gains and losses, and
we will then continue with a more formal framework that can be used
for campaign modeling. This framework justifies the creation of target-
ing models, as described in the next section, and lays a foundation for
campaign optimization.
3.2.1 Manufacturers and Retailers
A marketing campaign can be initiated and sponsored by either a man-

ufacturer or a retailer. In many cases, they both benefit from harvesting
more sales and more loyal customers. However, the way manufacturers
and retailers collaborate heavily depends on the business domain and
the marketing strategies for specific products or product categories.
The details of these collaborations are important for our purposes be-
cause they influence how programmatic targeting services can be pro-
vided or used in the retail world.
The first important consideration is the customer relationship man-
agement strategy of the customer-base owner, which is typically a re-
tailer. Mass market retailers, such as CPG retailers, typically welcome
manufacturers to participate in the marketing process by requesting
manufacturer-sponsored campaigns. Such campaigns help manufacturers
to increase their share in a category and are also beneficial for a re-
tailer. On the other hand, high-end retailers, such as fashion or cosmet-
3.2 business objectives 81
ics stores, position themselves as personal assistants and see a lot of

added value in their merchandising services. This category of retailers
cannot allow third parties to interact freely with their customer base.
Instead, they purchase the inventory of products upfront and manage
their marketing processes to sell it off over time and with maximum
profits.
The second consideration is that many retailers offer private-label
products, which leads to a conflict of interests with manufacturers. In
the case of a promotion service, retailers and manufacturers can negoti-
ate special rules to avoid destructive competition in such situations, for
example, by excluding customers who are highly loyal to the private
label from targeting.
Finally, retailers are interested in maximizing revenues in a category.
Encouraging customers to switch from high-margin products to dis-
counted ones can be harmful.
3.2.2 Costs
The costs of a promotional campaign can be borne by either a man-

ufacturer or a retailer. In either case, the retailers and manufacturers
are looking to compensate for the campaign costs by a higher volume
of sales. In the CPG world, for example, manufacturer-sponsored cam-
paigns are a very common practice. Such campaigns are initiated by
the manufacturer, and the retailer tracks coupon redemptions during
the campaign and then invoices the manufacturer for the total of the
redemption costs, which is typically the sum of the following compo-
nents:
distribution costs This includes coupon design and printing
costs, marketing agency fees, and fixed costs associated with a
campaign.
coupon redemption costs This is a total nominal value of all pro-

motions. This value can be estimated as a product of the total
number of promotions to be distributed, the redemption cost of
a single promotion, and the expected redemption rate.
clearing-house costs Store coupons have a life cycle that im-

poses additional post-redemption costs. When a consumer hands
the cashier a coupon at the checkout counter, the cashier puts
it into the cash drawer or a special envelope. At the end of the
day, the coupons are added up as if they were cash and packed
in bags. These bags are eventually shipped to a third-party
clearing house. The clearing-house clerks sort these coupons,
often by hand, invoice the manufacturer, and send a check to
the retailer. This process results in substantial clearing-house

costs because a major retailer can collect millions of coupons
and the clearing-house fee per coupon can be comparable to the
discount value. For instance, the clearing-house fee per coupon
was around $0.10 in 2016, while the discount value for most
CPG coupons was in the range of $0.50–2.00.
This structure can vary depending on the business domain and cam-
paign type, but the estimation of monetary costs is typically straightfor-
ward. However, marketing actions are almost always associated with
some non-monetary costs or losses that can be more challenging to esti-
mate. One of the most basic examples is the so-called email fatigue, that
is, a decline in email-opening rates and customer dissatisfaction caused
by too frequent or irrelevant emails. Although such losses can be diffi-
cult to quantify directly, we will see later that it is possible to correlate
monetary metrics such as revenues with the marketing actions in ques-
tion and, thus, to quantify and predict the negative effects. These loss
estimates can then be accounted for in the net profit equations.
3.2.3 Gains
The gains associated with a campaign can be viewed from several per-
spectives. The most straightforward element is the increase in sales vol-
ume. Both manufacturer-sponsored and retailer-sponsored campaigns
incentivize consumers to make purchases, at the expense of the cam-
paign costs, so the basic equation that describes the campaign gain will
be as follows:
profit Q pP V q C (3.1)
in which Q is the quantity sold, P is the baseline unit price, V stands

for variable campaign costs per unit (average redemption, distribution,
and clearing-house cost), and C is the fixed cost of a campaign. Simplis-
tically speaking, a campaign can be considered successful if the sales
volume induced by the campaign Qc exceeds the sales volume without
the campaign Q0 to an extent that is sufficient to cover the campaign
costs:
Qc pP V q C ¡ Q0 P (3.2)
Manufacturer-sponsored campaigns aim to achieve this goal in the

context of a given product and also to increase the market share of
the manufacturer in the corresponding product category in the long
run. At the same time, manufacturer-sponsored campaigns are typi-
cally beneficial for the retailer as well for the following reasons:
• Promotions stimulate return shopping trips. Manufacturers and

retailers share the goal of stimulating more shopping trips, so
promotional campaigns are often geared towards this mutually
beneficial goal.
• Promotions increase basket size. Some promotions are explicitly

designed to make people buy more of a given product. Other
types of promotions can decrease the consumer’s expenses and
release money for additional purchases.
• Promotions improve loyalty to the retailer. It is natural for a con-

sumer to perceive promotions as a result of a collaboration be-
tween the manufacturer and the retailer, so both are credited for
their effort in adding value and improving the consumer experi-
ence.
Consequently, the retailer benefits from a collaboration with a
manufacturer because of both improved revenues and loyalty effects.
This is the reason why most CPG retailers provide promotions as
a service to manufacturers. Retailer-sponsored campaigns typically
aim to promote private labels, promote entire product categories, or
stimulate the inventory turnover. From a promotional perspective, the
gains of retailer-sponsored campaigns are similar to the gains of the
manufacturer-sponsored campaigns described above. However, the
inventory turnover perspective is different and we will discuss it in
Chapter 6 in relation to price and assortment optimization.
The principle of sales volume maximization outlined in equation 3.2

is an important criterion for the campaign design, but it is, of course, a
very simplistic view of customer relationship management. We need to
do a more thorough analysis of campaign gains to better understand
the objectives that we can use in the design of targeting models and
campaigns. As a campaign aims to change the relationship with a con-
sumer, its objectives can be better understood by studying a customer
life cycle. We distinguish three main phases of interaction between a
consumer and a brand (manufacturer or retailer) that follow one an-
other, sometimes repeatedly:
• A consumer initially does not interact with the brand and prefers
other brands or totally different product categories. The main
goal of the brand at this phase is to acquire a new customer.
• Customers who interact with the brand can be incentivized to

buy more products. Promotion campaigns for these customers
typically follow up-sell or cross-sell methodologies. In the case
of up-selling, the promotion offers a stretch in comparison to the

normal quantities the customer purchases from the brand. In the
case of cross-selling, the promotion incentivizes the customer to
buy related products.
• Finally, a customer can stop interacting with the brand. This is
typically referred to as customer attrition, customer defection, or cus-
tomer churn. The cost of retaining an existing customer is typically
much less than acquiring a new one, so a brand can offer special
deals to customers who are about to churn.
These simple considerations provide a very important framework for
customer relationship management and promotion targeting in partic-
ular. First, let us note that consumer behavior and business objectives
are very different at each of the three stages of the life cycle, as depicted
in Figure 3.2.
Figure 3.2: Phases of a consumer life cycle.
Consumers who are at the first stage need to be acquired and con-
verted into customers by using marketing actions that are specifically
designed for this purpose. Customers at the second stage need to be
treated with incentives that are focused on maximization and growth
of consumption. Finally, the customers who are about to churn need
to be identified in a timely fashion and retained. These three objectives
– acquisition, maximization, and retention – are a very popular coordi-
nate system in marketing that can be used to orient individual cam-
paigns and structure campaign portfolios. A brand should be able to
distinguish consumers who belong to different phases, and this lays
the foundation for the targeting process. As we will see later, each of
these objectives can be mapped to a predictive model in a relatively
straightforward way, so this set of targeting objectives is well suited for
programmatics.
3.3 targeting pipeline 85
From the programmatic standpoint, a life-cycle-driven campaign de-

sign requires two fundamental problems to be solved. The first is the
identification of consumers who have a high propensity to move along
the life-cycle curve, as illustrated in Figure 3.3. If we assume that we
can quantify this propensity, we can determine the right consumers to
communicate with in order to achieve the objective and maximize the
gains.
Figure 3.3: Moving customers along the life-cycle curve.
Although targeting of the right consumers can potentially improve

the efficiency of marketing actions, it is not sufficient to quantify the
expected gain. Estimation of the expected gain is the second major
problem, and it requires not only prediction of the propensity of a
consumer to move to a certain point of the life-cycle curve but also
estimation of the total value that will be collected from a customer
after this point. This value corresponds to the area under the life-cycle
curve. The metric that we would ultimately like to measure is, however,
not the total gain, but the incremental impact of the marketing action
relative to the no-action alternative. This incremental gain corresponds
to the shaded area in Figure 3.4. In other words, a campaign model
needs to predict the incremental gain as a function of the expected
value with a no-action strategy and the expected value delivered by the
campaign, which, in turn, is defined in terms of propensity to respond
and potential impact. We develop a more formal framework for this
type of modeling and measurement in the next section.
3.3 targeting pipeline
Once the environment and business objectives are defined, we can dis-
cuss how a programmatic system can approach the problem of target-
Figure 3.4: Incremental impact of a marketing action. The upper life-cycle curve
corresponds to the aftermath of the action and the lower curve cor-
responds to the result with a no-action strategy.
ing and campaign management. This problem can be viewed as the

creation of a process that takes a marketing budget and business ob-
jectives as parameters, breaks them down them into campaigns, and
executes the corresponding marketing actions. This process can be de-
signed differently depending on how exactly the targeting system is
used, but a conceptual design of the process can often be represented
as a pipeline similar to the one shown in Figure 3.5.
Figure 3.5: A conceptual view of the targeting pipeline.
The pipeline starts with the available marketing budget that can be
allocated for different marketing activities. The first step of the process
is to determine how the budget should be distributed across the possi-
ble activities: what are the main objectives and how are these objectives
3.3 targeting pipeline 87
balanced? The output of this step is a set of objectives, such as acqui-

sition of new customers for product A and retention of customers for
product B, as well as budgeting parameters that should be followed
by the subsequent stages of the process. The second step is the evo-
lution of each objective into marketing campaigns, that is, campaign
design. A programmatic system can use a repository of campaign tem-
plates that are selected and parametrized based on the objectives. Each
campaign requires a number of decisions to be made to determine the
customers to be targeted, optimal targeting time, message parameters,
and so on. This generally requires the use of predictive models that
need to be trained and linked to campaigns. The models produce rel-
evance scores and other signals that can be used by the optimization
process to determine the optimal parameters of the campaign, such
as the list of customers to be targeted or maximum discount amount.
Finally, the campaign is executed and the collected data are used for
further optimization and measurement of the results. A programmatic
system should be able to run the pipeline both in simulation mode to
evaluate different strategies and in execution mode to do real target-
ing. To summarize, the targeting pipeline includes four major sets of
controls: budget allocation, campaign design, modeling, and execution
optimization. We discuss these controls in the next sections, starting
with models that are used as the basic building blocks, moving on to
campaign design and optimization, and finally considering the overall
budgeting and budget allocation.
The targeting pipeline can be viewed not only from the engineering
perspective outlined in Figure 3.5 but also from the end user (marketer)
perspective. This perspective is very important because it describes the
top-level functions and features of a programmatic system. The inter-
face of the system heavily depends on particular applications and busi-
ness environments, but we can consider a simple hypothetical example
to illustrate the basic principles. This hypothetical campaign manage-
ment flow is sketched in Figure 3.6.
This flow includes four principal steps. With the assumption that the
budget allocation steps are done in advance, the entry point to the cam-
paign creation workflow is the objective selection, which allows one to
specify the promoted products and high-level marketing goals. The
system uses historical data, best practices, and predictive models to
identify opportunities and propose campaign strategies. The expected
outcomes of the campaigns such as costs and gains are predicted, so
the marketer can select the optimal option. Once the campaign tem-
plate is selected and campaign parameters are estimated by the system,
the customer experience and creative assets, such as marketing mes-
sages, fonts, and images, are created or customized. This step clearly
Figure 3.6: A conceptual view of the campaign management workflow.
requires human involvement. Finally, a completely specified campaign

can be executed. Again, this is a hypothetical and oversimplified flow,
but it demonstrates what are we trying to achieve, ideally, in designing
a targeting system.
3.4 response modeling and measurement
Before we go deeply into modeling and campaign design, let us review

some basic principles of campaign response modeling and measure-
ments to get some idea of the role of modeling and optimization in the
targeting system. In this overview, we introduce new concepts, such as
response probability, but do not specify how exactly these values can
be modeled and predicted; these details are left for a later discussion.
Our goal here is to demonstrate how campaign costs, revenues, and
the statistical properties of customers fit together in one model.
Promotions and advertisements aim to change consumer behavior
and influence consumers’ decisions by urging them to make more
purchases, buy promoted products, and so on. Consequently, the suc-
cess of a campaign can be defined in terms of the response, which can
be measured with some simple metrics, such as the percentage of re-
deemed promotions, or with more complex measurements that include
direct and indirect, tangible and intangible gains. These metrics can be
3.4 response modeling and measurement 89
predicted before the campaign execution for optimization and decision-

making purposes or can be measured after the fact based on the col-
lected response data. These two problems are equally important, and
we discuss them separately in the subsequent sections by using the
principles of life-cycle-based modeling.
3.4.1 Response Modeling Framework
The response modeling framework is a simple generic framework that

helps to decompose the campaign modeling problem into several sub-
routines. This framework can be modified and extended to accommo-
date the complexity of real-life marketing campaigns. We assume a
relatively abstract setting in which a brand needs to optimize the dis-
tribution of a promotion or some other treatment across a population
of consumers by selecting the most promising candidates to receive
the treatment in order to maximize the overall value of the campaign.
We do not specify the exact meaning of the value at this point, defer-
ring this discussion until the following sections, but we assume it to be
some quantifiable measure that can be compared to costs. The acquisi-
tion, maximization, and retention tasks can be considered as variants
of this problem.
Recall that the basic marketing optimization problem is defined as
finding the strategy that maximizes the value function. In the case of
campaign response, we model the overall value of the campaign in
terms of the probability of response and the expected net value from a
customer. Our subject of optimization will be the set of customers who
receive the promotion, that is, the audience of the campaign:
Uopt argmax GpUq (3.3)

U P
in which P is the entire population of consumers, U is the subset

of consumers reached in the scope of the campaign, and GpUq is the
expected profit of the campaign, which is a function of the targeting
strategy that selects U from P. The expected profit of the campaign can
then be modeled as follows:
¸
GpUq Pr pR | u, T q pG pu | Rq Cq
P
u U (3.4)
p1 Pr pR | u, T qq pCq
in which PrpR | u, T q is the probability of a response to the treatment
(promotion) T from customer u, Gpu | Rq is the response net value
for customer u, and C is a cost of the promotion resource. The first
term corresponds to the expected gain from a responding consumer,
and the second term corresponds to the expected loss of sending a

promotion to which there is no response. The objective is to maximize
the expected profit by finding the subset of customers that are likely
to respond in the most profitable way. Equation 3.4 can be reduced as
follows
¸
GpUq PrpR | u, T q Gpu | Rq C
P
u U
¸ (3.5)
E rG | u, T s C
P
u U
in which E rG | u, T s denotes the expected value for a given con-

sumer on the assumption that the consumer will receive the promotion.
Consequently, the customer selection criteria can be simplified to
E rG | u, T s ¡ C (3.6)
because the expected net value is non-negative and all consumers

are assumed to be independent. Next, the optimal subset of customers
U can be determined as a subset that maximizes the value:
¸
argmax GpUq argmax E rG | u, T s C (3.7)

U P
U P P
u U
Note that this approach can be interpreted as the maximization of

the targeted net value relative to random resource distribution. To see
this, let us compare these two options with the assumption that a fixed
number of customers |U| will be participating in a campaign. The in-
cremental value of a targeted campaign relative to a campaign that dis-
tributes incentives among |U| customers selected at random is given
by
¸
argmax E rG | u, T s C | U | E rG | T s C

U P u U P ¸
argmax E rG | u, T s E rG | T s (3.8)
U P P
u U
¸
argmax E rG | u, T s
U P P
u U
in which E rG | T s is the average net value per customer over the

population. This average net value is constant; hence, it can be omitted
if the fixed cardinality |U| is assumed. On the other hand, we obtain
exactly the same result by reducing equation 3.7 under the assumption
that |U| is fixed, and, consequently, the cost can be dropped:
¸ ¸
argmax E rG | u, T s C argmax E rG | u, T s (3.9)

U P u U P
U P P
u U
In other words, the random selection of recipients represents the

baseline, and the value maximization problem is equivalent to the re-
distribution of promotions from one group of consumers to another.
It can be argued that the model defined by equation 3.7 is imperfect

because it favors consumers who are likely to respond to a promotion
but does not take into account customers who are likely to respond
anyway, thereby yielding the same profit even without the promotion
[Radcliffe and Surry, 1999; Lo, 2002]. Consequently, the actual profit
uplift of a promotion campaign in comparison to the no-action base-
line can be very small or even negative. Another way to think about
this problem is as an experiment in which we divide the set of cus-
tomers identified by targeting equation 3.7 into two groups, send the
promotion to one group but not the other, and then compare the out-
come from the two groups. It may be the case that customers in the
first group actively redeem the promotion and purchase the product
but that customers in the second group purchase the product equally
as much or even more. This campaign is clearly inefficient or even
harmful. To understand this problem better, let us separately consider
the following four possible strategies:
1. Select a set of customers |U| according to equation 3.7 and send
promotions to everyone in this group
2. Select a set of customers |U| randomly and send promotions to

everyone in this group
3. Select a set of customers |U| according to equation 3.7 but do not

actually send the promotions
4. Select a set of customers |U| randomly but do not actually send

the promotions
Each of these strategies yields a certain profit for the selected group
of customers |U|, so let us denote the profit of the i-th strategy as Gi .
Equation 3.7 maximizes the difference G1 G2 , that is, the lift from
targeting compared to random distribution. The alternative approach,
known as differential response analysis or uplift modeling, is to maximize
the uplift metric, defined as follows
uplift pG1 G2 q pG3 G4 q (3.10)
which measures not only the lift compared to random distribution but
also the lift compared to the no-action baseline for the same set of
customers [Berry, 2009]. In this case, equation 3.7 transforms into
¸
argmax E rG | u, T s E rG | u, Ns c (3.11)

U P P
u U
in which the second term corresponds to the expected net value

for customers who were not provided with the promotion. The dif-
ference between equations 3.7 and 3.11 can be illustrated by the fol-
lowing problem: Should a retailer offer a discount coupon on potato
chips to a person who buys potato chips every day? This question
would most likely will be answered in the affirmative according to
equation 3.7 because the person is likely to redeem the coupon. How-
ever, it is more probable that the customer would just buy the same
amount of potato chips for a lower price, which basically decreases the
retailer’s profit. Equation 3.11 alleviates this problem by incorporating
default customer behavior. By generalizing this example, we can cat-
egorize customers based on the difference between the probability to
respond given the treatment and the probability to respond without
the treatment, as shown in Figure 3.7.
Figure 3.7: Customer categorization with respect to uplift.
Analysis of the probability differences suggests four customer types

[Radcliffe and Simpson, 2007]. First, customers who have low proba-
bility of response, regardless of treatment, can be considered as Lost
Causes, who are probably the wrong targets for communications. Cus-
tomers who are likely to respond whether treated or not can be consid-
ered as Sure Things, who also seem to be the wrong targets. Customers
who tend to be driven away by treatment, often referred to as Do-not-
disturbs, should also be excluded from targeting. Finally, customers
who are more likely to respond only if treated are the most valuable
targets, the Persuadables. We will dive deeper into uplift modeling and
profitability optimization in the following sections.
3.4.2 Response Measurement
The response modeling framework provides a basic tool for response

prediction. The counterpart of this framework is a measurement frame-
work that can be used to evaluate the results of a campaign – did it
really help to acquire new customers, did it make existing customers
spend more, or did it help to improve retention rates? We want to
measure the effectiveness in terms of the ROI, which is defined as the
incremental gain delivered by the campaign relative to the no-action
strategy. This approach is consistent with the principles of life-cycle-
based targeting and uplift modeling that we stated in the previous
section.
The standard approach to measure the incremental gains is to com-
pare the performance of two groups of consumers: ones who received
the promotion (test group) and ones who did not receive it (control
group). If we have a targeting model in place, both groups are typi-
cally selected from the high-propensity customers to make sure that
the groups are statistically consistent and that the measured uplift re-
flects the impact of the promotion, regardless of the targeting strategy.
This approach is typically implemented by excluding a small percent-
age of customers from the targeted audience at the very end of the
targeting process, as illustrated in Figure 3.8.
Figure 3.8: Measuring promotion effectiveness with test and control groups.
The performance of the groups is compared over some period of

time that follows the campaign and is typically as long as several pur-
chase cycles for a given product category to obtain stable results. Note
that this approach does not require individual promotion redemptions
to be tracked – we do not care about the redemption rate and simply
need to compare the expenditures between the two groups. This can
be a very convenient property if the redemption data is not available.
We will return to the statistical details of measurements at the end of
this chapter.
3.5 building blocks: targeting and ltv models
Targeting models and lifetime value (LTV) models are the basic build-
ing blocks of the targeting process. The purpose of a targeting model
is to quantify the fitness of a given consumer for a given business ob-
jective in a given context. For example, a model can score the fitness
of a consumer for a potato chips promotional campaign given that the
promotion will be sent tomorrow by SMS. Models can be created for
different objectives and contexts, and a targeting system often main-
tains a repository of models attributed with the corresponding meta-
data, so that relevant models can be fetched according to the criteria.
For example, one can have a model for acquisition campaigns in the
potato chips category and another one for maximization campaigns in
the soda drinks category. Models are the basic primitives that can be
combined with each other, as well as with other building blocks, to cre-
ate more complex programmatic flows. Marketing campaigns can be
assembled from models, and marketing portfolios can be assembled
from campaigns.
It should be stressed that a programmatic system can use models
both in predictive and prescriptive ways. The most direct application is
prediction of consumer properties, such as propensity to respond to
an email or expected lifetime profit. Many models, however, express
the dependency between inputs and outputs in a relatively transpar-
ent way, so the system can contain additional logic that uses this pre-
scriptive insight or can at least make some recommendations to the
marketer. For instance, the parameters of a regression model that pre-
dicts a response can indicate a positive or negative correlation with
specific communication channels or other parameters, and this insight
can be used to make additional adjustments, such as limiting the num-
ber of communications in the presence of a negative correlation. In this
section, we consider three major categories of models that can be used
separately or together:
propensity models The idea of propensity models is to estimate
the probability of a consumer to do a certain action, such a pur-
chase of certain product. The output of such models is a score
proportional to the probability that can be used to make target-
ing decisions.
3.5 building blocks: targeting and ltv models 95
time-to-event models Propensity models can estimate the proba-

bility of the event, but they do not explicitly estimate when the
event is most likely to happen. This type of estimate is more
valuable in many marketing applications and requires a different
statistical framework to be used.
lifetime value models LTV models are used to quantify the value
of a customer and estimate the impact of marketing actions.
We start with a review of the data elements and data sources
used in modeling and then discuss several traditional methods that
can be viewed as heuristic propensity and LTV estimation models.
These methods typically make the assumption that the probability
of responding and the value of a customer are proportional to one
or several basic characteristics, such as the frequency of purchases.
The methods then group customers into segments so that an entire
segment can be included or excluded from a certain campaign. These
methods can be thought of as rule-based targeting. We then develop
more advanced models with statistical methods.
3.5.1 Data Collection
Data collection and preparation is one of the most important and chal-
lenging phases of modeling. Although a detailed discussion of data
preparation methodologies is beyond the scope of this book, it is worth
reviewing a few principles that can help to streamline the process and
avoid incorrect modeling. Targeting and LTV models generally aim
to predict consumer behavior as a function of observed metrics and
properties, so it is important to collect and use data in way that is con-
sistent with causal dependencies. From this perspective, data elements
can be arranged by tiers, with each tier depending on the previous
ones [Grigsby, 2016]:
primary motivations Consumer behavior is driven by fundamen-
tal factors such as valuation of a product or service, tastes, needs,
lifestyle, and preferences. Many such attributes cannot be ob-
served directly, but some data such as demographics or market-
ing channel preferences can be collected through loyalty program
registration forms and surveys or purchased from third-party
data providers.
experiential motivations The next tier of properties is created

by the interaction between a customer and a brand. These proper-
ties characterize the overall customer experience, including cus-
tomer satisfaction, loyalty, and usage patterns. Some of the ex-
periential properties can be directly or indirectly quantified by

using metrics like frequency of purchases.
behaviors The most important category of data is explicitly

observed behavioral data, such as purchases, website visits,
browsing history, and email clicks. These data often capture
interactions with individual products at specific points in time.
Behavioral data carry the most important signals needed for
modeling.
results Finally, customer actions result in directly observed financial

metrics, such as revenue or profits. It is important to keep in
mind that these metrics do not really explain the drivers of the
customer behavior and the ways it can be influenced; they merely
register the final outcomes.
The data described above should also be linked with additional di-
mensions, such as catalog data, seasonality, prices, discounts, and store
information. It is important to provide the ability to aggregate the data
at different levels of hierarchical dimensions to find the optimal level
of granularity. For example, a model can use data aggregated at the
product, category, or department levels.
The modeling process should generally be focused on revealing the
hidden properties and causality, not just analysis of the results. Al-
though analysis of financial results is important, it is typically a good
idea to establish the link between marketing actions and these results
through behavioral concepts. For example, a solution that expresses the
revenue as a function of advertising intensity is not necessarily insight-
ful and actionable enough. A solution that quantifies how advertising
influences customer loyalty and behavioral patterns (e. g., migration
from one customer segment to another) and then links customer prop-
erties to the revenue is likely to be more insightful and actionable.
3.5.2 Tiered Modeling
Targeting models score the relevance of a customer for a business objec-

tive and context, based on the features derived from a customer profile.
One of the most basic approaches is to use only one metric, such as the
monthly average dollar amount spent on a brand or category. This
metric can then be used in two ways. First, it is a measure of proxim-
ity between the customer and the promotion because promotions are
typically created for a specific brand and category. Consequently, the
most relevant promotions for a given customer can be selected based
on the brand and category with the highest monetary metrics. Second,
consumers can be sorted by the metric and the most valuable ones can
be selected for a given promotion. The classic example of this approach

is tiered segmentation, whereby consumers are assigned to gold, silver,
and bronze tiers depending on their score and heuristically selected
threshold, as illustrated in Figure 3.9.
Figure 3.9: Example of segmentation by a monetary metric (gold–silver–bronze

tiers). Customers are sorted by their spending in the category over
some fixed time period, and the top 5% are assigned to the gold tier.
The next 10% are assigned to the silver tier and the next 20% to the
bronze tier. The remaining customers are not eligible for promotions.
Each tier is attributed with the metrics, such as the average expected
response rate and average spending per customer, estimated based on
the historical data. For each promotion, the optimal subset of tiers
can be determined by running the promotion costs and tier metrics
through the response modeling framework. For example, it can be de-
termined that one campaign will be profitable if only the gold tier is
targeted and another campaign has maximum profit when the gold
and silver tiers are both targeted.
Single-metric segmentation can be elaborated by adding more met-
rics into the mix. As we discussed earlier, the consumer life cycle is an
important consideration for the design of promotional campaigns, so
the ability to target individual life-cycle phases is important. The life-
cycle phases are characterized by both the total spending in a category
and the loyalty to the brand, that is, the relative spending on a brand
compared with that on other brands, so we can classify customers into
segments by using these two metrics, as shown in Figure 3.10. This ap-
proach, known as loyalty–monetary segmentation, is used in traditional
manufacturer-sponsored campaigns.
Customers who are highly loyal to the brand and spend a lot of
money in the category are clearly the most valuable customers, who
should be rewarded and retained. Customers who spend a lot in the
category but are not loyal to the given brand are the best candidates for
trial offers and so on. Similarly to the tiered segmentation, promotions
can be assigned to the optimal subset of segments by starting from the
upper left corner of the grid in Figure 3.10 and evaluating the potential
outcome of including additional segments until the bottom right corner
is reached. This approach can be considered as a simplistic targeting
method that predicts consumer value based on two metrics – spending
Figure 3.10: Example of loyalty–monetary segmentation.
in the category and brand share of wallet. These are very coarse criteria
that can be improved by using predictive modeling methods.
3.5.3 RFM Modeling
Another popular segmentation heuristic is so-called recency–frequency–

monetary (RFM) analysis. It is similar to the loyalty–monetary approach,
but uses three metrics:
recency The number of time units that have passed since the cus-
tomer last purchased. This metric can be measured directly in
time units (e. g., months) or can be mapped to some score. For
example, customers can be sorted by the most recent purchase
date and then those in the most recent 20% are assigned a score
of 5, the next 20% a score of 4, and so on, until the last 20% get a
score of 1.
frequency The average number of purchases per time unit. Again,

the metric can be measured directly in units or scores.
monetary The total dollar amount spent per time unit. The mone-
tary metric is typically measured by using intervals or scores.
It is quite typical to use the same discrete scoring scale, say from 1
to 5, for all three metrics. In this case, the RFM model can be consid-
ered as a three-dimensional cube made up of cells, each of which is
determined by a triplet of metric values and corresponds to customer
segments. The targeting decisions can be made by selecting a subset

of segments from the RFM cube. One possible technique is to sum all
three metrics into a single score and select customers whose score ex-
ceeds a threshold – this corresponds to cutting a corner away from the
RFM cube.
RFM analysis is based on the empirical observation that recency, fre-
quency, and monetary metrics are often correlated with the probability
to respond and the lifetime value. Although this is a fair assumption,
the RFM approach is shallow because it measures the final outcomes
of the marketing processes and consumer actions, not the factors that
impact consumer behavior. As we will see in a subsequent section, a
more flexible solution can be obtained by using clustering.
3.5.4 Propensity Modeling
Simple segmentation models such as tiered segmentation and RFM

analysis can be viewed as particular cases of regression analysis with
a very limited number of features and heuristic assumptions about the
relationships between the metrics and the expected outcomes. Our next
step will be to build more formal scoring models.
The goal of propensity modeling is to find consumers who have a
relatively high probability of behaving in a certain way or committing
a certain action in the future. The number of actions that can be pre-
dicted and used in targeting is very high. Let us consider a few typical
examples:
propensity to try a new product Consumers who currently do
not buy a certain product but have a high propensity to buy it in
the future are good targets for acquisition campaigns.
propensity for category expansion Consumers who have

high propensity to switch from one category of products to
another or to try a new category are good targets for up-selling
or cross-selling campaigns. An example of such an audience
are consumers who are likely to switch from casual to luxury
products.
propensity to buy more Consumers who are likely to increase

their average purchase quantity of a product are the right targets
for maximization campaigns.
propensity to churn Customers who are likely to unsubscribe

from a service or stop buying a product can be targeted in
retention campaigns.
propensity to engage Propensity to engage is the probability of

responding to a marketing action, for example, to click on an
email link.
propensity to change shopping habits Each customer has

shopping habits that eventually determine their customer value
– how often the customer buys, which products, from which
categories, and so on. These habits are generally stable over
time, and once a brand manages to change a customer’s level
of engagement, this level tends to last. Consequently, brands
are generally interested in finding customers who are open to
changing their habits, for example, people who moved from
one city to another, graduated a school or university, just
married, and so on. The canonical example of such modeling
is Target’s attempt to predict customer pregnancies in the early
stages because births obviously change the way customers shop
[Duhigg, 2012].
Note that the major marketing objectives of acquisition, maximiza-
tion, and retention can be expressed by using propensity language. The
propensity-based approach is convenient from the response modeling
perspective because the campaign ROI can be estimated by multiplying
the expected gains and losses by the predicted outcome probabilities.
3.5.4.1 Look-alike Modeling

Look-alike modeling is one of the most important methods of propen-
sity modeling. Look-alike modeling is based on the observation that
propensity is essentially the probability of a customer moving from
one point on the life-cycle curve to another, so one can train a predic-
tive model by using profiles of consumers who exhibited this behavior
in the past and then evaluate the model against the current profile of a
given customer to estimate their propensity. For example, the profiles
of customers who used not to buy a given product for a while and then
started to buy it can be used to train a model that identifies customers
with a high propensity to try this product for the first time.
Look-alike modeling is a classification problem, so it requires the
profile features and response label to be specified. We assume that a
customer profile can include individual attributes, such as income or
household size, and a collection of behavioral events, each of which is
attributed with a time stamp. For each profile, we put all of the events
on a time line and specify three sequential time frames: the observa-
tion period, buffer, and outcome period. These frames are shown in
Figure 3.11. The observation period is used to generate features, and
the outcome period is used to generate the response label. These two in-
tervals may or may not be separated by a buffer. The buffer can be used
if one needs to predict events in the relatively distant future, instead of
predicting immediate events. For example, a model that predicts cus-
tomer churn should probably be trained with the outcome intervals
shifted into the future – it would be impractical to predict customers
who are likely to churn immediately because it gives no time to per-
form any mitigating marketing action.
Figure 3.11: Look-alike modeling time frames.
The model is trained on a set of historical profiles that contain both

the observation and outcome intervals. The model is then used to eval-
uate the current profiles, which, of course, contain only the observed
part, and the expected outcome is predicted.
The design of features and response labels heavily depends on a

particular domain. We focus on retail applications in this section, and
feature engineering for the online advertising domain will be discussed
in later sections. Look-alike modeling can generally use any customer
data that is available including demographics, purchases, and market-
ing responses, such as email opens, clicks, and promotion redemptions.
Profile features are typically specified as different combinations of time
frames, metrics, and filters that can be applied to the profile data. This
process is illustrated in Figure 3.12 for purchasing data. First, the fea-
tures can be calculated for different time frames within the observa-
tion period. These subperiods are typically counted back from the end
of the observation interval – last month, last three months, last six
months, and so on. Within the subperiod, different metrics such as
dollar amount or purchase frequency can be calculated, and different
filters such as category, brand, product, payment type, or part of the
week can be applied. Finally, the value can be expressed in units, such
as dollars or days, percentages, per-basket averages, or binary yes/no
variables. The curved line that connects the boxes in Figure 3.12, for
example, corresponds to the share of the bakery category in consumer
spending over the last 6 months relative to other categories. This ap-
proach allows the production of a relatively large number of features
that can be used in predictive model training and evaluation. The same
approach can be used for marketing response data and data from digi-
tal channels.
Figure 3.12: Example of profile feature engineering for purchasing data.
The response label is generated from the outcome period according

to the objective. For example, if the model is created to target customers
with a high propensity to try a given product, then the response la-
bel will indicate whether this product was purchased or not. Another
example is a retention campaign where the response label indicates
whether a customer defected or not. The set of training profiles can
also be prefiltered according to the objective. In an example with a
propensity to try a new product, the training set should include only
customers who did not buy the product during the observation period
(and then some of them started to buy it during the outcome period –
such customers are called natural triers). The same rule is applied for
model evaluation – a customer who already buys a promoted product
is immediately excluded as a non-fit.
example 3.1
Let us run through a small example that illustrates look-alike model-

ing. Consider the following scenario: a CPG retailer with bakery and

dairy departments started to work with a new brand of dairy desserts
six months ago, and the brand has now asked to run a manufacturer-
sponsored acquisition campaign. The targeting goal is to identify cus-
tomers with a high propensity to try the dessert. Let us assume that
a targeting system generated a training data set with 12 historical cus-
tomers profiles and 5 features, as shown in table 3.1.
Bakery Bakery Dairy Dairy

ID Credit Response
Total Weekend Total Weekend
1 150 10 150 140 1 1
2 210 20 120 110 1 0
3 190 190 210 20 1 1
4 270 250 190 0 1 1
5 180 180 190 10 1 1
6 260 250 230 20 0 1
7 270 30 210 210 1 0
8 150 40 150 50 1 0
9 90 70 120 100 0 0
10 30 0 200 200 1 0
11 190 190 250 10 1 1
12 10 0 30 0 1 0
Table 3.1: Training data set for look-alike modeling.
The Bakery Total and Dairy Total features are the total spending in
the corresponding categories during the observation period. The Bak-
ery Weekend and Dairy Weekend features are the total spending on
the weekend, so the amount spent on workdays equals the difference
between the Total and Weekday values in each category. The Credit col-
umn is the payment method, credit or cash. Finally, the Response vari-
able indicates whether the customer started to buy the dessert or not.
By inspecting this tiny data set visually, we can conclude that the natu-
ral triers of the dessert are mainly the customers who buy a lot of bak-
ery on the weekends and a lot of dairy on workdays. We choose here
to use logistic regression to build a look-alike model, although other
options including decision trees, random forests, and Naive Bayes are
often used in practice as well. By fitting the logistic regression, we get
the parameter estimates presented in table 3.2.
Note that bakery spending is positively correlated with the propen-
sity to try the product, whereas dairy spending is negatively corre-
Parameter Estimate
Bakery Total 0.0012
Bakery Weekend 0.0199
Dairy Total – 0.0043
Dairy Weekend – 0.0089
Credit – 0.4015
Table 3.2: Logistic function parameters for the training set in table 3.1.
lated1 . By evaluating the model for six profiles with different propor-
tions of bakery and dairy spending, we get the propensity score esti-
mates shown in table 3.3. One can see that only customers with high
spending on bakery and low spending on dairy have a high propensity
to try the product, regardless of the payment method. In real life, one
possible interpretation could be that dairy desserts are considered as
substitutes for bakery desserts by customers who actively buy in both
categories.
Bakery Bakery Dairy Dairy Propensity to

Total Weekend Total Weekend Credit Try
10 0 50 50 1 0.26
20 20 200 200 1 0.07
150 20 100 30 1 0.37
250 20 190 30 1 0.31
250 200 190 30 1 0.94
250 200 190 30 0 0.96
Table 3.3: Predicted propensity to try the product.
Note that we do not use the historical response data as features in

this example, that is, we do not take into account whether a customer
used to respond to promotions in the past or not. In real life, this is an
important signal for targeting accuracy, although it is valid to create
models without response features if the data are not available.
N
3.5.4.2 Response and Uplift Modeling
The most basic look-alike models, similar to those we just described
in the previous section, estimate the unconditional probability of a cer-
tain action being done. By including marketing communications into
the set of features, we can create a propensity model that estimates
1 See Chapter 2 for a detailed discussion of logistic regression. In this example, we skip
typical steps, such as model validation and diagnostics, for the sake of simplicity.
the conditional probability of the response (action) given a marketing

treatment. One possible methodology for creating such models is pi-
lot campaigns. The idea of this approach is to distribute a promotion
initially to a relatively small group of recipients, collect the responses,
and create a classification model that maximizes the difference between
treated responders and treated non-responders. This corresponds to a
look-alike model trained on a population of treated profiles by using
the response indicator as a training label. This model estimates the
probability of the response given the treatment as
PrpR | T , xq (3.12)
in which R is the response indicator, T is the treatment indicator, and x

is the profile feature vector. Once the model is created, it can be used to
execute a full-blown campaign, that is, to target customers with a high
propensity to respond when given the treatment. In certain cases, the
model can be created by using historical data for similar campaigns
without running a pilot. Thus, traditional propensity models are de-
signed to identify customers who are likely to respond to a promotion
or some other marketing communication. The downside of this ap-
proach is that such models can target customers who are likely to take
an action anyway, even without the treatment. In other words, it can
be the case that the propensity model predicts a high probability of re-
sponding, but once the campaign is executed, the observed difference
between the test and control groups, that is, the uplift, is insignificant
or the control group could even outperform the test group. We have al-
ready discussed this problem in the context of the response modeling
framework, but we now need to dig deeper and specify how exactly
this can be addressed in propensity modeling.
The problem with uplift arises from the fact that the basic propensity
modeling process described above accounts only for treated customers
and discards the information about the non-treated ones. This makes
it structurally impossible to model the uplift. We can work around
this problem by adding a control group to the pilot campaign. This
group includes randomly selected profiles of customers targeted and
not targeted in the pilot campaign. In this case, we can observe the
results for four distinguishable groups: treated responders, treated non-
responders, control responders, and control non-responders, as shown
in Figure 3.13.
The four observed groups enable us to create a model that maxi-
mizes the uplift, that is, the difference between the response rates in
the test and control groups:
upliftpxq PrpR | T , xq PrpR | C, xq (3.13)

Figure 3.13: Measurement groups for propensity modeling with uplift [Kane
et al., 2014].
in which the first term is the probability of a response after treatment

and the second term is the probability of a response by an individual
in the control group. These two probabilities can be estimated either
by using two separate classification models trained on the test and
control groups, respectively, or one model trained on a union of test
and control profiles with the treatment indicator included as a feature
[Lo, 2002]. The problem with the two-model approach is that separately
created models may have incomparable score scales and may select
features that are not actually predictive for the uplift, so this solution
often fails to achieve better results than the baseline non-incremental
propensity model [Radcliffe and Surry, 2011; Kane et al., 2014]. The
single-model approach can achieve better results but may require more
complex model design. For example, if logistic regression is used as the
underlying modeling method, the feature vector should include both
the profile features and profile–treatment interaction terms, so that the
model has the functional form
fpx, IpT q x, IpT qq (3.14)
in which IpT q is the indicator function and is equal to one if the cus-
tomer x has been treated and zero otherwise [Lo, 2002]. Consequently,
the uplift is estimated as
upliftpxq fpx, x, 1q fpx, 0, 0q (3.15)
It can be argued that even more accurate results can be obtained by

creating a multinomial model that predicts probabilities for each of the
quadrants in Figure 3.13 [Kane et al., 2014]. We can create such a model
to express the uplift as follows:
upliftpxq PrpR | T , xq PrpR | C, xq

PrpR | T , xq p1 PrpN | C, xqq (3.16)
PrpR | T , xq PrpN | C, xq 1
in which N denotes a no-response outcome. By using the Bayes rule
and the fact that PrpT | xq PrpT q because the test and control groups
are randomly selected, we get
PrpT R | xq PrpCN | xq
upliftpxq 1
PrpT | xq PrpC | xq
(3.17)
PrpT q PrpCq
1
By repeating the same transformations for the response probability
in the test group, we can also express the uplift as follows:
upliftpxq p1 PrpN | T , xqq PrpR | C, xq
1 PrpPrT NpT q| xq PrPr

pCR | xq (3.18)
pC q
From the sum of equations 3.17 and 3.18, we obtain the final expression
for the uplift estimation:
2 upliftpxq
PrpT q PrpCq
PrpT N | xq PrpCR | xq
(3.19)
PrpT q
PrpCq
in which the probabilities in the numerators are estimated by using
a single regression model. The uplift score can often be used as an
alternative to the response probability estimated by basic propensity
models – as we will see later, a targeting system can optimize the cam-
paign ROI by selecting the recipients from those individuals with the
highest uplift score as opposed to propensity scores.
3.5.5 Segmentation and Persona-based Modeling
Behavioral segmentation is the process of dividing customers into

groups or segments in such a way that customers within a segment are
similar to each other but dissimilar to customers in other segments.
From the marketing analytics standpoint, segmentation is typically
one of the most important, valuable, insightful, and complex projects.
It typically aims to define a small number of well-differentiated

segments with clear semantic meaning that can support strategic
decisions. The output of the segmentation process typically includes
segment profiles and segment models, also called clustering models.
A segment profile includes the distinctive properties and metrics
of the segment and some interpretation of what a typical customer
persona looks like. A simplistic example of segment profiles is shown
in table 3.4. The set with the most distinctive properties is usually
identified by running clustering algorithms on a set of historical
customer profiles, so each segment corresponds to a group of existing
customers and the segment profile is a set of statistical metrics for this
group. Although a segment is initially just a list of existing customers,
it can be converted into a clustering model that is essentially a
classification rule that maps any given customer profile to a persona.
The model-based representation of a segment is important because
it can dynamically assign customers to segments depending on their
profile features.
Segment 1 Segment 2 Segment 3

Convenience Casual Bargain
Persona
seekers buyers Hunters
% of market 20 50 30
% of revenue 40 40 20
Share of clothing 40 60 60
Share of electronics 50 20 10
Share of toys 10 20 30
Redemption rate 0.02 0.05 0.08
Table 3.4: Example of segments and segment metrics. Each segment can be inter-
preted in psychographic and behavioral terms. Convenience seekers,
for instance, seem to be less price sensitive and have fewer children
than consumers in other segments. This segment contains a relatively
small number of customers but makes a high contribution to the rev-
enue.
Note that behavioral segmentation is very different from RFM anal-

ysis, although the RFM method can also be viewed as a segmentation
technique. RFM analysis segments customers based on the observed
financial results, whereas behavioral segmentation aims to identify the
traits that cause this result. In many cases, the resulting features, such
as spending, are deliberately excluded from the profile features before
clustering to make sure that segments are created based on the behav-
ioral cause, not the financial outcome. The second important distinc-
tion is that RFM analysis and its variations use a fixed set of features,
whereas segmentation is a method to identify the most discriminative
features. These properties of behavioral segmentation are very impor-

tant for strategic marketing analytics because they help us to under-
stand the drivers of customer behavior (e. g., why one segment has
more churners than another) and differentiate marketing strategies for
each segment by using its distinctive characteristics. This differentia-
tion can be taken to great lengths, such as the assignment of dedicated
segment managers for each segment.
The programmatic perspective on segmentation is different from that
we just discussed for marketing analytics because programmatics is
more focused on execution and tactical aspects, rather than strategy.
A programmatic targeting system is more often a user of the outputs
produced by the behavioral segmentation process. First, persona tags
are often used as features in look-alike modeling and other targeting
rules and models; it does not matter how exactly these tags are created.
Persona tags carry an important signal about consumer behavior and,
hence, can have substantial predictive power for propensity modeling.
The second important application of segmentation outputs is segment-
level modeling. Propensity models created for an entire population of
customers can have limited accuracy because propensities can be de-
termined by different factors. For example, customers in one segment
can churn because of low product quality, whereas customers in an-
other segment churn because of high prices. Consequently, the model
repository can maintain specialized models for different combinations
of objective, product category, and customer segment.
3.5.6 Targeting by using Survival Analysis
Propensity modeling provides a powerful framework for estimating

the probabilities of potential outcomes of marketing actions. This ap-
proach, however, has a number of shortcomings. The first issue is that
the probability of the event does not straightforwardly translate into
the time-until-event, which is usually a more actionable metric. For
example, it can be more useful to know that a customer is likely to
make a purchase in 10 days and that this time can be reduced by 5
days by offering a discount, rather than to know that the conditional
probability of purchase by a customer given a discount is 0.8. Note the
we cannot work around this problem by building multiple propensity
models for adjacent time intervals because the intervals are interde-
pendent. For example, we cannot build separate purchase probability
models for January, February, and March because purchasing events in
February depend on purchases in January and so on. The second chal-
lenge is that we do not always observe the outcomes needed to create
the response labels for propensity modeling. For example, a look-alike
model for a retention campaign can be trained to differentiate between

the customers who have and have not churned. The training data set
will include a number of profiles of customers who have churned dur-
ing some time period in the past and also a number of profiles for
customers who have not churned in the same period. This approach
is not perfect because customers who have not yet churned may still
churn in the future, so it is more accurate to say that their outcomes
are unknown, rather than positive. This is a problem of the censored
observations that we have already discussed.
These limitations of propensity modeling can be addressed by us-
ing survival analysis introduced in Section 2.6.2 . Survival models can
properly handle censored data, predict the expected time-to-event (sur-
vival time), and specify how marketing actions and customer proper-
ties can accelerate or decelerate the events. Let us consider a numerical
example that illustrates the basic usage of survival analysis in a target-
ing system.
example 3.2
Consider the scenario of a retailer who sets up a promotional campaign

in a programmatic system. To determine the optimal properties of the
campaign, the system uses the data set for a previous similar campaign
presented in table 3.5. This data set includes 12 customer profiles with
3 features: an indicator of whether a customer had made a purchase a
week before the campaign announcement, the number of emails sent
to the customer in the scope of the campaign, and the discount amount
offered to the customer. The observed outcome is the time-of-purchase
measured in days from the campaign announcement. The campaign
ended after 20 days, so all three customers who did not purchase before
the campaign had ended are considered censored.
We use this small data set to fit a Cox proportional hazard model, as
described in Section 2.6.2.3. Recall that the Cox model is a semipara-
metric model with a nonparametric baseline survival function, which
describes how the purchase times are distributed, and a parametric lin-
ear model for individual risk ratios. The risk ratio describes whether
the “risk” of purchase for a given customer is higher or lower than
the baseline. The risk ratio is also expressed as a function of profile
features, so it is possible to quantify how different features impact the
expected time-to-purchase. By fitting the Cox model, we get the follow-
ing model for the risk ratios:
logpriskq 1.957 Previous Purchase
0.510 Number of Emails (3.20)
0.323 Discount
Previous Number of
ID Discount, % Purchase Time
Purchase Emails
1 0 2 5 5
2 0 2 0 10
3 0 3 0 20 (censored)
4 1 1 0 6
5 1 2 10 2
6 1 3 0 15
7 1 4 0 20 (censored)
8 1 5 5 6
9 0 2 10 8
10 1 5 5 13
11 0 0 0 20 (censored)
12 1 2 5 8
Table 3.5: Training data set for survival analysis. The censored records corre-
spond to customers who did not make a purchase during the first 20
days after the campaign announcement.
The interpretation of this model is that the previous purchase and

discount are negatively correlated with the time-to-purchase and the
number of emails is positively correlated. In other words, additional
discounts decrease the time-to-purchase, but additional emails increase
it. This indicates that sending more emails actually harms the cam-
paign, so the emailing strategy and relevance of messages should as-
sessed and fixed. This part of the model is useful but does not yet pro-
vide additional insights over standard propensity modeling. The more
interesting part is survival functions. The Cox model can produce a
survival function for any given value of the feature vector, and each
function corresponds to the cumulative distribution of the purchas-
ing times. In this example, the feature vector is a three-element vector
with the previous purchase indicator, number of emails, and discount
depth expressed as a percentage. Examples of the survival functions
are shown in Figures 3.14 and 3.15. All of the curves have the same
shape, but they are scaled according to the risk ratio estimated based
on the feature vector x. One can see that the number of emails pushes
the survival curve up, which justifies and quantifies the inefficiency of
the communications. Meanwhile, the discount pushes the curve down,
which indicates that discounts decrease the time-to-purchase.
It is important not only to obtain the survival curves but also to es-
timate the statistical properties of the time-to-purchase. Recall that the
survival functions Sptq are directly related to the cumulative distribu-
tion functions of the time-to-purchase Fptq:
Sptq 1 Fptq (3.21)

Figure 3.14: Survival curves for different numbers of emails. The purchase indi-
cator and discount depth are equal to zero for all curves.
Figure 3.15: Survival curves for different discount depths. The purchase indica-
tor and numbers of emails are equal to zero for all curves.
Consequently, we obtain the distribution functions from the Cox

model as well. This enables us to estimate the average or median
time-to-purchase, as well as confidence intervals and other statis-
tical properties. As the survival function can be obtained for any
combination of the independent variables, we can estimate the av-
erage or median time-to-purchase for each customer separately and
then use these values in marketing rules (e. g., send a notification
one day before the expected time-of-purchase) or targeting scores
(e. g., target the ten percent of customers with the longest expected
time to purchase). We can also quantify the impact of the indepen-
dent variables in terms of the average or median time-to-purchase.
For example, it is possible to estimate how the average number

of days until purchase decreases with every percent of discount.
N
Survival analysis can be applied to a number of marketing activ-

ities. The most typical applications include estimating the right tim-
ing for a message in replenishment campaigns, estimating time-until-
churn in retention campaigns, and estimating the total number of pur-
chases over some time interval for lifetime value modeling. Similarly
to propensity models, survival models can be created for different
products, categories, and customer segments. The expected times-to-
purchase produced by these models can be compared with each other
and then the most relevant products and offerings can be selected
based on the ratio of times.
3.5.7 Lifetime Value Modeling
The last building block we will consider is estimation of the customer

lifetime value, commonly abbreviated as LTV, CLV, or CLTV. The goal of
LTV modeling is to estimate the total amount of money that a brand
is likely to make from a given customer over the lifetime of their rela-
tionship. The exact structure of the LTV analysis heavily depends on
the business model of a given brand, but it is possible to create basic
LTV models that can be customized with brand-specific profit and loss
terms.
The LTV is an important building block in campaign design and
marketing mix management. Although targeting models can help to
identify the right customers to be targeted, LTV analysis can help to
quantify the expected outcome of targeting in terms of revenues and
profits. The LTV is also important because other major metrics and
decision thresholds can be derived from it. For example, the LTV is
naturally an upper limit on the spending to acquire a customer, and
the sum of the LTVs for all of the customers of a brand, known as the
customer equity, is a major metric for business valuations. Similarly to
many other problems of marketing analytics and algorithmic market-
ing, LTV modeling can be approached from descriptive, predictive, and
prescriptive perspectives. We start with the basic descriptive approach
and then develop more advanced models.
3.5.7.1 Descriptive Analysis

The LTV typically accounts for all revenues derived from a customer
and variable costs associated with the relationship, and it can option-
ally include the cost of customer acquisition. One of the most basic
ways to estimate the lifetime value of customer u is to sum the average

expected profits for some time interval in the future:
Ţ
LTVpuq pR C q T pR C q (3.22)

t 1
in which time t is measured in some units, typically months, R and

C are the average expected revenues and costs, respectively, per cus-
tomer per time unit, and T is the expected lifetime or projection hori-
zon. The average expected revenues and costs are typically estimated
based on historical data, such as transaction histories and campaign
budgets. This estimate is not personalized (it is averaged over all cus-
tomers) and is thus relatively straightforward. The revenues and costs
can differ sharply for different customer segments, so it is very com-
mon to estimate R and C for each segment separately and then calcu-
late segment-specific LTVs if the customer segment (persona) is known.
The lifetime duration T can also be selected heuristically based on a typ-
ical relationship duration or planning horizon, often 24 or 36 months.
Basic LTV formula 3.22 does not account for several major effects.
First, it does not explicitly account for customer retention. Although
we can adjust the time horizon T according to the average customer
lifetime, it can be more convenient to include the customer retention
rate r into the formula as a parameter. For example, an annual retention
rate of 0.8 means that 20% of the present customers will terminate the
relationship within a year. Second, the LTV is typically measured for
relatively long time intervals of 2–3 years, so we might need to account
for the fact that money in the present is worth more than the same
amount in the future, that is, for a discount rate d. The discount rate
reflects the cost of tying up the capital for a period of time. For example,
an annual discount rate of 0.15 means that the present value of $1
should be considered equivalent of $1.15 to be received in one year.
Accounting for the discount rate gives the net present value of the LTV.
By taking these two factors into account, we can estimate the customer
net profit as pR Cq for the first time period, pR Cq r{p1 dq for
the second time period, and so on, which eventually adds up to the
following definition of the LTV [Berger and Nasr, 1998] :
LTVpuq
Ţ
pR Cqrt1
t1
p 1 d q t 1 (3.23)
This formula is widely used and can be considered as a standard

definition of the LTV. This expression, of course, does not include all of
the effects that can be found in real life, and it can be extended to reflect
other processes and parameters that influence the LTV. For example,
we can model the net profit m not simply as a constant value R C but
as a value that gradually increases over time as the relationship with
the consumer matures:

mt m0 pmM m0 q 1 ekt (3.24)
in which m0 is the net profit at the beginning of the relationship, mM

is the potential maximum of the profit, and k lnp2q{τ is the profit
growth rate specified in terms of the halfway time to maximum value.
The halfway time τ determines how quickly the profit approaches the
potential maximum – for every τ time unit, the difference between the
current value of the profit and the maximum decreases by half. The
net profit mt can then be inserted into expression 3.23 instead of the
constant R C value.
example 3.3
We continue here by considering a numerical example of LTV calcula-

tions. We assume the following parameters of the model:

• The net profit at the beginning of the relationship, m0 $100
• The potential maximum of the profit, mM $150
• The halfway time to the maximum profit value, τ 3 years
• Retention rate, r 0.9
• Discount rate, d 0.1

By substituting these parameters into equations 3.23 and 3.24, we
obtain the result shown in table 3.6. The nominal net profit in the first
column grows in accordance with equation 3.24 and passes the halfway
point of $125 after the third year. The expected net profit is a product
of the nominal net profit and total retention rate rt1 . Finally, the dis-
counted net profit is obtained by multiplying the expected net profit by
the discount multiplier p1 dqt1 . The LTV is the sum of the annual
discounted net profits.
Note that this analysis not only produces the total LTV but also
shows the LTV dynamics over time. From the partial sums of the dis-
counted net profit for one, two, and more years, we can draw an LTV
curve against time. If the curve saturates quickly, it means that most
of the value is extracted at the beginning of the relationship and long
relationships do not bring a lot of additional value. If the curve grows
steadily over a long time, it means that customer accounts remain prof-
itable in the long run.
N
Net Retention Expected Discount Discounted

Year
profit rate net profit multiplier net profit
1 $100.00 1.00 $100.00 1.00 $100.00

2 $110.31 0.90 $99.28 0.91 $90.26
3 $118.50 0.81 $95.99 0.83 $79.33
4 $125.00 0.73 $91.13 0.75 $68.46
5 $130.16 0.66 $85.40 0.68 $58.33
LTV $396.38
Table 3.6: Example of the LTV calculation for a horizon of five years.
The descriptive LTV model is similar to RFM analysis in the sense

that it merely extrapolates the average revenues observed in the past
into the future. It allows for some level of personalization if calcu-
lated for individual customer segments, but it does not predict how
customer properties and marketing actions can influence the lifetime
value.
3.5.7.2 Markov Chain Models

The descriptive LTV model does not provide us with much flexibility
when it comes to complex customer journeys with multiple states of
acquisition, maximization, and retention. At the same time, the pres-
ence of multiple states suggests that we can model a customer journey
as a random process or, more specifically, a Markov chain. The idea of
this approach is to define the set of customer states based on observed
customer properties, such as recency of purchase, estimate the proba-
bilities of transition between the different states and the corresponding
profits and losses, and then estimate the LTV based on the expected
customer path in the state graph [Pfeifer and Carraway, 2000].
A key part of the Markov chain approach is how the states and tran-
sitions are defined, so we choose to describe this method by using an
example. Consider the case of a retailer who determined from its data
that the recency of the last purchase is a good indicator for customer
churn – customers who made a purchase last month make a purchase
next month with a probability of p1 0.8, customers who made their
last purchase two months ago have a probability of p2 0.4 to pur-
chase again, three months corresponds to a probability of p3 0.1,
and, finally, customers who are inactive for four months are very un-
likely to return. These sets can be modeled by using a Markov chain, as
depicted in Figure 3.16. The chain has four states – one state for each
value of recency and one for defected customers. Customer who do

not make purchases move along the chain from left to right one step at
a time until they reach the defunct state. A purchase resets the process
and moves the customer back to the initial state.
Figure 3.16: Example of a Markov chain for LTV modeling. The white circles
correspond to three different values of recency. The black circle rep-
resents the defunct state.
This Markov chain corresponds to the following transition matrix:

p1 1 p1 0 0

p2 1 p2
P
0 0 (3.25)
1

p3 0 0 p3
0 0 0 1
Each row of the matrix corresponds to the current state and each
column corresponds to the next state. Each element of the matrix is the
probability of a customer moving from the current state to the next one.
The probability that a customer who is currently in state s will end up
in the state q after t months can then be calculated as the ps, qq element
of the matrix Pt , according to the standard properties of the Markov
chain. This gives us a simple way to estimate the customer journey in
probabilistic terms if the current state is known.
From an economic standpoint, each state of the chain corresponds to
profits and costs. For example, the marketing strategy may be to spend
some budget C on each active customer (e. g., send a printed catalog)
and stop doing so after the customer moves into the defunct state. The
first state is also associated with the revenue R of the purchase. Let
us introduce a column vector G so that the net profit of the i-th state
corresponds to its i-th element:

RC

G
C
(3.26)

C

0
It is easy to see that the matrix product PG produces a vector of

expected net profits for each state after one time period. For example,
the expected net profit for the first state will be
E rprofit | state 1s p1 pR Cq p1 p1 qC p1 R C (3.27)
because we are definitely spending the cost C and have p1 chance to

make a profit. Similarly, the expected profits for the second time period
are given by P2 G, and so on. Consequently, the LTV can be estimated
as a sum of such expected values over several time periods, and, if we
also adjust each period by the discount rate d, we get the following
expression:
Ţ
V
1
Pt G
p1 dqt1
(3.28)
t 1
with column vector V containing the LTV estimates for each initial
state. The LTV of a customer is also estimated as one of the elements
of this vector based on the current customer states, that is, the recency
value in this example. This result can be compared to the standard
descriptive LTV model in equation 3.23 – we essentially replace the
static net profit and retention rate parameters with a time-dependent
probabilistic estimate.
Let us conclude the example by estimating the LTV for several dif-
ferent values of time horizon T . As before, we assume the transition
probabilities of p1 0.8, p2 0.4, and p3 0.1. Let us also assume
that the expected revenue of one purchase is R $100, monthly cost
of marketing communications is C $5, and monthly discount rate
is d 0.001. By evaluating expression 3.28 for these parameters and
different values of the time horizon T , we get the following sequence
of LTV vectors:

$75.0 $135.5 $184.0

V T 1 $35.0 VT 2 $48.4 VT 3 $53.4 (3.29)
$9.5 $10.4 $10.5
$0.0 $0.0 $0.0
We can see that the LTV heavily depends on the initial customer
state. For a time horizon of three months, that is, VT 3 , the LTV of
a customer who made a purchase a month ago is $184.0. After two
months, the expected LTV drops to $53.4, and finally to $10.5 after the
third month.
We can extend the Markov chain method to accommodate more com-

plex customer states and marketing strategies. For example, the proba-
bility of the next purchase is often correlated not only with the recency
but also with the frequency of past purchases. In this case, each dis-
tinct pair of recency and frequency values can be modeled as a sep-
arate state in the chain. We can also replace the algebraic estimation
of the LTV given by expression 3.28 with Monte Carlo simulations of
the consumer journey, which can provide even more flexibility in how
gains and loses are modeled. In this case, we randomly select the ini-
tial state according to the state frequencies estimated from the data; we
then traverse the graph, flipping a coin in each state to decide in which
direction to go and recording the profits and losses encountered on the
way. By repeating this process many times, we obtain multiple samples
of the expected lifetime value. The advantage of this approach is that
the statistical properties of the LTV, such as mean, variance, or con-
fidence intervals, can be straightforwardly assessed by analyzing the
histogram of obtained samples. It also enables us to include additional
business logic and parameters in each state, which can be difficult to
encode with transition matrices.
3.5.7.3 Regression Models

The Markov chain model improves the descriptive LTV model by re-
placing the static retention rate and average expected profit with time-
dependent and state-dependent estimates. The limitation of this ap-
proach is that the number of states grows exponentially with the num-
ber of customer properties that we include in the model. We can take
one step back and note that, conceptually, both descriptive and Markov
chain models estimate the LTV in terms of the customer’s probability
to stay with the brand and the expected net profit from the customer.
This can be expressed as follows:
Ţ
LTVpuq ppu, tq mpu, tq (3.30)

t 1
in which ppu, tq is the probability of customer u staying until time

t and mpu, tq is the net profit from the customer in time period t. The
conventional descriptive model estimates both factors by using static
retention rate and average profit, whereas the Markov chain model es-
timates the same factors with probabilistic analysis. A more flexible
solution for problem 3.30 can be obtained by creating regression mod-
els for both factors. The advantage of this approach is that regression
models can use a wide range of independent variables created from a
customer profile and, thus, enable predictive and prescriptive capabili-
ties.
One can see that survival analysis is a natural choice for the reten-
tion probability factor in equation 3.30. This probability directly corre-
sponds to the customer’s survival function Su ptq, so the model can be
rewritten as
Ţ
LTVpuq Su ptq mpu, tq (3.31)

t 1
The survival model is trained to estimate the time to churn, and it

requires the churn event to be defined. These events may be tracked di-
rectly (if a customer explicitly unsubscribes from a service) or heuris-
tically by using some business rule (for example, all customers with
five or more months without purchase activity are considered lapsed).
Survival analysis efficiently solves the problem of the retention proba-
bility estimation by proper handling of censored data, and the ability
to estimate personalizes survival functions parametrized by customer
properties, such as recency and frequency of purchases.
The net profit values mpu, tq can be estimated in several different
ways with different accuracies. One of the most basic approximations
is to estimate the average net profit value for each customer segment
(persona) and use this static value for all customers within a segment.
More elaborate regression models can be created by including season-
ality and customer profile features.
3.6 designing and running campaigns
Targeting and LTV models, the basic building blocks of a targeting

system, provide a solid foundation for efficient marketing decisions.
A marketing campaign, however, is typically a flow with multiple ac-
tions and decisions geared to achieve a certain objective. This flow may
require multiple models to be wired together and optimizations to be
done with multiple signals and constraints taken into account. A target-
ing system often has some sort of repository for campaign templates,
where each template describes a certain flow of actions and decisions.
This flow is typically designed for a certain objective, but it can be
parametrized with different targeting models, budgeting constraints,
user experience properties, and so on. Forecasting of the campaign
3.6 designing and running campaigns 121
ROI and optimization of the execution parameters and thresholds to

balance the costs and profits is an important part of the campaign de-
sign process, and the corresponding routines and models can be con-
sidered as part of the template. In this section, we consider several
types of campaigns and their relationship to the previously discussed
basic models.
3.6.1 Customer Journeys
From an economic standpoint, interactions between a customer and

brand can often be viewed as a collection of transactions that can
be characterized in terms of total amounts, purchased items, margins,
clicks on the website, and so on. The problem of marketing optimiza-
tion can also be viewed in a transaction-centric way so that all compo-
nents of the marketing mix become focused on the optimization of in-
dividual transactions, in terms of their probabilities and margins. The
notion of the customer life cycle puts this optimization into a broader
context but still focuses on the brand’s concerns and objectives rather
than the customer experience. This approach is incomplete in many
marketing environments, including retail, where interactions with a
customer are experience-centric and the success of a brand is deter-
mined by its ability to deliver a superior long-term customer experi-
ence, rather than optimizing individual transactions.
One popular approach to customer experience analysis and model-
ing is based on the notion of customer journey maps. A journey map
tells the story of the relationship between a customer and the brand.
The map can describe the entire arc of the engagement, similarly to
the life-cycle curve, or can be focused on a specific scope, such a single
purchase. The map is typically visualized as a diagram with the steps
or stages of the engagement and the transitions between them. A very
simplified example of a customer journey map is shown in Figure 3.17.
This map depicts the flow of a single transaction but puts it in the con-
text of the customer experience and long-term interactions with the
brand.
The journey starts with triggers, which include browsing for a new
idea and product, preparation for special events such as birthday par-
ties, getting a promotional email, or the necessity to replenish a con-
sumable product. The trigger is followed by researching product in-
formation and selecting the purchasing channel. The interaction then
continues in the scope of the selected channel, including browsing of
a specific product and checkout, and completes with post-purchase ac-
tions such as product return or writing a customer review. In real life,
customer journey maps are typically much more complicated and in-
Figure 3.17: Example of a customer journey map.
clude a lot of details about the customer behavior and decision-making

process, distribution of customers by different states and branches, and
so on. In addition, customer journey maps are often created for each
customer segment because the journeys can differ substantially across
the segments.
A marketing campaign typically has a certain footprint on the jour-
ney map, in the sense that a campaign attempts to influence the path of
a customer. In Figure 3.17, for example, customers who drop out of the
offline purchasing funnel are provided with an offer to win them back.
Each campaign can be viewed as a type of template that can be applied
to a specific situation in the customer journey. A programmatic system
can have a repository of campaign templates, where each template in-
cludes the rules that prescribe when the campaign actions should be
triggered and how the situation should be handled, along with mod-
els to estimate the parameters of the required actions and forecast the
outcomes. A template can describe a single action or a whole set of ac-
tions that can be executed at different points in time, by using different
channels, and by taking the observed feedback into account.
Analysis of customer journeys and the creation of journey maps is

usually a strategic project that often includes extensive analytical re-
search, customer surveys, and marketing strategy development. Thus,
the creation of customer journeys maps and campaign templates is not
the responsibility of a programmatic system. We generally assume that
these artifacts are created elsewhere and then entered into the system.
The responsibilities of the system, however, include estimation and op-
timization of the template parameters and dynamic selection of the
most optimal templates.
3.6.2 Product Promotion Campaigns
One of the most basic types of targeted campaign is a sales promotion

for an individual product. Examples of such promotions include ad-
vertisements without monetary value, dollar-off coupons, buy-one–get-
one (BOGO) coupons, and free product samples. In the CPG domain,
this class of campaign is often referred to as free-standing inserts (FSI),
named so because of coupon booklets that are inserted into local news-
papers. In the most basic form, a standalone promotion corresponds
to a simple customer journey with a trigger (promotion) and purchase
(redemption). As we will see later, this approach is not necessarily the
most efficient one, but it is applicable to all objectives:
• For acquisition, a brand can send BOGO or dollar-off coupons
to customers who are heavy category buyers but do not buy this
particular brand.
• For maximization, a brand can send conditional promotions like

Buy 3, Get $1 Off to existing customers.
• For retention, a brand can send BOGO or dollar-off promotions

to customers who have decreased their consumption relative to
previous purchasing cycles.
The response modeling framework provides some guidance on how
such promotions can be targeted by using predictive models that esti-
mate the probability of response, but there are many additional aspects
that need to be covered, including the targeting process, budgeting
rules, and selection of promotion parameters.
3.6.2.1 Targeting Process

A targeting system can be used in both batch and real-time modes,
depending on the environment and campaign nature. Some promo-
tions can be distributed by sending millions of emails in one shot, so
a targeting system can prepare a list of customers to be targeted in

advance. Other promotions are targeted in near real time because of
rapidly changing customer profiles or context. For example, a promo-
tion may or may not be offered to customers depending on the content
of their shopping baskets just before checkout. The real-time approach
is generally more flexible, and a properly designed real-time targeting
system can also simulate the batch mode by evaluating the targeting
rules and models for the entire customer database. Consequently, we
focus on the real-time targeting case and consider a process that re-
ceives a single customer profile and the corresponding context as the
input and produces a list of promotions that should be offered to this
customer.
We also assume that the system has a database of available promo-
tions that can potentially be offered. This database includes promotions
from all campaigns that are currently active. Each promotion needs
to be attributed with properties such as business objective, promoted
product, and category, so the targeting system can use this informa-
tion to link the promotion with the proper targeting models and rules.
This is the reason why promotion creation and targeting are closely
related to each other, in the sense that each targeting step or feature
requires a counterpart in the campaign configuration and promotion
attributes. We will go though the targeting process and discuss both
how promotions are selected from the set of available options and the
methodology for creating and attributing promotions with the proper-
ties and conditions needed for targeting.
The targeting process can be thought of as a sequence of three steps.
First, the system takes all available promotions and selects those that
are valid for a given context and customer. Next, promotions are scored
to produce a list sorted according to the fit to the objective. Finally,
the optimal set to be offered to the customer is selected by applying
budgeting limits and other constraints. This process is sketched in Fig-
ure 3.18. The initial filtering of promotions is typically based on busi-
ness rules and conditions, so we refer to it as hard targeting. On the
other hand, promotion scoring typically uses predictive models that
produce a continuous score, so we call this stage soft targeting.
The goal of the hard-targeting stage is to select promotions that qual-
ify for a given context. Promotions created in the targeting system are
typically associated with conditions that must be met by a given con-
text to activate the promotion. The purpose of these conditions is to
encourage certain consumer behaviors and ensure the basic economic
goals of the promotion. The hard-targeting conditions essentially de-
fine the campaign template, that is, the point in the customer journey
Figure 3.18: The promotion targeting process.
where the promotion should be applied. Consider the following typical

examples:
• Quantity condition. Activates a promotion when the customer
purchases a certain quantity of a certain product, brand, or cate-
gory in a single transition or over a certain period of time. This
condition is often used in maximization campaigns to stretch a
consumer, that is, to give an incentive to buy more then usual.
For example, a customer who typically buys two packs of yogurt
can be offered a Buy 4, Get 1 Free promotion.
• Non-buyer condition. Activates a promotion for customers who

have not bought a product or brand for a certain period of
time. This condition can be used in retention and acquisition
campaigns to separate active brand customers from inactive and
prospecting consumers.
• Channel condition. Activates a promotion when a customer inter-

acts with a brand or retailer via a certain channel. For example, a
customer can be rewarded for visiting a store three times a week.
• Retargeting condition. Activates a promotion based on previ-

ously offered or redeemed promotions. For example, customers
who have received but have not redeemed promotions via digital
channels can be contacted by using in-store channels.
• Location condition. Activates a promotion based on the customer

location as determined from mobile device data, store location,
in-store beacons, or IP address.
• Availability condition. Some promotions can be temporarily de-

activated if the corresponding products are out of stock or un-
available through a given marketing channel.
The hard-targeting step produces a set of promotions that can poten-
tially be offered to a consumer. The goal of the soft-targeting stage is
to select the most relevant offers and filter out options that are likely to
be inefficient. Soft targeting is often done by using propensity models.
A targeting system can maintain a repository of models where each
model is trained for a certain business objective and product category
and is attributed accordingly. As each promotion is also attributed with
similar properties, the system can dynamically link models to promo-
tions. Scoring models can be combined with special conditions that
complement the logic encapsulated in the model. For instance, the ba-
sic look-alike acquisition model identifies customers who are similar
to natural triers, but it does not ensure that a promotion will not be of-
fered to those who already buy the product. In contrast, maximization
and retention promotions typically should not be offered to customers
who do not consume the promoted product. These additional checks
can be implemented as a condition.
3.6.2.2 Budgeting and Capping

Once the set of candidate promotions is prepared and ranked, the sys-
tem needs to select the final set of promotions that can be offered to
the customer. This step can include several controls to manage differ-
ent aspects of a campaign. First, the number of promotions received
by a customer in the scope of a single campaign, as well as the total
frequency of communications with the customer (number of messages
per time unit), should be limited. These rules, often referred to as pres-
sure rules or frequency capping rules, typically use thresholds selected
heuristically or by means of experimentation. Next, the campaign bud-
get or the total number of issued promotions is typically limited. The
targeting system, however, often needs to determine the optimal num-
ber of promotions to maximize the campaign ROI. It can be the case
that this number is far below the limit specified by the marketer and
consumption of the budget up to the limit can make losses. From the
propensity modeling standpoint, the profitability optimization prob-
lem can be viewed as finding the propensity scoring threshold that
maximizes the profit if all customers with a higher score are targeted
and all other customers are not. We have already shown how the trade-
off between campaign costs and profits can be modeled by using the
response modeling framework, and we now consider an example that
provides more practical details.
example 3.4
Consider the case of a retailer who has 100,000 loyalty card holders.
The retailer plans a targeted campaign where each promotion instance

costs $1 and the potential profit of one response is $40. The average re-
sponse rate for this type of campaign and product category estimated
from historical data is 2%. On the basis that we have created a propen-
sity model that estimates the response probability for each customer,
we can score all card holders and sort them by the scores. The result
can be summarized by splitting the customers into “buckets” of equal
size where the first bucket corresponds to the customers with the high-
est scores and the last bucket corresponds to those with the lowest
scores. The targeting problem can then be defined as finding the opti-
mal number of top buckets to include in the targeting list, or, equiva-
lently, finding the threshold score that separates these top bucket from
the bottom ones. We use bucketing for the sake of convenience in this
example and this approach is often used in practice as well, but there is
nothing to stop us from doing the same calculations for individual cus-
tomers, that is, having as many buckets as customers. Let us assume
that we have 10 buckets or deciles, so that each bucket contains 10,000
customers; consequently, the average expected number of responders is
200 per bucket. In other words, we are likely to get 200 responses from
each bucket if we randomly assign customers to buckets. This number
is shown in the second column of table 3.7, and the third column con-
tains the cumulative number of responders, which reaches 2,000 or 2%
of the customer base in the bottom row.
Decile Responses, Random Responses, Targeted Lift

Bucket Total Pr Bucket Total
1 200 200 0.060 600 600 3.00
2 200 400 0.057 570 1,170 2.93
3 200 600 0.038 380 1,550 2.58
4 200 800 0.017 170 1,720 2.15
5 200 1,000 0.010 100 1,820 1.82
6 200 1,200 0.007 70 1,890 1.58
7 200 1,400 0.006 60 1,950 1.39
8 200 1,600 0.003 30 1,980 1.24
9 200 1,800 0.001 10 1,990 1.11
10 200 2,000 0.001 10 2,000 1.00
Table 3.7: Example of campaign lift calculations.
Next, let us assume that the lowest response probability scores gener-
ated by the propensity model in each bucket are those presented in the
fourth column. By multiplying the bucket size by this probability, we
get the expected number of responses in the case of the targeted distri-
bution presented in the next two columns. The total number of respon-
ders still adds up to 2,000, of course. The ratio between the number of
responders in the case of targeted and random distributions is called
lift, and it is the key metric that describes the quality of the targeting
model. The lift is typically visualized by using a lift chart similar to the
one in Figure 3.19. This chart shows two lines that correspond to the
cumulative number of responses: the straight line corresponds to the
random distribution and the raised curve to the targeted distribution.
Figure 3.19: Lift chart for the targeting model.
To determine the number of buckets to be targeted, we need to esti-

mate the campaign ROI. Each promotion costs $1, so the random dis-
tribution strategy is not profitable because each bucket causes a loss:
$40 response profit 10,000 recipients 0.02 response rate

$1 per customer 10,000 recipients
$2,000
The targeted campaign, however, will be profitable for the first three
buckets because of the high response rates, as summarized in table 3.8.
One can see that including more buckets initially increases the cam-
paign ROI but it then starts to decrease and eventually becomes nega-
tive.
The campaign ROI is maximized for outreach to three buckets, that

is, the top 30% of the population. This corresponds to all customers
with a propensity score above 0.038. The targeted campaign ROI is
plotted in Figure 3.20. Note that the maximum possible budget, which
Decile Cost Profit Targeted

Random Targeted ROI
1 10,000 -2,000 14,000 14,000
2 10,000 -2,000 12,800 26,800
3 10,000 -2,000 5,200 32,000
4 10,000 -2,000 -3,200 28,800
5 10,000 -2,000 -6,000 22,800
6 10,000 -2,000 -7,200 15,600
7 10,000 -2,000 -7,600 8,000
8 10,000 -2,000 -8,800 -800
9 10,000 -2,000 -9,600 -10,400
10 10,000 -2,000 -9,600 -20,000
Table 3.8: Example of campaign profitability calculations.
corresponds to sending a promotion to each and every customer, does

not maximize the ROI. On the contrary, it causes a loss of $20,000.
Figure 3.20: Targeted campaign ROI as a function of the outreach.
It is important to note that we used the basic response probability,

instead of uplift modeling, in this example. In practice, this can result
in poor campaign performance because high response rates do not
guarantee uplift in customer spending or consumption. In other words,
a control group in each bucket can perform equally well or even better
than the targeted group in the same bucket. One can work around this
problem by replacing the response probabilities in table 3.7 with the
uplift scores discussed in Section 3.5.4.2.
N
The ROI maximization principle allows estimation of the optimal

baseline parameters of a campaign, such as the total number of promo-
tions to be distributed and the scoring threshold. In the real world, it
can sometimes be beneficial to deviate from the baseline, especially for
real-time applications when the set of customers who will actually in-
teract with the system is not known in advance. Consider the following
scenario. A system runs a promotional campaign with a fixed budget
and spreads this budget evenly over the campaign’s time frame. This
suggests that we should use some fixed distribution rate, for example,
100 promotions per hour. However, what should we do if the campaign
is running over this target rate (in our example, 100 promotions have
already been issued during the last hour) but we encounter a consumer
with a very high propensity score? It can be reasonable to go over the
budget at this point and then slightly decrease the rate later to get back
on track. This behavior can be implemented by dynamic adjustment of
the scoring thresholds depending on the deviation from the target dis-
tribution rate. This idea is illustrated in Figure 3.21. We define the tar-
get distribution rate and and two margins, ε and ε , that determine
the maximum acceptable deviation from the target line. Note that the
target does not necessarily have to be a straight line, and one can use
a more sophisticated curve that takes into account weekends, working
hours, and so on. The actual distribution rate is constantly measured
and controlled by the system to stay within the margins.
Figure 3.21: Dynamic scoring threshold for budget control.
The scoring threshold can then be expressed as a function of the de-

viation from the target line at the current moment of time t0 . If we are
substantially under budget (under the ε line), the scoring threshold
can be set to the minimum, which corresponds to the lowest affinity
L0 between the consumer and promotion that is sufficient to make the
offer. If we are substantially over budget (above the ε line), then the
threshold should be set to the maximum possible scoring value to stop
the distribution completely. These two extreme points can be connected
by some growing function, as illustrated in Figure 3.21. Consequently,
we become more and more demanding of consumers as we approach
and cross our budgeting limits, and we lower the bar when we do not
encounter enough high-quality prospects.
3.6.3 Multistage Promotion Campaigns
Standalone promotion campaigns, such as the distribution of trial or

maximization coupons, are widely used in practice. However, it can
be argued that this strategy can be inefficient because it has a very
short and limited impact on the customer journey [Catalina Marketing,
2014]. It is sometimes possible to design more sophisticated campaigns
with multiple phases that influence the customer journey over a longer
period of time. Let us consider an example of a CPG maximization
campaign with the following design:
• The first phase of the campaign is an announcement that aims to
inform the customers about the offer. For example, a brand can
distribute the following message via available marketing chan-
nels: Buy Q or more units of product X and save on your next shopping
trip. The more you buy, the more money you save.
• The second phase is distribution. A targeting system tracks the

transactions and issues dollar-off coupons to customers who
qualify for the targeting condition, that is, the purchase of Q or
more units in this example. The discount amount of the coupon
is determined dynamically based on the purchased quantity –
as announced, the more consumers buy, the more they save. At
this stage, a consumer is incentivized to buy more units to get a
coupon as a reward.
• The third and the last phase is redemption. On the second shop-
ping trip, the consumer buys the promoted product to redeem
the coupon issued in the previous stage. The consumer is incen-
tivized to buy a product to redeem a coupon and get a discount.
This campaign template can be thought of as a customer journey
with three steps: trigger, purchase, and redemption. It can be argued
that this approach is more efficient than standalone promotions be-
cause it has a more durable impact on customer loyalty and lower
costs per unit moved [Catalina Marketing, 2014]. The dynamically de-
termined discount value in the second stage is an interesting detail
because the targeting system needs to optimize this value and fore-
cast how it will influence the campaign outcomes. This aspect is not
addressed by the targeting and budgeting processes discussed in the
previous sections. Let us consider an example that demonstrates how a
targeting system can heuristically evaluate different promotion param-
eters and forecast the campaign outcomes by using just basic statistics.
More formal discount optimization methods will be discussed later in
Chapter 6 in the context of price optimization.
example 3.5
Consider the case of a promotional campaign that follows the three-

phase scenario described above. The goal of the targeting system is
to choose a reasonable value for the quantity threshold Q and, in the
second phase, to determine the discount amount based on the quantity
that is actually purchased. We can start with a histogram of purchase
quantities for the promoted product calculated for the time interval
equal to the duration of the campaign. Let us denote the number of
transactions with exactly q units of the promoted product purchased
as Hpqq, so the historical histogram is as follows:
Hp1q 4000 (32%) Hp4q 1000 (8%)

Hp2q 5000 (40%) Hp5q 600 (5%)
Hp3q 2000 (16%) Hp6q 0
We want to stretch customers who buy relatively small quantities,

so the system can select threshold Q to be above the majority of trans-
actions. In this example, the value of 3 would be a reasonable choice
because 72% of transactions are under this threshold. Consequently,
the system will offer a discount coupon to customers who buy more
than 3 units. The discount value depends on the quantity actually pur-
chased. We will discuss this in more detail in Chapter 6, but we can
assume for now that the discount values are static configuration pa-
rameters. For example, suppose that the minimal discount value is 15%
and it increases by 5% at each level. This means that customers who
buy 3 units get the discount of 15%, those who buy 4 units get 20%,
and those who buy 5 units get 25%. Let us denote the number of units
at level i as qi and the corresponding discount value as di . Once all
of these parameters are determined, the system can forecast the cam-
paign outcomes. This can be done separately for each discount level.
The expected number of coupons generated at level i can be estimated,
based on the previously created histogram, as
couponspiq Hpqi q (3.32)

The expected number of redemptions can then be estimated by using

a response model that includes the discount depth as a feature:
redemptionspiq couponspiq rpdi q (3.33)
in which rpdi q is the average response rate predicted by the model. The
cost of coupons at level i can then be estimated as
costpiq pproduct price di qi cq redemptionspiq (3.34)
in which c stands for the additional costs associated with a coupon,

such as distribution and clearing-house costs. The campaign efficiency
can be predicted as the ratio between the total number of redemptions
and the total costs summed over all levels (cost per redemption).
N
3.6.4 Retention Campaigns
Retention campaigns aim to save customers who are likely to leave.

This type of campaign is widely used in telecommunications, insur-
ance, banking, and other subscription-based domains where the conti-
nuity of a relationship is critical. The problem of customer churn, how-
ever, is relevant for most non-subscription businesses as well, including
retail. One of the key reasons why retention activities are important is
that acquisition of new customers can be much more challenging and
expensive than the retention of existing ones. According to some stud-
ies, the acquisition cost per consumer can be 10–20 times higher than
the retention or reactivation cost because of lower response rates and
other factors [Artun and Levin, 2015].
A retention campaign can be defined as a follow-up with customers
who are at risk. The follow-up and risk, however, can be defined very
differently depending on the campaign design. Examples of risk in-
clude the risk of subscription cancellation and the risk of switching
to a different supermarket chain. The definition of risk depends on the
business model, nature of the product or service, and usage patterns. A
software service provider, for example, can be concerned about the risk
of subscription cancellation but may also observe a significant number
of customers who create an account but do not download the client
application. Thus, the risk of not downloading the application can be
recognized and addressed by a dedicated retention campaign. Exam-
ples of follow-up actions include reminder emails, distribution of edu-
cational materials, requests to review the recently purchased product,
and special offers and discounts.
Compared to promotional campaigns, the design of retention cam-
paigns puts more emphasis on the lifetime value and uplift. The in-
corporation of lifetime value projections is important because invest-

ments in the retention of customers with low value would be mean-
ingless. Uplift modeling is important because targeting of the wrong
customers can be counterproductive for several reasons [Radcliffe and
Simpson, 2007]. First, many customers at risk are dissatisfied already,
and additional communications, especially intrusive ones like phone
calls, can catalyze the churn process. Second, some retention commu-
nications can remind customers that they have an opportunity to leave,
which makes them reconsider their relationship with the brand and
look around for alternative options. This makes it important for the
communications to be focused and the outcomes to be constantly mea-
sured by using control groups.
Retention campaigns are typically assembled by using the standard
building blocks, but there exist different design methods. One of the
most basic approaches is to target based on the propensity to churn.
This model can be created by using standard propensity modeling
methods with a training data set assembled from active and churned
customer profiles. This approach is seemingly simple, but it has pit-
falls that should be discussed. As we have previously discussed, a
marketing action can be described in terms of two conditional prob-
abilities – the probability to respond if treated and the probability to
respond without treatment. In the case of retention, the response event
corresponds to churn. All customers can be categorized with respect
to these two probabilities, as depicted in Figure 3.22.
Figure 3.22: Categorization of customers from the perspective of a retention

campaign.
If the overall retention strategy is focused, that is, the customers are
not normally treated with retention offers, a propensity model trained
to predict churn outcomes will actually predict the probability of churn

given no treatment as
scorepxq Prpchurn | N, xq (3.35)
in which x is the customer profile feature vector. A retention campaign

driven by this probability focuses on the rightmost vertical slice of
the square in Figure 3.22, which includes many Persuadables but also
many Lost Causes. If the retention strategy is broad, that is, almost all
customers are treated to some degree, the model will actually estimate
the propensity to churn under treatment as
scorepxq Prpchurn | T , xq (3.36)
which corresponds to the horizontal area at the top of the square in Fig-
ure 3.22, which contains many Do-no-disturbs and Lost Causes. This
aspect of modeling should be taken into account when the population
for model training is selected. The retention campaign can also use sur-
vival analysis to estimate the time-to-churn, which can be more conve-
nient for choosing the right moment for treatment than the probability
of churn.
Targeting customers based on their probability to churn does not

take into account the long-term outcomes of the campaign. These out-
comes can be quantified in terms of lifetime value because every reten-
tion saves the LTV of the corresponding customer and every churn is a
loss of this LTV. If the probability to churn and LTV are estimated for
a given customer, the product of these two values is the expected loss.
We can expect that the ratio between the saved revenues and the cam-
paign costs is maximized by treating those customers with the highest
expected loss, so this measure can be used as a targeting score:
scorepxq Prpchurn | N, xq LTVpxq (3.37)
The LTV can be estimated based on the average customer spend or

by using the more advanced LTV models described earlier. This model
can be customized depending on the business model to account for
the costs and profits associated with the different possible outcomes.
For example, the expected retention gains, churn losses, and campaign
costs can be separately estimated. The expected loss model specified by
equation 3.37 is widely used in practice because of its simplicity and
reasonably good efficiency.
The main shortcoming of the expected loss model is that it uses only
the probability to churn and not the churn uplift, that is, the difference
between the treated and non-treated churn probabilities:
upliftpxq Prpchurn | T , xq Prpchurn | N, xq (3.38)
A positive churn uplift means that the treatment amplifies the churn,
that is, the treatment has a negative effect. High uplift corresponds to
the upper left corner of the square in Figure 3.22. A negative churn up-
lift means that the treatment decreases the churn, which corresponds
to the lower right corner in Figure 3.22. Consequently, we want to tar-
get customers by using the inverse of the uplift as a score:
scorepxq upliftpxq savabilitypxq (3.39)
This metric is also called savability because it estimates the propen-

sity to react positively to the retention activity. Uplift/savability can
be modeled by using the methods described in Section 3.5.4.2, includ-
ing the two-model and single-model approaches. Similarly to other
applications of uplift modeling, the savability-based approach helps to
separate those customers who are likely to stay only if they are treated
from other groups, thereby increasing the efficiency of the retention
campaign. However, it is important to keep in mind that this approach
also inherits the typical disadvantages of uplift modeling. It includes
more complicated modeling and higher variance of estimates because
the uplift is the difference of two random variables. The uplift can also
be combined with the expected loss technique to take into account the
long-term impact on the harvested LTVs:
scorepxq savabilitypxq LTVpxq (3.40)
Once the targeted scores are calculated, the optimal targeting depth,
that is, the percentage of the population to be targeted, can be deter-
mined by using the ROI maximization method described earlier in Sec-
tion 3.6.2.2. The campaign can then be executed with the same target-
ing process as that used for product promotional campaigns.
3.6.5 Replenishment Campaigns
Retention campaigns are most relevant for subscription-based busi-

nesses, such as telecommunication services, insurance, software, and
banking. In the retail domain, the subscription-based model is less
frequently used, but many products are routinely replenished so the
engagement model becomes similar to a subscription. Examples of re-
plenishable products are numerous: food, cosmetics, office supplies,
3.7 resource allocation 137
accessories like water filters, and many others. Replenishment cam-

paigns aim to drive repeat purchases and decrease purchasing cycles by
sending reminders, recommendations, and specialized promotions.
From the campaign design standpoint, the distinctive features of re-
plenishment are the emphasis on communication timing and purchas-
ing habits. The communication timing is important because the replen-
ishment notifications should be aligned with the individual purchasing
cycles – it is not efficient, for example, to send a notification immedi-
ately after a customer has purchased the product. Connection with the
purchasing habits is also important because the notification message
should be consistent with the products and categories typically pur-
chased by the recipient.
Let us start with a very basic approach that can be implemented by
a targeting system. First, the system estimates the average duration of
a purchasing cycle for each replenishable product or product category.
The replenishment campaign is then executed repeatedly, for exam-
ple, on a daily basis. For each execution, the system goes through the
customer profiles and determines the last purchasing date for replen-
ishable products. This date is compared with the estimated duration
of the purchasing cycle, and a notification is sent to customers who
are apparently approaching the end of the cycle. The message can be
personalized based on the most recently or most frequently purchased
products found in the purchasing history.
One of the main limitations of the above approach is an estimate of
replenishment cycles that is too rough. One possible improvement is
to break down the estimate not only by product category but also by
customer segment or persona, to account for the differences between
customers. In other words, the cycle duration is estimated for each pair
of a category and a persona. More accurate results can be obtained by
using survival analysis to estimate the time-to-purchase, as we have
already discussed. The survival model also allows the determination
of the factors that positively or negatively contribute to the time-to-
purchase, such as discounts or replenishment notifications themselves,
so the message content and frequency can be adjusted accordingly.
3.7 resource allocation
The problem of optimal targeting can be viewed as a resource allo-

cation problem where some limited resource, such as sales coupons,
needs to be allocated across customers. So far, we have focused only on
this type of allocation and have ignored the fact that marketing activi-
ties may be required to make many other allocation decisions. The cor-
porate marketing resource allocation strategy generally includes alloca-
tion between marketing and non-marketing activities, products, prod-

uct life-cycle stages, markets and territories, business objectives, mar-
keting channels, and communication types [Carpenter and Shankar,
2013]. Many of these allocation decisions, such as allocation between
marketing and research activities, are very strategic and thus cannot be
addressed by a programmatic system. Some other decisions are more
tactical, and the system can include modules that automate or at least
facilitate the decision-making process. It should be kept in mind, how-
ever, that targeting is one of the most tactical and technical allocation
problems, and the automation of other allocation decisions is increas-
ingly more complex and challenging.
The modeling and optimization of how resources are allocated
across marketing activities and capabilities is known as marketing mix
modeling (MMM). It can be viewed as the statistical analysis of how
the different components of the marketing mix, such as promotions
and prices, impact business performance metrics, such as sales and
revenues. In this section, we focus on two resource allocation problems,
allocation by channel and allocation by business objective, and discuss
how these problems can be addressed by using MMM methods.
3.7.1 Allocation by Channel
A targeting system often has multiple marketing channels at its dis-

posal, and each channel has its own costs structure, audience, and
efficiency. The direct mail channel, for example, may have a much
higher cost per message than the email channel, but it may provide
higher response rates for certain categories of customers. This requires
marketing communications to be optimized with respect to channels.
One possible approach to this problem is optimization at a customer
level, where the channel is selected by using a response model that
accounts for channel-specific response probabilities and costs. Another
approach is to optimize the global budget allocation across channels
to maximize the revenue. This is sometimes referred to as channel mix
modeling. These two methodologies can be viewed as bottom-up and
top-down solutions, respectively, and both are important.
Channel mix modeling is a set of statistical analysis methods that
focuses on the following descriptive and predictive questions:
• What percentage of revenue (or other performance metric) is
driven by each channel or communication type?
• How will an increase or decrease in a given channel spend affect

the revenue?
• What is the optimal budget allocation across the channels?

Intuitively, we can expect these questions to be answered by a re-

gression model that expresses the metric of interest as a function of the
channel activity. The challenge is that the dependency between the ac-
tivity and the observed metric can be complicated for several reasons.
First, we can directly measure the channel activity only as the current
number of emails or online advertising impressions, but the customer
responses are typically delayed and spread over time. Second, multi-
ple campaigns can overlap when they run in parallel, but we can only
observe the cumulative effect. Finally, the dependency between the in-
tensity of the channel activity and the magnitude of the response is
often nonlinear because of saturation effects. One popular channel mix
model that accounts for these effects is the adstock model [Broadbent,
1979].
The key assumption made by the adstock model is that each given
sales period retains a fraction of the previous stock of advertising. By
assuming, for now, that we have only one advertising channel, let us
denote the intensity of the channel activity measured in dollars spent
or the number of messages in time period t as xt , the business metric
of interest, often the sales volume or revenue, as yt , and the current
effect induced by the activity on the business metric as at . The effect
variable at is called the adstock. The adstock model assumption can
then be expressed as
at xt λ at1 (3.41)
in which λ is the decay parameter that corresponds to the fraction of
the effect carried over the time period. For example, a parameter of
0.4 means that the treatment from one period ago has 40 percent of
its effect during the current period. In other words, the adstock model
assumes that each new marketing activity increases loyalty and aware-
ness to the new level, but loyalty gradually fades until it is boosted
again by the next portion of activity. By expanding recursive equa-
tion 3.41, we get
at xt λxt1 λ2 xt2 λ3 xt3 ... (3.42)
Note that this is essentially a smoothing filter applied to the input
sequence. In practice, we can always assume that the treatment effect
is finite and limited by n periods, so we rewrite the adstock transfor-
mation of the original sequence as
ņ
at xt λj xtj (3.43)

j 1
The observed business metric is then estimated as a linear function
of the adstock:
pt
y wat c (3.44)
in which w is the weight and c is the baseline value given no adstock.

In the case of multiple channels, we assume that the adstock is additive,
so the full model specification is a linear regression over the adstocks:
ņ
pt
y wi ait c (3.45)

i 1
whereby each channel is modeled with its own decay parameter λi ,

so the full model requires estimation of the baseline parameter c, n
decay parameters λi , and n channel weights wi . We can fit the model
by solving the following problem for the observed samples yt :
¸
min
c, w, λ
yt ypt 2 (3.46)
t
The fitted model allows us to estimate the impact of increasing or de-

creasing the channel budgets and to measure the relative contribution
of each channel to the target metric as:
zit °wwi aita (3.47)

j j jt
This value can be averaged over time to obtain the average relative
channel contribution. The efficiency of the channel can be measured as
the ratio between the absolute channel contribution and the channel
budget, that is, the number of units sold generated by each dollar spent
on marketing activities through this channel. The following example
illustrates how the adstock model can be created and used.
example 3.6
Consider a retailer who uses two marketing channels: email and SMS.
The retailer can measure and control the intensity of marketing com-
munications through each of the channels by setting budgeting and
capping rules. The retailer also observes the sales volume. A data sam-
ple with these metrics is plotted in Figure 3.23 for 20 sequential time
intervals (we have omitted the table with numerical values for the sake
of space).
The adstock model can be fitted by solving problem 3.46 with numer-
ical optimization methods. By setting the length of the decay window
n to 3, we get the following estimates for the model parameters:
baseline: c 28.028
email: λemail 0.790 wemail 1.863
SMS: λsms 0.482 wsms 4.884
Figure 3.23: Data for adstock modeling: sales volume, email activity, and SMS
activity.
We can use these parameters to calculate the adstock for each of

the channels according to expression 3.43. The structure of the sales
volume can then be visualized as three layers stacked one on top of the
other: the baseline sales volume determined by constant c, the email
contribution estimated as the email adstock scaled by wemail , and the
SMS contribution estimated as the SMS adstock weighted by wsms . This
structure corresponds to expression 3.45, and the result is plotted in
Figure 3.24. This decomposition enables us to estimate the efficiency of
each channel and optimize budget allocation.
N
The basic adstock model accounts for overlapping marketing activ-

ities and decay effects but not for the advertising saturation that we
mentioned earlier. In general, an increase in the treatment intensity
increases the outreach of the campaign, which increases the demand.
The dependency between the intensity and demand, however, is not lin-
ear. It typically follows the law of diminishing returns, so that spending
more dollars on the marketing activity at some point yields a lower in-
cremental demand. The adstock model can account for this saturation
effect by nonlinear transformation of the intensity variable. One pos-
Figure 3.24: Decomposition of the sales volume into the layers contributed by
different marketing channels.
sible choice is to use the sigmoid (logistic) function, so that recursive

adstock equation 3.41 is redefined as follows:
at 1 1
exppµ xt q
λ at1 (3.48)
in which µ is an additional model parameter that controls the steepness

of the logistic curve. The adstock model can be extended or modified
in many ways to account for additional effects that can be found in
practice. For example, we typically need to account for the seasonality
of demand, which can be done by extending the model with additional
variables. At this point, channel mix modeling can take advantage of
the demand modeling techniques discussed in detail in Chapter 6. An-
other example is that the geometric lag assumption made by the ad-
stock model is somewhat restrictive because the time lag can have a
more complex shape. In fact, the model described by equations 3.42
and 3.43 is known in econometrics as the Koyck distributed lag model,
which, in turn, belongs to the family of distributed lag models [Koyck,
1954]. This family provides a number of more flexible alternatives, in-
cluding a polynomial distributed lag model, which can be both more
flexible and easier to estimate than the Koyck model [Almon, 1965;
Hall, 1967].
3.7.2 Allocation by Objective
A programmatic system can use both LTV growth and acquire–

maximize–retain goals as the input objectives for the targeting
optimization. The ROI can be estimated in all of these cases by using

the LTV uplift or immediate net profit uplift, in accordance with the
response modeling framework. The choice between these objectives
and global optimization of the ROI is a strategic question that does
not necessarily need to be addressed by a programmatic system.
However, the system can provide some guidance on how to allocate
budgets across the objectives to maximize the overall ROI [Blattberg
and Deighton, 1996].
We have discussed in Section 3.5.7 that the retention rate is a major
factor that influences the LTV, so the LTV can be considered a function
of the retention rate. We can make the assumption that the retention
rate is, in turn, a function of the marketing budget spent on retention
activities. For example, we can model the dependency between the
budget and rate as follows:

r rmax 1 ekr R (3.49)
in which R is the retention budget per customer, rmax is an estimate

of the maximum retention rate (ceiling) that we can achieve given the
unlimited budget, and kr is a coefficient that determines how fast the
rate approaches the ceiling. Similarly, we can model the acquisition
rate a (response rate for an acquisition campaign) as a function of the
acquisition budget:

a amax 1 eka A (3.50)
in which A is the acquisition budget per customer, amax is an estimate

of the maximum response rate, and ka is a parameter that controls the
sensitivity of the rate to budget changes. Consequently, the acquisition
net profit for a given customer can be defined as
a LTVprq c (3.51)
in which c is the acquisition cost per prospect. The overall optimization
problem for budgets A and R can then be defined as follows:
max Np pa LTVprq cq Nc LTVprq
A, R
(3.52)
subject to A R ¤ total budget
in which Np is the total number of available prospects and Nc is the
total number of current customers. The first term of equation 3.52 cor-
responds to the revenues from new customers and the second term cor-
responds to the revenues from existing customers, so it is effectively a
revenue optimization problem. Equation 3.52 defines the optimization
problem in terms of aggregated and averaged values, but it can be eas-
ily rewritten as a sum of individual LTVs over all customers to enable
more accurate estimations with predictive models.
3.8 online advertisements
The principles of promotion targeting discussed in the previous sec-

tions are geared towards consumer packaged goods and the traditional
retail environment. It is clear that many of these principles hold for
other marketing environments, but implementation heavily depends
on the available data and exact definition of business objectives, which
can vary across environments. We continue here by analyzing online
advertising, which is perhaps the most important and well-developed
application of algorithmic marketing and is an excellent example of
an environment where the technical infrastructure and data flows are
so complicated that the business objectives cannot be understood and
achieved without careful examination of the technical capabilities and
limitations.
3.8.1 Environment
The online advertising environment is very complex and diverse be-

cause it represents a marketplace where thousands of companies sell
and buy ad inventory, offer and utilize technical systems that automate
the buying process, and control and measure the quality and effective-
ness of advertising campaigns. Additional complexity comes from the
fact that, although most of the terminology and standard offers are
generally accepted across the industry, there are many variations and
cross-cutting solutions that appear as the industry evolves. The high
complexity of the online advertising ecosystem makes it difficult to
capture all important aspects of the environment in a single view, so
we will start with a simplified model, shown in Figure 3.25, to support
our discussion of business objectives and economic goals.
Figure 3.25 depicts the relationships between the following key enti-
ties that constitute the online advertising landscape:
• A brand, also commonly referred to as a marketer, is a seller of
products or services. The brand invests money into advertising
campaigns and expects to obtain a return on the investments by
improving certain aspects of sales and customer relationships.
• An advertiser or agency runs advertising campaigns on behalf

of the brand. The advertiser generally tries to achieve the same
goals as the brand, but its exact strategy depends on the pay-
ment model established between the brand and the advertiser, as
well as the methodology used to measure the performance of the
campaign. The brand can work with multiple agencies that may
compete against each other in the scope of one campaign.
3.8 online advertisements 145
Figure 3.25: Online advertising environment.
• The advertisers can reach an internet user, who is a current or

potential client of the brand, through different channels. Examples
of channels include ad banners on a web page, paid results on
a search-engine result page, and online video ads, among others.
In a general case, the set of channels is not limited to internet
channels and can include other media, like TV ads or printed
catalogs.
• Each channel is represented by multiple publishers, for example,

websites. Publishers sell their ad inventory, that is, the available
slots that can contain the actual ads.
• The publishers and advertisers are connected by means of an

ad exchange. The exchange receives ad requests from publishers
when a piece of inventory becomes available (e. g., an internet
user opens a web page) and distributes the requests across ad-
vertisers, who, in turn, can buy the available ad slot and show
the ad to the user. The exchange is often organized as a Vick-
rey (second-price) auction that processes each ad request in real
time, so the exchange is commonly referred to as a real-time bid-
ding (RTB) process.
• A user is the recipient of the ads delivered via the channels. The
user can interact with multiple channels and publishers over
time, receiving ad realizations known as impressions. From the
brand perspective, the user either eventually converts to produce
some desired outcome, such a purchase on the brand’s website,
or does not convert. Consequently, there is a funnel of sequential
impression events Ai for each user that ends with the outcome
Y, as shown in Figure 3.25.
• Finally, the impressions and conversions are tracked by an attri-
bution system. We consider the attribution system as an abstract
entity that can trace the user identity across channels and pub-
lishers and can keep records of which user received which im-
pression from which advertiser at which point in time. The pur-
pose of the attribution system is to measure the effectiveness of
the ad campaign and provide insights into the contributions of
individual channels, advertisers, and user segments. The attribu-
tion system typically collects the information by using tracking
pixels attached to ad banners and conversion web pages; users
are identified with web browser cookies. However, the attribu-
tion process can consume additional data sources, such as pur-
chases in brick-and-mortar stores, correlate this data with online
profiles by using loyalty or credit card IDs, and measure causal
effects across online and offline channels.
In the environment model above, the brand relies on the attribution
system to measure the effectiveness of individual advertisers and ad
campaigns as a whole. The metrics produced by the attribution sys-
tem directly translate into the advertiser’s fees and the brand’s costs
and revenues, so we will spend the next section examining attribution
models and their impact on advertiser’s strategies.
3.8.2 Objectives and Attribution
Similarly to the case with promotions, the business objectives of the

brand are driven by a desire to shift the relationships with certain con-
sumers from one level to another:
brand awareness The marketer is generally interested in making
its brand recognizable to potential customers and making it as-
sociated with a certain category of products, such as soft drinks
or luxury cars, even if this does not immediately translate into
conversions.
customer acquisition Acquisition aims to attract prospects who
do not interact with the brand and drive them to conversion.
retargeting Retargeting, also known as remarketing, focuses on

prospects who have already interacted with the brand, where
there is potential to develop the relationship with them. A typi-
cal example is internet users who visited the brand’s website one
or more times but did not convert.
These primary objectives can be complemented by additional con-
straints that are important for the brand. For example, the brand
might not be willing to advertise on websites with adult, violent,
or hateful content. Ideally, the contract between the brand and the
advertiser should be designed in such a way that the advertiser is
paid for achieving the objectives above. More specifically, the desired
properties of the contract can be described as follows:
• The targeting and bidding processes should be driven by the
business objective of the campaign (e. g. , brand awareness, acqui-
sition, or retargeting) and be restricted by additional rules such
as brand safety.
• The effect of the campaign should be measurable, and the metrics

should accurately reflect the value added by the advertiser. In
other words, the metrics should answer the question “What will
happen with the business objective if the advertiser is removed?”
Note that this question directly relates to the uplift modeling
discussed earlier in this chapter.
• It should be possible to answer the above question about adver-

tiser removal for the case of multiple advertisers working for the
same brand. Credits should be attributed to advertisers propor-
tionally to their contribution to the total value increment.
Unfortunately, it is not straightforward to define a contract that fully
meets the above criteria. The business objectives can be formalized in
different ways, and the measurement of the incremental value also rep-
resents a non-trivial statistical and technical problem. Let us first de-
scribe several basic methods that are widely used in practice, and we
will then discuss the outstanding questions and limitations that should
be addressed in later sections.
From the brand perspective, the overall effectiveness of the campaign
can be measured by using the cost per acquisition (CPA), which is de-
fined as the total cost of the campaign Ccamp divided by the total num-
ber of conversions Nconv :
CPA
Ccamp
(3.53)
Nconv
Conversion, however, can be defined in different ways. One possible
approach is to count post-view actions, that is, to count the users who
visited the brand’s site or made a purchase within a certain time in-
terval (for example, within a week) after they received an impression.
A more simple method is to count the immediate clicks on advertise-
ments, which is referred to as the cost per click (CPC) model. From the
advertiser perspective, it makes sense to split the cost of the campaign
into a product of the number of impressions and the average cost of
one impression, so the CPA metric can be expressed as follows:

Nimpr E cimpr
CPA CR E
1 (3.54)
cimpr
Nconv
in which Nimpr is the total number of impressions delivered by the

campaign, E cimpr is the average price paid by the brand for one
impression, and CR is the conversion rate. The advertiser’s margin is
the difference between the price paid by the brand and the bid value
placed in the RTB, so we can define the advertiser’s equivalent of the
CPA as follows:

CPAa CR
1
E cimpr cbid (3.55)
We need to specify contracts for cimpr and cbid in order to evaluate

the above expressions for CPA and CPAa . The price cimpr paid by the
brand is typically fixed, although there are two different types of con-
tracts that are used in practice:
• Cost per action , also known as cost per acquisition (CPA) or pay per
acquisition (PPA), contract. The brand pays a fixed fee for each
conversion measured by the attribution system.
• Cost per mile (CPM) contract. The brand pays a fixed fee for each
impression, but eventually measures the overall CPA by using
the attribution system.
Both approaches are equivalent in the sense that the advertiser has
to minimize the CPA metric to satisfy the client, even for CPM con-
tacts. The fixed fee implies that the CPA metric in equation 3.54 can
be optimized by maximization of the conversion rate CR. However, the
bid value cbid in equation 3.55 is not fixed and directly influences the
conversion rate, so optimization of the CPAa metric requires joint op-
timization of CR and cbid .
The final area we need to cover is attribution in the case of multi-
ple advertisers. The most basic approach is last-touch attribution (LT),
which gives all the credit to the last impression that preceded the con-
version. Consequently, the goal of the advertiser under the LT model is
to identify customers who are likely to convert immediately after the
impression.
The CPA and LT assumptions – we will refer to these settings as the

CPA-LT model – provide a reasonably complete and formal problem
definition that can be used for targeting process optimization. However,
the CPA-LT model is overly simplistic and has a number of issues and
limitations:
• There is no explicit relationship with the business objective. The
model does not distinguish acquisition, awareness, or retargeting
goals. In fact, CPA-LT principles are geared towards consumers
with a high propensity to purchase, which implies a heavy bias
towards retargeting and tactical acquisition rather than aware-
ness and strategic acquisition.
• The model suggests optimization of the response, not uplift. This

can lead to meaningless results under certain circumstances. For
example, a targeting method that identifies only high-propensity
users who are likely to convert without any impressions will have
a very good performance under the CPA-LT model, although this
is unlikely to be a good approach from the ROI standpoint.
• Last-touch attribution encourages advertisers to cheat and pig-

gyback on each other’s efforts. For example, an advertiser can
buy a lot of low-quality inventory, such as ad slots at the bot-
tom of web pages, to “touch” as many use users as possible (the
so-called carpet bombing).
We will discuss how to optimize targeting and bidding strategies
under the CPA-LT model in the next section, and we will then investi-
gate how the shortcomings of this model can be addressed with more
sophisticated attribution and controlled experiments.
3.8.3 Targeting for the CPA-LT Model
The basic goal of targeting under the CPA-LT model is to identify users
who are likely to convert shortly after the impression. Similarly to the
case of promotion targeting, we use a variant of look-alike modeling to
solve this problem, but we want to explicitly account for the informa-
tion about a user’s response to advertisements as opposed to selecting
natural buyers based on purchase histories. In particular, we want to
account for the performance of the currently running advertisement,
which means that we have to dynamically adjust our targeting method
based on the observed results. In other words, we want to build a self-
tuning targeting method.
We can assume that the advertiser has the following data for each
consumer profile:
• Visited URLs. The advertiser listens to bid requests and other

partner data sources that allow the user’s browsing history
to be captured. URLs can include both the domain, such as
google.com, and the address of a particular page.
• User attributes. The advertiser can receive additional information

about the user along with the URLs: for example, the proper-
ties of a browsing device and applications, geographical location,
time spent on a page, and some others.
• Bids and impressions. The advertiser can track the bids it made
for a given user and the impressions delivered to the user.
• Ad clicks. The advertiser can get information on how the user

interacted with the delivered ads.
• Conversions. The brand can provide the advertiser with informa-

tion about the conversions on its website.
• Additional brand data. The brand can provide additional data,

such as products browsed by a user on the brand’s website.
The features for predictive modeling can be engineered based on these
data elements. Visited URLs and derived characteristics, such as re-
cency and frequency, are known to carry a lot of predictive informa-
tion about conversions. A major challenge, however, is a high number
of observed URLs, in that a model that consumes a binary vector where
each element, zero or one, indicates whether a user visited a URL or
not, might have millions of dimensions.
The straightforward approach for the problem of self-tuning target-
ing is to start the campaign with random targeting, that is, to bid for
random people, wait for a sufficient number of conversions, and then
train the scoring model by using converted users as positive examples
and non-converted users as negative examples, as shown in Figure 3.26.
This approach, however, is not optimal because conversion events are
very rare in the case of random targeting and the dimensionality of
user profiles, as discussed above, is very high, so the creation of a suf-
ficient training data set by using random bidding at the beginning of a
campaign can be impractically expensive [Dalessandro et al., 2012a].
There are many different techniques that can help to improve the
basic approach described above. In the rest of this section, we closely
follow the staged targeting methodology described in the work of [Da-
lessandro et al., 2012a] and [Perlich et al., 2013], which provides a com-
prehensive practical solution for self-tuning targeting. The approach
is to perform the targeting process in three sequential steps: calculate
Figure 3.26: Desirable sampling for the targeting task. The shaded circles corre-
spond to positive and negative examples.
the brand proximity, incorporate the ad response, and incorporate the

inventory quality and calculate the bid amount.
3.8.3.1 Brand Proximity

The goal of this step is to estimate the probability of conversion Y re-
gardless of the ad impact, that is, to calculate the unconditional brand
proximity PrpY | uq for user u. If historical information about visitors to
the brand site is available before the campaign begins, the advertiser
can create models for a converting user by selecting converted profiles
as positive examples and random profiles of internet users as negative
examples. Note that this sampling is different from the desired sam-
pling depicted in Figure 3.26. This step is essentially look-alike mod-
eling that uses visited URLs as features and conversions as labels to
model the unconditional brand proximity:
ϕpuq PrpY | uq
PrpY | URL1 , . . . , URLn q
(3.56)
in which URLi are binary labels equal to one if the user visited the
corresponding URL and zero otherwise. The advertiser can use differ-
ent definitions of the URL and conversion to build multiple models
ϕu1 , . . . , ϕuk that capture different indicators of proximity:
• The URLs can be aggregated into clusters, and labels URLi can
be replaced by per-cluster binary labels that indicate whether the
user visited some URL from a cluster or not. This reduces the di-
mensionality of the problem, which can be helpful if the number
of available conversion events is relatively small. The distance be-

tween URLs needed for clustering can be calculated based on the
inventory quality scores that are discussed later in this section.
• A conversion can be defined as a visit to the brand’s site, a pur-

chase after an impression, or any purchase.
The brand proximity model can be used to score users at the be-
ginning of a campaign when the actual data about ad responses are
not yet available. The next step will incorporate the new data when
available and adjust the scores.
3.8.3.2 Ad Response Modeling

The goal of the response modeling step is to estimate the conditional
probability of conversion PrpY | u, aq for an ad a. This step basically
does the same thing as the baseline approach described at the begin-
ning of this section – the advertiser uses the proximity model ϕ to
target users at the beginning of the campaign, but, in addition to that,
the ads are shown to a small number of random people to obtain
the desired sampling shown in Figure 3.26. The difference from the
baseline method is that we can now use the outputs of the previous
step as features rather than high-dimensional raw URLs, which makes
the learning process more efficient. Brand proximities can be supple-
mented with additional user information features fu1 , . . . , fur , such as
browser type and geographical location, so the model can be described
as follows:
ψa puq PrpY | u, aq
PrpY | ϕu1 , . . . , ϕuk , fu1 , . . . , fur q
(3.57)
The key difference between the models for unconditional proximity

ϕ and conversion propensity ψ is sampling: the family of ϕ models
is constructed to classify users as converted or non-converters regard-
less of the advertisements, whereas the ψ model classifies users as re-
sponders or non-responders and depends on the advertisement. The
scores produced by models ϕ, however, have high predictive power
for response, providing reasonable initial values for ψ and making re-
estimation of ψ more effective as the actual response data arrives.
3.8.3.3 Inventory Quality and Bidding

The final step is to incorporate additional information that is not cap-
tured in the scores produced by the model ψ and to determine the
actual bid price to be submitted to the ad exchange. With the assump-

tion that the ad exchange is a second-price auction, the optimal bid
price can be calculated as the expected value of conversion vpY q:
bopt E rvpY qs PrpY | u, aq vpY q (3.58)
The value of conversion vpY q can typically be assumed to be constant
for all users and is incorporated into some baseline bid price bbase
set by the advertiser and depending on the contract with the brand
and properties of the exchange. Consequently, propensity scores can
be considered as multipliers to scale the baseline price.
Propensity scores produced by the model ψ are generally sufficient
to make targeting and bidding decisions. The bid price for a given user
can be calculated as
bpuq bbase s1 pψa puqq (3.59)
in which s1 pq is some scaling function for the score ψ. In particular,
s1 pq can map all scores below a certain threshold to zero (no bidding),
and the threshold can be determined based on the desired number of
impressions and other considerations, as we discussed earlier in the
context of promotions.
The targeting process we have described so far considers the user
profiles and advertisements but not the context of the impression, that
is, the inventory. The quality of the inventory is important for several
reasons [Perlich et al., 2013]:
• The inventory carries information about the user’s purchasing
intent and the relevance of the ad for the user. For example, hotel
advertisements will have higher conversion rates on travel web-
sites than on news sites.
• The perception of an advertisement depends on the context. For

example, users who are reading complex technical materials may
pay less attention to ads than visitors to entertainment sites, some
ad slots may be poorly positioned and require users to scroll
down the page, etc.
Consequently, the advertiser can expect to get better results by using
the probability ωa pu, iq PrpY | u, a, iq in which i is the inventory.
The ratio ωa pu, iq and its expectation over all inventories ωa puq
Ei r ωa pu, iq s can be used as a measure of inventory quality because
it shows how much better or worse inventory i is in comparison to an
average inventory. This metric can be used as an additional multiplier
to scale the bid:

ωa pu, iq
bpuq bbase s1 pψa puqq s2
ωa pu q
(3.60)
Note that although the notation we have used implies that ωa puq
equals ψa puq, the advertiser can use different data samples and mod-
els to estimate ω and ψ depending on the available data and other
considerations. The steepness of the scaling functions s1 pq and s2 pq
determines the trade-off between conversion rates and the advertiser’s
CPA. Steep scaling functions (e. g., zero if the argument is below the
threshold and a very high value otherwise) generally maximize the
conversion rate, but these can be suboptimal from the CPA standpoint.
Scaling functions that are close to the identity function optimize the
CPA as it follows from theoretical equation 3.58 but can be suboptimal
in terms of conversion rates.
3.8.4 Multi-Touch Attribution
The obvious limitation of last-touch attribution is that the efforts that

preceded the last impression are neglected. One can work around this
by using more elaborate attribution methods that distribute the credit
according to a position of the advertiser in the funnel. Several exam-
ples of such static models are shown in table 3.9. Static weight-based
attribution, however, does not help to estimate the contribution of in-
dividual advertisers to the overall campaign effect. We need to create
an algorithmic attribution method that measures the actual contributions
and enables the brand to reward the best advertisers or channels and
remove the worst ones.
Model A1 A2 A3 A4 A5
First impression 100% – – – –

First click – 100% – – –
Last touch – – – – 100%
Linear 20% 20% 20% 20% 20%
Position-based 35% 10% 10% 10% 35%
Time decay 10% 15% 20% 25% 30%
Table 3.9: Static attribution models. The table shows the percentage of credit
assigned to each of five impressions A1 , . . . , A5 .
Let us assume that the brand works with a network of advertisers

or channels C tC1 , . . . , Cn u. We can think of this network as a set of
states that can be traversed by a user before conversion, as illustrated in
Figure 3.27. We can define the causal effect of channel Ck as the differ-
ence between the probability of conversion for the full set of channels
and the probability of conversion if channel Ck is removed:
Vk PrpY | Cq PrpY | CzCk q (3.61)
To evaluate this expression, we can enumerate all possible subsets of

the set CzCk and estimate the causal effect for each subset separately
[Dalessandro et al., 2012b]:
¸
Vk wS,k PrpY | S Y Ck q PrpY | Sq (3.62)
z
S C Ck
Coefficients wS,k model the probability distribution of particular re-

alizations of S, that is, the probability of a user traversing a certain
sequence of channels. By assuming a uniform distribution of all se-
quences, we have

1
wS,k |C||S| 1 |C1 | |S|! p|C||C|!1 |S|q! (3.63)
because we draw sequences of length |S| from the set CzCk with car-
dinality |C| 1. For example, the causal effect of channel C3 in the
network C tC1 , C2 , C3 u is given by the following equation:
V3 13 pPrpY | C1 , C2 , C3 q PrpY | C1 , C2 qq
1
pPrpY | C1 , C3 q PrpY | C1 qq
6
(3.64)
pPrpY | C2 , C3 q PrpY | C2 qq
1
3
pPrpY | C3 q PrpY | ∅qq
The attribution formula 3.62 can be difficult to evaluate in prac-
tice because long sequences of channels have relatively low realization
probabilities, which impacts the estimation stability [Dalessandro et al.,
2012b; Shao and Li, 2011]. It can be reasonable to discard all sequences
S longer than 2 channels to produce a more simple and stable model
[Shao and Li, 2011]:
¸
Vk

wS,k PrpY | S Y Ck q PrpY | Sq
z
S C Ck

w0 PrpY | Ck q PrpY | ∅q (3.65)
¸
w1 PrpY | Cj , Ck q PrpY | Cj q

j k
Figure 3.27: Example of a network with three channels.
The baseline probability of conversion PrpY | ∅q can also be dis-

carded because it is equal for all channels, and the coefficients are de-
fined as
C| 1 1 1

w0
|
| C | |C |
1
0
C| 1 1 1

(3.66)
w1
|
|C| p|C| 1q|C|
1
1
We can therefore express the causal effect as
Vk |C1 | PrpY | Ck q

¸ (3.67)
PrpY | Cj , Ck q PrpY | Cj q
1
p|C| 1q|C| jk
The probability of conversion PrpY | Ck q can be estimated as the
ratio of converted users who passed through channel Ck to the total
number of users who passed through the channel. The second-order
probabilities PrpY | Cj , Ck q can be estimated in the same way but for a
pair of channels.
Equations 3.62 and 3.67 describe a practical solution for multi-touch
attribution. However, it is worth noting that there are alternative ways
to approach the same problem. For example, one can build a regres-
sion model that predicts conversions based on traversed channels and
compare the magnitudes of the regression coefficients [Shao and Li,
2011].
3.9 measuring the effectiveness 157
3.9 measuring the effectiveness
The effectiveness of marketing campaigns is inherently challenging to

measure because each consumer has unique properties, changes over
time, and interacts with the brand and marketing media in their own
way, so the attribution of any improvement or degradation to a par-
ticular marketing action can always be debated. Marketers typically
cannot strictly prove the effectiveness of the action, but they can try
to arrange an experiment or analyze the collected data in such a way
that the actions in question and the outcomes are properly isolated, so
that the causal effect cannot be attributed to external factors. This can
be considered as proof of a statistically significant causal relationship
between the actions and outcomes.
This problem statement enables us to leverage a huge statistical the-
ory that was developed in other fields long before algorithmic market-
ing appeared. Importantly, experimentation frameworks developed in
areas such as biology and healthcare are specifically adapted to deal
with scenarios that are structurally similar to marketing campaigns.
3.9.1 Randomized Experiments
Consider a basic marketing campaign that distributes promotions or

advertisements to prospects to make them convert. Although our ulti-
mate goal is to estimate the causal relationship between the treatment
and conversions, we can start with very basic questions and gradually
build a statistical framework that estimates the causal effects.
3.9.1.1 Conversion Rate

One of the most basic questions that we can pose is the measurement
of simple metrics, such as the conversion rate. Based on the assump-
tion that the total number of individuals who received a treatment n
is known and the number of converted individuals k among these n
recipients is measured, we can estimate the conversion rate as
R
k
(3.68)
n
The obtained estimate may or may not be statistically reliable, depend-
ing on the number of individuals and conversions. If these numbers
are small, we can expect the measured rate to have high variance and
to change drastically if the same campaign is run multiple times. If
the numbers are high, we can expect more consistent results. The re-
liability of the estimate can be measured in different ways by using
different statistical frameworks. In this book, we generally advocate
Bayesian methods and Monte Carlo simulations because of their con-

sistency and flexibility, so we use this approach for randomized exper-
iments as well. Although it is not necessarily the most simple solution
for the most basic problems, it helps us to establish a framework that
can be efficiently extended for the more complex scenarios that we will
consider later on.
Because the total number of promotions n is a non-random value
chosen before the experiment, our goal is to understand the distribu-
tion of the conversion rate given the observed number of conversions
ppR | kq. If this distribution is known, we will be able to estimate the
probability that the results of hypothetical repeated experiments would
deviate significantly from the observed value and thus measure the re-
liability of the estimated rate. According to Bayes rule, the distribution
in question can be decomposed as follows:
ppk | RqppRq
ppR | kq
p pk q
(3.69)
in which ppk | Rq is the likelihood, that is, the probability, of observing

k conversions given that the conversion rate value is known to be R,
and ppRq is the prior distribution of conversion rates. The probability
ppkq can be viewed as a normalization factor because the data point
k is given; hence, this term just ensures that the rate distribution is a
probability distribution, that is, that the integral over the entire range
is 1. So this value can be expressed as follows:
»
ppkq ppk | RqppRqdR (3.70)
In words, we start with a prior belief about the rate distribution ppRq,
and the observed data, that is, the number of conversions k, provide
evidence for or against our belief. The posterior distribution ppR | kq is
obtained by updating our belief based on the evidence that we see.
As the posterior rate distribution includes two factors, ppk | Rq and
ppRq, we need to specify these two distributions. Under the assumption
that the conversion rate is fixed, the probability that exactly k individ-
uals out of n will convert is given by a binomial distribution, with a
probability mass function of the form

p p k | Rq
n
k
Rk p1 Rqnk
(3.71)
k!pnn! kq! Rk p1 Rqnk
The second factor, the prior distribution ppRq, can be assumed to
be uniform or can be estimated from historical campaign data. Let us
consider the case of uniform distribution first. If the prior distribution

ppRq is uniform in the range from 0 to 1, the posterior distribution
ppR | kq has the same form as the likelihood given by equation 3.71,
but it is now a function of R, not k, so the normalizing constant will be
different. We can denote this constant as cpn, kq and obtain
ppR | kq Rk p1 Rqnk cpn, kq (3.72)

This distribution is known as a beta distribution, and there is a stan-
dard notation for it. In this notation, the posterior can be expressed
as
ppR | kq beta pk 1, n k 1q (3.73)
in which the beta distribution is defined as
beta pα, βq
1
Bpα, βq
xα1 p1 xqβ1
»1 (3.74)
Bpα, βq xα1 p1 xqβ1 dx
0
The distribution of the conversion rate given n treated individuals
and k conversions is described by the beta distribution.
If the prior distribution is not uniform, it can also be modeled as the
beta distribution:
ppRq beta px, yq (3.75)
in which parameters x and y can be estimated, for example, based on
historical data. In this case, the posterior distribution is still the beta
distribution:
ppR | kq 9 ppk | Rq ppRq
9 Rk p1 Rqnk beta px, yq
9 Rk x1 p1 Rqnk y1
(3.76)
9 beta pk x, n k yq
It is said that the beta distribution is the conjugate prior to the bi-
nomial distribution: if the likelihood function is binomial, the choice
of a beta prior will ensure that the posterior distribution is also beta.
Note that beta p1, 1q reduces to a uniform distribution, so result 3.73
obtained for the uniform prior is a particular case of expression 3.76.
We can now estimate the probability that the conversion rate R lies
within some credible interval ra, bs as
»b
Prpa R bq beta pk 1, n k 1q dR (3.77)
a
Equation 3.77 can be evaluated analytically, but we can also estimate

the credible interval for a conversion rate by using Monte Carlo simu-
lations. In this case, the estimation process can be described as follows:
1. The inputs are n, k, and the desirable confidence level 0 q
100%.
2. Generate a large number of random values with distribution
beta pk 1, n k 1q.
3. Estimate the q{2-th and p100 q{2q-th percentiles of the gener-
ated values to obtain the desired credible interval. For example,
we can be 95% confident that the estimate R lies in between the
2.5% and 97.5% percentiles.
Examples of beta distributions for different values of n and k, as well
as the corresponding credible intervals, are shown in Figure 3.28. The
simulation approach can look excessively complicated for the assess-
ment of basic metrics such as the conversion rate, but its advantages
will become more apparent as we move to more complex cases.
p | q
Figure 3.28: Examples of the posterior distribution p R k for the uniform
prior and different sample sizes n. The mean k n { 0.1 for all
samples. The vertical lines are the 2.5% and 97.5% percentiles of the
corresponding distributions. We start with the uniform prior, and
the more samples we get, the narrower the posterior distribution
becomes.
3.9.1.2 Uplift
The conversion rate by itself is not a sufficient measure of the quality of
a targeting algorithm or the effectiveness of a marketing campaign. As
we discussed earlier in this chapter, the effectiveness is typically mea-

sured as the uplift, which is the difference between conversion rates in
the test and control groups. The conversion rate in the control group
is considered as the baseline, and the uplift can be estimated as the
conversion rate in the test group measured against the baseline rate:
L 1
R
(3.78)
R0
in which R0 is the baseline conversion rate and R is the conversion rate
for the campaign in question. From a statistical standpoint, we also
want to measure the reliability of this estimate, that is, the probability
PrpR ¡ R0 | dataq (3.79)
to ensure that the obtained results are attributed to the impact of the
campaign in question relative to the baseline, not to some external un-
controlled factors. The standard way to tackle this problem is random-
ized experiments. The approach is to randomly split the consumers who
can potentially be involved in the campaign into two groups (test and
control), provide the test group with the treatment (send promotion,
show ads, present a new website design, etc.), and provide the control
group with the no-action or baseline treatment. Random selection of
test and control individuals is important to ensure that the observed
difference in outcomes is not caused by a systematic bias between the
two groups, such as a difference in average income. Running the test
and control in parallel is also important to ensure equality of the test
conditions for the control groups, which might not be the case, for
example, in a comparison of new data with historical data.
The design of randomized experiments for targeted campaigns is
illustrated by Figure 3.29. The high-propensity customers identified
by the targeting algorithm are divided into test and control groups,
and the test group receives the treatment. The number of positive and
negative outcomes is measured for both groups: nT and nC are the
number of individuals and kT and kC are the number of conversions
in the test and control groups, respectively. The uplift is measured by
comparing the conversion rate of the test group kT {nT with that of the
control group kC {nC .
We now want to assess the probability PrpRT ¡ RC q or, equivalently,
to find a credible interval for uplift L. We can calculate this in a similar
manner to that we used for the credible interval of the conversion rate
in expression 3.77, but now we need to account for the joint distribution
for RT and RC :
¼
Prpa L bq LpRT , RC q PrpRT , RC qdRT dRC (3.80)

a L b
Figure 3.29: Sampling in randomized experiments.
If the randomized experiments are properly designed and executed

to achieve independence between the test and control groups, we can
assume that the joint probability above can be split into individual
distributions of conversion rates:
PrpRT , RC q PrpRT | kT , nT q PrpRC | kC , nC q (3.81)
At this point, we can apply the same simulation approach that we

used for the individual conversion rate. Conversion rates RT and RC
follow the beta distribution, so we can generate the uplift samples by
drawing two conversion rates from the corresponding beta distribu-
tions and calculating the ratio. The process will be as follows:
1. The inputs are values kT , nT , kC , and nC , measured from the
observed data, and the desirable confidence level is 0 q
100%
2. Generate a large number of values L by computing each sample

as follows:
a) Draw RT from the distribution beta pkT 1, nT kT 1q
b) Draw RC from the distribution beta pkC 1, nC kC 1q
c) Compute L RT {RC 1
3. Estimate the desired credible interval for L by taking the q{2th

and p100 q{2qth percentiles for the generated values.
The above approach works for many practical scenarios, for exam-
ple, promotional campaigns, advertisements, and testing of arbitrary
improvements, such as a new design of a website. Randomized ex-
periments, however, impose certain limitations on how a campaign is
executed, and this can be a problem in some cases. In particular, the

requirement for a control group can incur additional expenses – we
will study this issue in detail in the next section.
It is important to note that measurement of the revenue uplift does
not necessarily require a conversion rate to be measured or even indi-
vidual conversions to be tracked. We just need to know the total rev-
enues generated by the test and control groups over a period of time
after the campaign and estimate the uplift as a ratio between these two
revenue values. This can be the only way to measure the uplift if the
conversion information is not available.
3.9.2 Observational Studies
Randomized experiments can be used in the online advertising en-

vironment to measure the conversion uplift delivered by campaigns.
Randomized methods, however, require one to be very careful with
control group selection to make sure that there is no systematic bias
between the test and control groups. The standard approach to achieve
unbiased randomization is to leave picking of the control users until
the very end of the ad delivery pipeline and to sample the users af-
ter the targeting and bidding stages, as shown in Figure 3.30. The test
users are exposed to the actual ad impressions and the control users
are exposed to some dummy ads, such as public service announce-
ments (PSA), so the uplift between the groups is a measure of the ad
impact.
The presence of an ad exchange, however, introduces a major chal-
lenge because impressions for the control group do not come for free
and have to be purchased, just like the actual impressions. The ques-
tion that arises is whether control group selection can be moved to
before the bidding stage, as shown in Figure 3.31.
This approach effectively means that we do not do a controlled ex-
periment anymore, because the bidding process – which bids are won
and which are lost – is not controlled, and, consequently, it can induce
an arbitrary bias in the test group compared to the control group. We
can only observe bidding outcomes and conversions and measure the
causal effect of the ad by doing statistical inference. This leads us to the
large theory of observational studies and causal inference, which was
under intensive development for decades and is driven by the neces-
sity to analyze processes that are not under the control of researchers.
Our problem with the bidding bias closely matches the problem of
treatment effect under non-compliance in clinical trials. The causal effect
of a treatment can be evaluated by using randomized experiments and
comparing subjects from the test group who received the treatment
Figure 3.30: Uplift estimation in online advertising by using randomized exper-

iments. RT and RC are the conversion rates in the test and control
groups, respectively.
with the control group. Although subjects can be assigned to the test
and control groups randomly, some people in the test group cannot
be exposed to the treatment because of compliance issues. The split
into compliant and non-compliant subgroups after randomization cor-
responds to the win–lose split in the bidding process when it follows
control group selection, so we can leverage the studies dedicated to
clinical trials with non-compliance.
The problem of uplift estimation with observational studies can
be approached by using different techniques. We start with a basic
method that illustrates how some concepts of causality theory can be
applied to the problem [Chalasani and Sriharsha, 2016; Rubin, 1974;
Jo, 2002].
We can see in Figure 3.31 that we have at least three conversion rates
that can be measured directly: RC for the control group, RL T for the lost
bids in the test group, and RW
T for the users who got actual impressions.
Our goal is to find the conversion rate RWC , which can be interpreted as
a potential conversion rate of the users who would have been won even
if they were not provided with impressions. This value is hypothetical
because we cannot go into the past, revoke the impressions we already
delivered, and see what would happen. However, it can be estimated
Figure 3.31: Uplift estimation by using observational studies.
from the known data under certain assumptions. First, we can note
that the ratio γ between the number of users who were won and the
number of users who were lost is directly observable. By assuming that
the distribution of “winners” and “losers” is the same in both test and
control groups, we can claim that
RC γ RW
C p1 γqRLC (3.82)
in which RW L
C and RC are the conversion rates that we can expect from
the control users who could be won or lost, respectively, if reassigned
to the test group. The second assumption we can make is that RL C RT
L
because both groups contain only “losers” who have not been exposed
to the ad, so we do not expect any bias between them. Consequently,
we can express RW C by using the known values as follows:

RW
C γ1 RC p1 γqRL
T (3.83)
Finally, the uplift can be estimated as the ratio between the observed
RW
T and inferred RC .
W
The reliability of the uplift estimate can be evaluated by using the

same simulation approach we used for randomized experiments. This
requires us to generate samples according to the uplift distribution,
which can be quite challenging to specify because it is a composition
of several random processes: control group selection, bidding, and con-
versions. We observe only a few bits of information for each realization
of this complex process (assigned group, bidding outcome, and con-
version outcome), but we do not observe the intrinsic properties of the
users and other latent factors that determine the joint probability dis-
tribution of the observed outcomes. In the rest of this section, we will
discuss a statistical framework that combines the idea of potential out-
comes discussed above with advanced simulation methods to infer dis-
tributions of different campaign properties, including, but not limited
to, the uplift [Chickering and Pearl, 1996]. We describe the framework
in two steps. First, we specify the model of the random processes of
interest. Second, we discuss how the model can be evaluated by using
simulations.
3.9.2.1 Model Specification

We can account for the latent factors and their impacts by using the
graphical model presented in Figure 3.32. Each node represents a ran-
dom variable, and the arrows indicate the dependencies between the
nodes. Random variables Z, A, and Y correspond to randomization,
bidding, and conversion. More specifically, binary variable Z P t0, 1u
takes a value of one if the user is assigned to the control group and zero
otherwise, variable A P t0, 1u takes a value of one if we won the bid
and showed the ad and zero otherwise, and, finally, variable Y P t0, 1u
equals one if the user converted and zero otherwise. The random vari-
able S corresponds to the user state and, possibly, other latent factors
that influence the advertiser’s ability to win the bid and get a response
after the impression.
Figure 3.32: Graphical model for an observational study with latent factors.
At a very high level, we want to understand the joint distribution

Prpz, a, y, sq and integrate over it to obtain a credible interval for the up-
lift L. The question is how to decompose the distribution Prpz, a, y, sq to
make it computationally tractable. The graphical model in Figure 3.32
makes certain assumptions that can be used for decomposition : Z
and S are considered independent because randomization must not
be influenced by external factors, and Z and Y are conditionally inde-
pendent given A and S because conversions can be influenced only
through events A. This leads to the following decomposition of the
probability density:
Prpz, a, y, sq Prpzq Prpaq Prpa | z, sq Prpy | a, sq (3.84)
We now need to specify the state random variable S and its role in
the densities Prpa | z, sq and Prpy | a, sq. The idea behind the latent fac-
tors is to capture “the state of the world” that is not observed directly
but can influence outcomes like the uplift. This concept can be consid-
ered as a counterpart of the potential outcomes that we discussed at
the beginning of this section, because if we can infer the state from the
observations then we can evaluate the potential outcomes for different
preconditions. For example, if we know that a given user can never be
won on the exchange, then we can predict the outcomes for assigning
this user to both the test and control groups.
The latent state can be modeled differently depending on the avail-
able data, metrics of interest, and general understanding of the domain.
We use a standard model that illustrates how the latent states can be de-
fined as functions of the observed data and how metrics like uplift can
be derived from the states [Heckerman and Shachter, 1995; Chickering
and Pearl, 1996].
From a campaign efficiency standpoint, we are interested mainly in
two properties of the user: compliance with the advertising method
(ability or inability to win a bid) and response to the advertisement
(converted or not). These properties correspond to the probabilities
Prpa | z, sq and Prpy | a, sq discussed above and can be considered
as the user’s internal state that systematically influences the outcomes
obtained for the user. We can enumerate the possible states separately
for compliance and response and specify a condition for each state that
indicates whether the state is possible for the observed tuple pz, a, yq
or not.
The set of possible user states is a Cartesian product of the
compliance and response behaviors, which gives us a 16-element
Compliance Condition Description
C1 a0 User can never see the ad
C2 az User sees the ad every time we bid and

only if we bid
C3 a z User sees the ad if and only if we do not

bid
C4 a 1 User always sees the ad
Table 3.10: User compliance states and conditions. States C3 and C4 should
never be the case in the scenario that we consider, but they can occur
in other environments such as omni-channel advertising.
Response Condition Description
R1 y0 User never converts

R2 ya User converts only after impression
R3 ya User converts only without impression
R4 y1 User always converts
Table 3.11: User response states and conditions.

set ts1 , . . . , s16 u, in which si iterates through all pairs Cp , Rq of
compliance and response behaviors listed in tables 3.10 and 3.11:
S P ts1 , . . . , s16 u

p q 1 ¤ p, q ¤ 4
(3.85)
sp 4 q 1 Cp , Rq ,
Consequently, the random variable S is a 16-state random variable

drawn from the set of 16 possible states.
We directly observe binary tuples zj , aj , yj for each user j, but user
states sj are never observed directly. However, if we infer the state, it is
possible to evaluate potential metrics of interest, such as uplift, based
on the inferred states. More specifically, we are interested not in the
individual user states but in a vector of state shares
µ pµ1 , . . . , µ16 q (3.86)
in which each share µi is the ratio between the number of users in the
corresponding state si and the total number of observed users. The
metrics can then be defined as functions of µ. For example, the uplift
Lpµq can be defined as the ratio between the sum of the four µi values
that correspond to states with the R2 response component (and any

compliance component) and the sum of another four µi values that
correspond to the states with the R3 response component. It is possible,
however, to define different functions of µ to answer other questions.
3.9.2.2 Simulation
By assuming the model specified above, we can express the credible
interval of the metric Lpµq via the posterior distribution of the random
vector µ:
»
Prpa Lpµq bq Lpµq ppµ | dataqdµ
(3.87)
a L µ p q b

in which data represents all observed tuples zj , aj , yj . Let us denote
the vector of user states as

s s1 , . . . , sm (3.88)
in which m is the number of observed users. The distribution of the

state shares µ can then be considered as a random function of user
states s, which, in turn, are also random variables that are not observed
but are probabilistically inferred from the data. Consequently, we have
to consider the joint distribution of 16 variables in µ and m variables
in s:
»
Prpa Lpµq bq Lpµq ppµ, s | dataqdµds
(3.89)
a L µ p q b
The simulation approach requires the distribution ppµ, s | dataq to
be estimated based on the observed data, so we will be able to draw
vectors µ from this distribution. Once the vectors are generated, it is
possible to calculate samples Lpµq and estimate their distribution. The
question we have to answer now is how to draw samples from the
distribution ppµ, s | dataq for the sake of the simulation. We do not
know the functional form of the distribution, but statistical methods
do exist that can help us with generating samples from the distribution
without specifying it explicitly.
Gibbs sampling is a widely used method of drawing samples

from a multivariate distribution [Geman and Geman, 1984]. Let us
assume that we need to draw samples from a multivariate distribution
ppx1 , . . . , xn q. The Gibbs sampler exploits the fact that this multivariate
distribution can be split into n conditional distributions
ppxi | x1 , . . . , xi1 , xi 1 , . . . , xn q, 1¤i¤n (3.90)

It can be the case that we cannot sample points directly from the mul-
tivariate distribution, but sampling from the conditional distribution
is possible. The idea in Gibbs sampling is that, rather than probabilis-
tically picking all n variables at once, we can pick one variable at a
time with the remaining variables fixed to their current values. In other
words, each variable is sampled from its conditional distribution with
the remaining variables fixed:
xi ppxi | x1 , . . . , xi1 , xi 1 , . . . , xn q, 1¤i¤n (3.91)
This is an iterative algorithm that repeatedly draws samples from the

conditional distributions by substituting previously generated samples
into the conditions. For example, consider the basic case of two vari-
ables x1 and x2 . The variables are first initialized to some values that
can be sampled from the prior distribution and are then updated at
each iteration i according to the following rules:
piq ppx | xpi1q q
x1 1 2
piq
x p px | x
p i 1 q
q
(3.92)
2 2 1
This process may need a certain number of iterations to converge, and

then it starts to produce points that follow the distribution ppx1 , x2 q.
This method is very powerful in practice because the conditional dis-
tributions are often much easier to specify than the joint distribution
of interest. A generic version of the Gibbs sampler is provided in algo-
rithm 3.1.

p0 q
Initialize x1 , . . . , xn
p0q from the prior distribution
for iteration i 1, 2, . . . do
draw x1
piq p x pi1q pi1q
x , x , . . . , x
p i 1 q
1 2 3 n
draw x2
piq p x piq pi1q
x ,x p i 1 q
, . . . , xn
2 1 3
...
piq p x piq piq

x , x , . . . , xpiq

draw xn n 1 2 n1
end
Algorithm 3.1: Gibbs sampler.
Let us now come back to the distribution ppµ, s | dataq and inves-
tigate how the Gibbs sampler can be used to draw samples from it.
As the sampler draws each element of µ and s separately, we can

separately specify the estimation routines for pps | µ, dataq and
ppµ | s, dataq.
For the first probability, we can leverage the assumption that the
users are independent, so the posterior probabilities of the user states
are given by
p p sj
si | µ, s, dataq 9 ppaj , yj | zj , si q µi (3.93)
in which ppaj , yj | zj , si q is the likelihood of observing the outcomes

zj , aj , and yj given the state si . We can assume that the likelihood
is equal to one if the observed outcomes agree with the conditions of
state si and zero otherwise. Consequently, the likelihood of state si for
user j can be estimated based on the known values of aj , yj , and zj
and the state conditions from tables 3.10 and 3.11. For a model with 16
states, we estimate a vector of 16 probabilities for each user. This vector
is then multiplied by the prior probability of the state µi , in accordance
with the right-hand side of expression 3.93. The resulting vector of 16
numbers defines the multinomial distribution from which sample sj
can be drawn.
The second part is the conditional distribution ppµ | s, dataq. Let us
denote the number of times state si occurs in s as ni . Because µ is the
vector of state shares, that is, each element µi is the empirical probabil-
ity of state si , the vector of counters ni has a multinomial distribution
with parameter µ. Consequently, the likelihood of observing vector s
given the state shares µ is
¹
µn
i
i
(3.94)
i
and thus the posterior distribution of the state shares is

¹
ppµ | s, dataq 9 i Prpµq
µn i
(3.95)
i
The last step is to specify the prior distribution Prpµq. Recall that we
have used a beta distribution for the prior in randomized experiments
because the likelihood had a binomial distribution and the beta distri-
bution is a conjugate prior to the binomial. In a similar way, we now
have a multinomial likelihood and its conjugate prior is the Dirichlet
distribution (see Appendix A): if we choose Prpµq to be the Dirich-
let, the posterior distribution described in expression 3.96 will also be
Dirichlet. More formally, we can express the prior belief as a set of
counters n0i , which are used as parameters of the prior Dirichlet distri-
bution, and the posterior can then be expressed as
¹
ppµ | s, dataq 9 i Dirpn1 , . . . , n16 q
µn i 0 0
i
¹ n 0 n 1
9 i
µii (3.96)
i
9 Dirpn01 ni , . . . , n016 n16 q
The above equations can be plugged directly into the Gibbs sampler:
we generate the samples of µ by using expression 3.96, generate m sam-
ples of s by using equation 3.93, and then repeat this process iteratively
until we have enough realizations of vector µ to evaluate the credible
interval of Lpµq.
3.10 architecture of targeting systems
Targeting systems can be implemented differently depending on the

particular industry and applications. However, some logical compo-
nents are common for the majority of targeting systems. In this section,
we consider a canonical architecture that includes all of the major log-
ical blocks needed to create targeted advertisements or promotions.
This architecture assumes that the system functions in the request–
response mode, that is, it receives real-time requests that contain some
context information, such as consumer ID or channel ID, and returns
one or more offers generated for this particular context. We consider
this application and design to be the most universal and important;
however, it can be adapted to other applications, such as batch email
generation.
A high-level logical architecture for a targeting system is shown in
Figure 3.33. This architecture assumes that there are three major sub-
systems, each of which contains multiple components.
3.10.1 Targeting Server
The targeting server encapsulates most of the logic related to process-

ing of the incoming requests and crafting the response with adver-
tisements or promotions. It can be thought of as a pipeline with the
following typical stages:
conditions According to the targeting process that we previously
described, one of the first steps is to validate explicit restric-
tions for all candidate advertisements or promotions. Examples
3.10 architecture of targeting systems 173
Figure 3.33: High-level architecture of a targeting system.
of such conditions are the presence of certain products in the

current shopping basket, certain geographical locations, etc.
a/b testing Evaluation of multiple targeting strategies in parallel

is a standard technique that helps to combat the inaccuracy of
data and models, as well as to measure the performance of the
campaigns. The targeting server is able to assign different adver-
tising methods to different users and report performance metrics
for each method separately, so the optimal strategy can be se-
lected later on. A/B testing can be done for different aspects of
the targeting process and consumer experience, such as different
scoring methods, different text messages or images, etc.
Strategy selection is typically sticky with regards to consumers
– once a certain strategy is selected, it is saved to the consumer
profile and is consistently used for all requests related to this
person. This helps to achieve a consistent user experience and
create disjointed consumer segments for different strategies.
It is also typical to compare one or more experimental strate-
gies with a control group, which is a group of consumers who
receive an old or basic experience that can be considered as a
baseline strategy or no strategy at all. The performance of the ex-
perimental strategies can then be measured as uplift relative to

the baseline.
scoring The targeting server scores the incentives that have passed
the previous step by evaluating propensity models associated
with the incentive against the context, including the consumer’s
historical profile. Propensity models can be dynamically selected
for an incentive based on metadata and business rules to avoid
manual binding of a model to each and every incentive.
budgeting rules The final response is created based on a list of

valid scored candidate incentives by applying budgeting rules
and other constraints that limit the number of impressions for a
given consumer, channel, or campaign.
3.10.2 Data Management Platform
The data management platform can be viewed as an operational

database that stores customer profiles and other data required for
targeting, including campaign configurations. The following items
can be considered as the main components of the data management
platform:
profile repository The repository that stores historical data about
individual consumers. This can include both raw data, such as in-
dividual orders or website events, and aggregated statistics (fea-
tures) that can be directly used in the evaluation of propensity
models. The repository can be populated by using context data
from the targeting server and external data sources.
model repository The model repository stores the pool of propen-

sity models that are used by the targeting server for scoring.
These models are created and updated by the analytics platform
described below.
campaign repository The campaign repository contains configu-

ration details of campaigns, including graphical assets, condi-
tions, budgeting limits, etc.
3.10.3 Analytics Platform
The analytics platform collects, consolidates, and stores the customer

profile data, as well mix-ins needed for modeling, data preparation,
and reporting purposes. Examples of such mix-ins include product cat-
alog information, sales data, and store data, among others. One of the
3.11 summary 175
main outputs of the data preparation process is the profile features that
can be used for training and evaluation of predictive models. The data
consolidated and cleaned by the platform can be provided with differ-
ent service level agreements for different consumers. For example, pro-
file features can be created in batch mode for analytical purposes, but
the same features are needed for real-time model evaluation in the tar-
geting server. Thus, some data preparation modules can be designed
to work in different modes. This aspect is illustrated by the feature
preparation block in Figure 3.33, which is used for both modeling and
real-time data aggregation for the data management platform.
One of the main functions of the platform is to produce propensity
models by running machine learning algorithms over the data that
come from the targeting server and external data sources. The platform
can provide tools for manual model creation and automated model
updates (re-training). In addition to that, the analytics system performs
measurements and provides capabilities for reporting and exploratory
data analysis.
Finally, the analytics platform can host a planner. This is a key com-
ponent of a programmatic marketing system that designs and opti-
mizes promotion and advertising campaigns. The planner uses histor-
ical data and statistics, business rules, best practices, and heuristics
to determine the optimal strategies (duration and type of incentives,
channels, propensity models, etc.) based on the input objective and
additional constraints like budgeting limits. It can also forecast the per-
formance of a campaign by using historical data. The planner might
have several functional blocks including
investment planner The investment planner provides a high-
level view of market opportunities derived from historical data.
It helps the end user to set the right business objectives and dis-
tribute the budget amounts across different strategic directions
and campaigns. It can be considered as a global optimization
tool.
campaign planner The campaign planner optimizes the individ-

ual promotional and advertising campaigns that are suggested by
the investment planner. It calculates optimal time frames, spend-
ing rates, etc.
3.11 summary
• Promotion and advertising services focus on the problem of target-

ing, that is, finding the optimal match between consumers and offer-
ings. Depending on the application, the system may need to identify
the offerings for a given consumer or the consumers for a given of-
fering.
• Promotion and advertising services are typically driven by revenue

optimization objectives, but it is important to focus on customer ex-
perience optimization as well.
• The main business environments for promotion and advertising ser-

vices are sales promotion and online advertising. The main actors
and entities in the sales environment are consumers, manufacturers,
retailers, and marketing campaigns.
• The business objectives of manufacturers and retailers can be mod-

eled in terms of campaign costs and gains. The immediate objectives
include campaign profitability, and the strategic objectives can be
described in terms of the customer life cycle. The key strategic objec-
tives are the acquisition of new customers, maximization of existing
customers, and retention of customers who may be lost.
• A targeting system can be designed as a pipeline that starts with

resource allocation across the objectives, followed by fitting of cam-
paign templates with the objective, linking targeting models, and,
finally, executing the campaigns.
• The response modeling framework fits together the campaign costs,

revenues, and statistical properties of the customers. The main prin-
ciple of response modeling is to maximize the uplift, that is, the
incremental profit gain. The uplift can be measured after the fact by
using test and control groups.
• The main building blocks used in targeting systems are propensity

models, time-to-event models, and lifetime value models. The most
basic examples of targeting models are loyalty tiers and RFM anal-
ysis. These approaches focus on the financial outcomes and do not
delve deeply into the drivers of the customer behavior.
• The goal of propensity modeling is to find consumers who have a

relatively high probability to behave in a certain way, for example, to
buy a new product. Look-alike modeling is one of the main methods
of propensity modeling.
• It is often more convenient to measure the time-to-event than the

probability of the event. This can be done by using survival anal-
ysis. Survival models, similarly to regression models, can express
the time-to-event as a function of independent variables such as cus-
tomer properties or discount depth.
3.11 summary 177
• Lifetime value models estimate the total amount of money that a

brand is likely to make from a given customer over the lifetime of
their relationship. LTV modeling can be done with descriptive and
predictive approaches.
• A marketing campaign can be assembled from multiple building

blocks. A campaign template can include targeting conditions, scor-
ing models, and budgeting and capping rules. A campaign typically
corresponds to a certain point or set of points in a customer journey
and aims to influence it. Examples of campaign templates include
product promotional campaigns, multistage campaigns, and reten-
tion and replenishment campaigns.
• Targeting can be viewed as resource allocation to customers, but the

resources can also be allocated to channels, objectives, territories,
and other criteria.
• Many methods and principles of promotion targeting are applicable

in other domains, such as online advertising, but each domain has its
own business objectives and implementation challenges. The online
advertising objectives are often defined in terms of cost-per-action
and attribution across the competing advertisers.
• Online advertising uses a large number of targeting methods and

techniques. Many of them are based on look-alike modeling, notions
of brand proximity, response probability, and inventory quality.
• The effectiveness of promotion and advertising campaigns is typi-

cally measured by using randomized experiments involving test and
control groups. In some environments, including online advertising,
the control groups are associated with additional costs or missed
gains, so the effectiveness can be measured with the more advanced
methods of observational studies.
4
SEARCH
Targeted promotions and advertisements are focused on the problem

of finding the right audience for a given product or service. An equally
important counterpart of targeting is product discovery, that is, the task
of providing customers with convenient services and interfaces for
browsing the assortment and searching for the products they need.
Targeting and discovery services are the two principal programmatic
marketing tools that can be leveraged to improve different aspects of
customer awareness of a product, service, or brand.
The product discovery problem revolves around the notion of pur-
chasing intent. In some cases, customers can explicitly express their in-
tent by entering a search query or specifying desired product attributes
in a different way. In other cases, the intent is not explicitly expressed,
and a programmatic service has to infer it from known customer at-
tributes and behavior. These two scenarios are often distinguished and
thought of as two different categories of services – search services help
customers to find products that meet explicitly expressed requirements,
whereas recommendation services do not require users to express their
search intent in order to make suggestions. However, the boundary be-
tween search and recommendations is blurry. A basic search service
can use only an explicitly entered query to find products. A more ad-
vanced solution can use additional pieces of information about a user
to personalize search results. In certain applications, these implicit sig-
nals can become more important than the explicit ones and a search
service structurally metamorphoses into recommendations. From this
perspective, search and recommendation services can be compared to
a store sales associate who can find a specific product upon request or
can suggest options from only basic information about the customer
and their needs.
Search and recommendations are important not only as functional
services but also as a fundamental capability to combine multiple
strong, weak, or noisy signals to correctly understand customer needs
179
180 search
and determine products that match these needs. This capability is the
key to building efficient customer-facing services and applications.
The first product discovery service that we will consider is the search.
The purpose of a search service is to fetch offerings that are relevant
to the customer’s search intent expressed in a search query or with
selected filters. This type of problem is addressed by information re-
trieval theory, so we have a wide range of theoretical frameworks and
practical search methods at our disposal. The primary goal of this sec-
tion is to put together, align, and adapt the frameworks and methods
that are relevant to marketing applications. Some of these methods are
borrowed from the toolkit of generic search methods developed in in-
formation retrieval theory, and others were developed specifically for
marketing and merchandising. We will be taking a practical approach
to search methods and will focus on industrial experience, techniques,
and examples, rather than information theory. At the same time, we
will try to avoid implementation details, such as data indexing, as
much as possible and will stay focused on the business value deliv-
ered through relevant search results.
We will start this chapter with a review of the environment and eco-
nomic objectives. We will then demonstrate that the problem of rele-
vant search can be expressed in terms of features, signals, and controls,
similarly to other programmatic services. We will review a number of
methods for engineering, mixing, and tuning these signals and controls
in manual mode, and we will then discuss how predictive analytics can
be leveraged for automated optimization.
4.1 environment
A search is one of the most natural and convenient interfaces between

a human and a computer. This makes search an integral part in a wide
range of services and applications that belong to different domains
and use search functionality in very different ways. These environ-
ments can vary dramatically in both economic objectives and technical
properties, such as data volume, which can dramatically influence the
design and implementation of search services. Let us consider a few
major examples:
web search The web is a collection of web pages that contain tex-
tual and media information, so it is not surprising that the web
search has initially been approached as a text analysis problem.
The scoring criteria of the first search engines were primarily
based on the page content, which created a lot of opportunities
for website owners to cheat the system by tampering with hid-
4.1 environment 181
den fields and other relevance-boosting techniques. The search

engines were forced to develop a totally new strategy that used
cross-references between websites and their trustworthiness as
the main scoring signal, so web domains and pages that are ref-
erenced by hyperlinks from trustworthy and popular sites are
scored highly and the relevance of non-referenced resources is
sharply discounted. This approach, pioneered by Google as a
PageRank algorithm, became an important distinguishing fea-
ture of web search in comparison with most other search envi-
ronments. The large scale of the web and necessity of indexing
enormous amounts of data has also influenced web search meth-
ods to a very significant extent.
merchandising search In many search applications, organic

relevance is complemented or overridden by business rules that
ensure business objectives and constraints are met. A prominent
example of such environments is the merchandising-oriented
search, which can be found in eCommerce, retail, and other
consumer applications, such as hotel booking or restaurant
search services. These applications require search functions that
can improve profits by boosting high-margin products, sell out
expiring inventory, or promote sponsored offerings. Merchan-
dising search should also take into account domain-specific
terminology and usage patterns to recognize the user’s intent
and common idioms.
expert search A large number of search services and applications

in the law, medicine, research, and industry domains fall into the
category of expert search. Used by professionals, an expert search
requires deep understanding of the domain including jargon and
latent relationships between the concepts. It also requires the ac-
commodation of specialized usage patterns and relevance defini-
tions that can be very different from typical consumer needs in
web and merchandising search. For example, it can be critical for
a lawyer or patent professional to find and examine every docu-
ment relevant to the topic in question, whereas merchandising or
web search users are typically interested in a few of the most rel-
evant results. An expert search is also referred to as an enterprise
search in certain contexts.
Despite all these dissimilarities, search solutions for different do-
mains share many common principles and methods. We choose to fo-
cus on merchandising search in this chapter because these are best
aligned with the agenda of this book, although most of the building
blocks that we describe are perfectly applicable in other domains.
182 search
Let us start with a description of a minimal environment for a mer-

chandising search, as depicted in Figure 4.1, that introduces the key
entities and assumptions:
• The main purpose a merchandising search service is to provide
customers with a simple interface that finds items such as prod-
ucts, services, or facilities through a free-text query and addi-
tional contextual information, which can include the customer
profile, browsing device type, geographical location, and so on.
In other words, the input of a search service is assumed to be
made up of the pair of query and context, and the output is
a list of entities in order from the most relevant to the least rele-
vant. This basic functionality can be complemented by additional
search tools and features that we will discuss in detail later on.
Figure 4.1: Example of a merchandising search environment.
• Each item that can be searched for typically represents a mix of

structured or semi-structured data records that can be collected
from multiple sources. For example, an online retailer can assem-
ble its product items from the textual descriptions provided by
manufacturers, user ratings and reviews, sales data, inventory
data, store locations, prices, and in-house metadata, such as cate-
gory hierarchies.
• A search engine produces the search results by matching query

or context features with entity features. This process can be con-
figured by using multiple relevance controls that determine how
features are created from the original data, how they are matched,
and how different signals are blended together to produce the fi-
nal results.
• The search engine and interface application record the interac-

tions with the users. The interaction history typically includes
multiple metrics and parameters ranging from search queries to
page scrolling timings. This information can be fed into the rele-
vance tuning process that adjusts the relevance controls to meet
the objectives.
The environment we defined above is focused exclusively on a free-
text search where a user enters some free text into a form and gets back
a ranked list of results. In a real-world search system such as an eCom-
merce website, this basic functionality is often extended with many
additional tools and capabilities including facets, automatic query com-
pletion, and result sorting options. The core functionality of a search
service almost completely depends upon the notion of relevance, that
is, the differentiation between relevant and non-relevant items. We will
spend most of this chapter discussing different and numerous aspects
of this question. Our first step will be to explore how relevance can be
related to economic objectives and measured.
It is often said that the main objective of a search service is to un-

derstand the user’s intent and deliver results that are relevant to this
intent. Although this is generally true (and already very challenging to
achieve), this view of the objective is limited and does not reflect many
criteria that are important for a good search service. We can attempt to
build a more comprehensive framework by using the basic profit equa-
tion as a starting point. Consider an online retailer who sells a certain
assortment of products, so the total profit can be expressed as
¸
G qj pj vj (4.1)
j
in which j iterates over all products, q is the quantity sold, p is

the price, and v stands for variable costs, which can include whole-
sale price and distribution costs. Although we cannot formally connect
equation 4.1 to an intangible concept such as user intent, we can make
a few heuristic assumptions that help to establish this link and work
well in practice.
First, we can assume that all quantities sold qj are roughly propor-
tional to the organic relevance and overall ergonomics of a search ser-
vice. That is, given that the number of search users is fixed, poor rel-
evance leads to a relatively large fraction of users who cannot find
products that match their purchasing intent, whereas good relevance
leads to a relatively low fraction of users who are not able to find the
184 search
right product. Consequently, relevance can be viewed as a plain multi-

plicative coefficient for the profit.
The second opportunity that becomes apparent from equation 4.1 is
that we can try to redistribute quantities qj to sell more high-margin
products at the expense of the low-margin ones. We can exploit the fact
that purchasing intents typically exhibit some flexibility: a customer
might initially be looking to purchase one product but willing to sub-
stitute it with a different one, or they might simply browse available
offerings with a broadly defined intent and accept the first suitable one.
Given the fixed number of search users, we can promote high-margin
products by boosting them in the search results to capture as many
potential buyers as possible and, consequently, by pushing down those
products with low margins, thereby negatively influencing their sale
volumes. It is important to note that we can account not only for the
margin (the difference between the cost of a product and its selling
price) but also for the costs of alternatives associated with a product.
For example, fashion retailers often practice clearance sales at the end
of a season to free up space for the new collection, so the promotion of
expiring products also becomes an objective.
We already stated that relevance is directly related to the total profit
because it influences the number of conversions. We should, however,
pay more attention to the negative effects associated with relevance
and ergonomics defects. From the user viewpoint, the search process
includes multiple steps, such as making an initial query, browsing and
checking suggested results, query reformatting, and so on. Poor rel-
evance or ergonomics can force users to repeat searches many times,
which may or may not negatively impact conversion rates but is likely
to create a negative user experience and, in the long run, to decrease
the total number of users; this translates into a lower total quantity
sold, according to equation 4.1.
The bottom line is that the objectives of a search service and its qual-
ity can be viewed, at least, from the following three standpoints:
1. Relevance
2. Merchandising controls
3. Ergonomics and customer satisfaction

We discuss each of these three aspects in the subsequent sections and
then delve into the implementation details to see how a search service
can achieve these goals.
4.2.1 Relevance Metrics
The relevance of search results can be defined as a measure of how well

the results satisfy the user’s search intent. The search intent, typically re-
ferred to as the information need in information retrieval texts, cannot be
completely formalized in most applications, especially in merchandis-
ing, so the standard way to measure relevance is an expert assessment
of the search results to classify each item as relevant or irrelevant. For
example, relevant search results for the query sore throat treatments can
be subjectively defined as remedies for this symptom ranging from
hot tea to medications, not merely a set of items containing the words
from the query. We will assume for now that the relevancy grades are
set by experts from the search service team, but, as we will see later
in this chapter, a search system can collect and analyze certain metrics
about user behavior, such as click rates for items in the result set, to
automatically estimate relevancy.
Let us consider an ideal case, when we are assessing a single pair
of a search request and the result, and the total number of items is
small enough to be manually classified. We can define the following
three values: D is the total number of items relevant for a given intent
indexed in the system, S is the number of items in the search result set,
and R is the number of relevant items in the result set. The relationship
between these three values is shown in Figure 4.2. The quality of search
results can then be measured in terms of two metrics, precision and
recall, defined as follows:
precision
R
(4.2)
S
recall
R
(4.3)
D
We typically need both of these metrics to describe the result set or
search method. On the one hand, recall measures the completeness of
the search results regardless of the result set size, so one can always
achieve the maximum possible recall of 1.00 by returning the entire
collection of items. On the other hand, precision measures the density
of relevant items in the result set and tells us nothing about the relevant
items that have not been fetched.
The difference between the two metrics, however, does not mean
that they are independent. First, let us review Figure 4.2 again. It sug-
gests that we can change the recall from 0 to 1 by stretching the search
results rectangle in the vertical direction; meanwhile, the precision re-
mains constant. This behavior is almost never the case for real data.
186 search
relevant results (R) search results (S)
all items
relevant items (D)
Figure 4.2: The relationship between relevant items, search results, and relevant
results.
One of the main reasons is that we describe items and specify queries
by using dimensions that are not necessarily aligned with the shape of
the item set defined by the search intent. Let us consider the example
depicted in Figure 4.3. A retailer has a large collection of shoes that are
described by using properties such as price and category. A user who
searches for affordable quality shoes might consider items scattered along
the diagonal from inexpensive dress shoes to expensive sandals to be
generally relevant. A search system, however, might not be able to re-
peat this shape. If it treats the terms affordable and shoes in a strict way,
it can achieve high precision but relatively low recall by returning, for
example, average running shoes. Loosening of the criteria increases the
recall but also scoops irrelevant items, such as expensive dress shoes,
which decreases the precision1 .
This pattern is almost constantly present in search applications, so
we typically have to choose between high-precision, low-recall search
methods and the low-precision, high-recall alternatives. Merchandis-
ing search is heavily biased towards high-precision search methods be-
cause the primary goal is to provide a user with a reasonable number
of relevant results that can quickly be reviewed.
The basic precision and recall metrics provide a useful conceptual
view of relevance but have many limitations as quantitative measures.
First, precision and recall are set-based metrics that cannot be straight-
forwardly applied to ranked search results. This limitation is critical for
merchandising search that strives to provide the user with a few valu-
able results sorted by relevance. One possible approach to account for
ranking is to go through the items in the result set from top to bottom,
1 This problem is not specific for search and often comes up in machine learning, especially
in deep-learning applications. For example, a set of photographic images is a very “curly”
area in the space of all possible two-dimensional matrices. Such sets embedded into high-
dimensional spaces are referenced as manifolds.
price
expensive
high precision
low recall
cheap low precision

high recall
category
dress shoes running sandals
shoes
Figure 4.3: Precision–recall trade-off.
calculate the precision and recall at each point, and plot a precision–
recall curve. This process is illustrated by the example in Figure 4.4. Let
us assume that we have 20 items in total and 5 of them are relevant.
The result set starts with a relevant item, so the precision is 1.00 and
recall is 1⁄5. The next two items are not relevant, so the recall remains
constant but precision first drops to 1⁄2 and then to 1⁄3. Continuing this
process, we get a precision of 1⁄4 and recall of 1.00 at the twentieth item
in the result set. The precision–recall curve is thus jagged but typically
has a downward-sloping concave trend.
The precision–recall curve provides a convenient way to analyze
search quality for a single query, but we often need a compact metric
that expresses the overall performance of a search service as a single
number. One standard way to do this is to determine the mean average
precision (MAP), which first averages the precisions at each relevant
item and then averages this value over all of the queries we use in the
evaluation. If the number of queries is Q, the number of relevant items
for query q is Rq , and the precision at the k-th relevant item is Pqk ,
then
Rq
1 ¸ 1 ¸
Q
MAP Pqk (4.4)
Q

Rq
q 1 k 1
188 search
precision Rank Precision Recall

1 1.00 0.20
1.00 2 0.50 0.20
3 0.33 0.20
4 0.50 0.40
5 0.60 0.60
6 0.50 0.60
7 0.57 0.80
8 0.50 0.80
9 0.44 0.80
10 0.40 0.80
0.50 11 0.36 0.80
12 0.33 0.80
13 0.31 0.80
14 0.29 0.80
15 0.33 1.00
16 0.31 1.00
17 0.29 1.00
18 0.28 1.00
recall 19 0.26 1.00
0.00 20 0.25 1.00
0.50 1.00
Figure 4.4: Precision–recall curve. The solid circles correspond to relevant items
and the empty circles correspond to irrelevant items in the result set.
For example, the MAP for the single query illustrated in Figure 4.4
is the mean of the five precision numbers for each of the five relevant
items:
MAP (1.00 + 0.50 + 0.60 + 0.57 + 0.33) 0.6

1
(4.5)
5
We typically cannot assess and enumerate all of the relevant results
for a given query in web or merchandising search applications because
the total number of relevant results is huge; even if we can do the
assessment, it does not make sense because no user will be interested
in reading all of the results. Consequently, the average precision above
is often calculated not for all of the items relevant for a query but for
the relevant items in the result set on the basis that the size of the result
set is fixed [Manning et al., 2008].
A popular alternative to MAP is discounted cumulative gain (DCG),
which abandons the notion of precision and recall completely and
gives center stage to ranking [Järvelin and Kekäläinen, 2000]. Let us
consider a search result that contains K items with each item graded
with a relevance value R, so we have K relevance grades Rk . The rele-
vance values can be binary (one for relevant and zero for nonrelevant
items) or continuous or discrete with multiple levels. The cumulative

gain of the result is defined as the sum of the relevance grades:
Ķ
CG Rk (4.6)

k 1
Cumulative gain is similar to precision, given that K is fixed, but it

allows the differentiation of documents based on their usefulness by
setting multilevel grades R. This is the reason why the metric is called
cumulative gain – it attempts to assess the usefulness of the search re-
sult. The shortcoming of cumulative gain is that it does not include
the order of the items in the consideration, so changes in the ordering
of search results do not affect the gain value. This can be fixed by pe-
nalizing for relevant results appearing low in a search results list, that
is, relevance grades are discounted proportionally to their positions in
the list. This leads to the notion of DCG, which uses a logarithm of the
result position as a discounting weight:
Ķ
DCG R1
Rk
log2 pkq
(4.7)

k 2
It is more common, however, to use a slightly different definition

of DCG that strongly emphasizes relevant items by assigning them
exponentially high weights [Burges et al., 2005]:
2Rk 1
Ķ
DCG
log2 pk 1q
(4.8)
k 1
The magnitude of the DCG calculated according to formula 4.8 can
vary depending on the number of results K. To compare DCG metrics
obtained for different queries, we need to normalize them. This can be
done by calculating the maximum possible DCG, called the Ideal DCG,
and dividing the actual DCG by this value to obtain the normalized
DCG (NDCG):
NDCG
DCG
(4.9)
Ideal DCG
The ideal DCG can be estimated by sorting the items in a search
results list by relevance grades and applying formula 4.8 to calculate
the corresponding DCG. Consequently, the NCDG is equal to one for
an ideal ranking. For example, let us consider a search results list with
six items that are scored by an expert on a scale of 0 to 4, with 0
meaning nonrelevant and 4 meaning most relevant:
4, 3, 4, 2, 0, 1 (4.10)
190 search
The value of the DCG calculated for this search result in accordance
with formula 4.8 is 28.56. The ideal ordering for this search result is
4, 4, 3, 2, 1, 0 (4.11)
and the corresponding ideal DCG is 29.64. Consequently, the NDCG

for search results list 4.10 is 28.56/29.64 = 0.96.
4.2.2 Merchandising Controls
Merchandising controls are tools that enable merchandisers and other

business users to reshape search results in accordance with business
needs not covered by organic relevance. The boundary between
merchandising controls and relevance, however, is very blurry because
many business rules can be viewed as enhancements of content-based
relevance and, conversely, many relevance tuning methods can be
viewed as business rules that improve the search results by adding
some domain knowledge. For example, a merchandiser can create a
trigger that redirects all users who enter the query insulated jackets
to a manually curated category for seasonal jacket sales. On the one
hand, this aims to achieve the business goal of promoting products
that are on sale. On the other hand, it can be argued that this manually
curated content is a better match for the user’s search intent than a
standard search result. Most search systems provide a rich toolkit of
merchandising controls that can include the following capabilities:
boost and bury As we will see in the next sections, organic rel-
evance is typically calculated by matching different item and
query properties, mixing the resulting scores together, and rank-
ing the items by the final score. A merchandiser can adjust or
override this logic by tweaking relevancy scores in a way that
promotes desirable items and demotes undesirable ones. This ca-
pability is often referred to in the merchandising world as boost
and bury. The boost and bury control can often be expressed as
a scoring formula that blends different item properties. This ap-
proach can be illustrated with an example that boosts new, dis-
counted, or high-rated products and buries items that lack of
these properties:
score 0.2 newness 0.4 discount 0.4 rating (4.12)
given that each item is attributed with the corresponding new-

ness, discount depth, and user rating grades measured on some
scale. The score calculated this way can override the relevance
score, or the two scores can be summed or multiplied.
filtering The primary purpose of filtering is to remove undesirable

items from a search result set. Examples of filtering include re-
moval of out-of-stock items and removal of irrelevant items that
have cropped up in a search result because of data or scoring
issues.
canned results It can be difficult to achieve a certain ordering

of items by using boost and bury formulas, so a merchandiser
might be willing to put a manually selected set of items at the
top of the search results. The injection of such canned items is
often triggered by certain keywords in a query.
redirection Redirection is a capability similar to canned results,

but it completely replaces organic search results by redirecting
the user to a manually curated category of products or some spe-
cial content, such as an interactive fashion magazine.
product grouping Efficient usage of the display space is an impor-

tant goal for merchandising search. It is important not only to
provide users with relevant search results but also to present the
available assortment in the best possible way, given that display
space is limited. For example, it might be beneficial to replace
closely related products or product variants, such as different
sizes and colors of the same model of jeans, by a single represen-
tative to free up more space for other models and avoid cluttering
the search results with similar items.
From an economic standpoint, some merchandising controls can be
viewed as market segmentation techniques. Consider the example of
boosting high-margin or luxury products: it is essentially an attempt
to segment customers by their price sensitivity, so price-insensitive
customers spend little effort picking high-margin items at the top and
price-sensitive customers have to spend more time going through
pages of results to find a better value price.
4.2.3 Service Quality Metrics
The relevance of search results assessed by an expert and measured

with metrics like the NDCG does not guarantee the acceptable perfor-
mance of a search service. We need to define performance indicators
that can be measured and monitored in real applications to ensure a
high level of user experience and business efficiency. The quality of
a search service can be related to relevancy algorithms, data quality,
ergonomics of the user interface, and robustness of the technical im-
192 search
plementation. Let us review a few key performance indicators that are

often collected for search services:
conversion rate Conversion rate is by far the most important per-
formance indicator for merchandising search. It can be defined
as the ratio of user sessions that used the search service and con-
verted to the total number of user sessions that used the search
service. A user session is typically equivalent to a web session in
this context. The conversion rate is a critically important metric
because it is directly related to revenues and the user’s ability to
find the desired products.
click-through rate The ratio of users who clicked on specific

search results to the total number of users who used the search
is an important metric for the relevance of results.
time on a product detail page Although a high click-through

rate is a generally positive indicator, a large number of users
quickly glancing at the details page and then returning back may
indicate poor relevance or problems with interface ergonomics,
such that a user is not able to recognize relevant items by their
summaries in the search results list.
query modification rate There is a good chance that a user

who modifies a query multiple times is unable to get satisfactory
search results.
paging rate Frequent usage of pagination and clicks on low-ranked

search results may indicate relevance problems.
retention rate The fraction of users who continue to regularly use

the search. Retention rate is typically calculated for some period
of time, such as a week or month, as
EN
retention rate (4.13)
S
in which E is the number of regular users at the end of a period,
N is the number of new users acquired during that period, and
S is the number of users at the start of a period.
search latency The time it takes to process a search query and re-
turn a result has a big impact on the user experience. Many retail
and web search companies have reported impressive statistics on
this matter. For instance, Amazon reported that every 100 mil-
lisecond increase in the page load time results in a 1% loss in
4.3 building blocks: matching and ranking 193
sales, and Walmart showed that every second of latency improve-

ment increases the conversion by 2% [Kohavi and Longbotham,
2007; Crocker et al., 2012].
These metrics can be broken down by dimensions, such as marketing
channel (e.g., mobile, desktop, or tablet), for relevance fine tuning. We
will return to the question of relevance tuning, including both manual
and automatic tuning processes, in section 4.7, but we first need to
delve into the details of relevance scoring.
4.3 building blocks: matching and ranking
The search relevance problem can be considered as a classification

problem because it aims to distinguish relevant items from nonrelevant
ones. At the same time, it is a very special case of classification because
of its focus on textual data and ranking. These features enable very ef-
ficient heuristic methods that can achieve excellent relevance without
training a classifier by using machine learning techniques. Although
the machine learning approach is also possible (and we will discuss
it in the next sections), the basic matching and ranking methods are
sufficient for most merchandising search applications and also provide
a solid methodology for feature engineering in cases when machine
learning methods are applied. We will spend this section reviewing
basic search methods that can later be assembled into more compli-
cated and comprehensive relevance solutions that may or may not use
machine learning for parameter tuning.
At a very high-level, we can describe a search as a calculation of
some similarity metric between an item and a query, so all items can
be ranked according to this metric or excluded from the search result
set. Similarly to other classification problems, this requires the repre-
sentation of an item, a query, and, optionally, other pieces of contextual
information, such as user profile details, as features and then the calcu-
lation of one or more scores, which we will call signals hereafter, that
indicate how well the item features match the query features. The sig-
nals are later combined together to make a final decision about whether
an item should be included in the result list (matching) and about its
position in the list (ranking). Figure 4.5 illustrates this flow.
Consequently, the designer of a search service has to address multi-
ple controls at each of these stages: how to engineer features out of the
raw data, how to match the features to produce signals, and how to
mix signals together to achieve good relevance. Let us start to consider
these controls with a few basic techniques and then gradually increase
the complexity by chaining multiple blocks together and adding more
variables into the equation.
194 search
Figure 4.5: A high-level view of a search flow and its main controls.
4.3.1 Token Matching
In a merchandising search, items are typically quite complex entities

with multiple textual and numerical attributes, such as name, descrip-
tion, price, and brand. The data can also contain a significant amount
of structural information, such as a category hierarchy and different
size/color product variants associated with one logical product. Let us
put this complexity aside for a moment and consider the simple case
of a retailer who represents each product as a document with a sin-
gle description field that contains plain text, so an example with two
products would be as follows:
Product 1
Description: Pleated black dress. Lightweight look
for the office.
Product 2
Description: Fiery red dress. A black ribbon
at the waist.
The most basic thing we can do to search over such documents is to
break the descriptions up into words and allow only for single-word
search queries, so a product will be included in a result set only if
its description contains the same word as the query. The process of
breaking up a text into words or other elements such as phrases is
called tokenization, and the outputs – words in our cases – are called
tokens. For English language, tokenization is usually done by using
spaces and punctuation marks as delimiters, so the documents above
will produce the following tokens:
Product 1: [Pleated], [black], [dress], [Lightweight],
[look], [for], [the], [office]
Product 2: [Fiery], [red], [dress], [A], [black]
(4.14)
[ribbon], [at], [the], [waist]
Consequently, the query black will match both products and the
query red will only match the second product. Clearly, this method
provides only a matching capability and the items in the result set are
not ranked.
Although this token matching method is extremely simple and lim-
ited, it illustrates the main search controls that we discussed earlier.
First, each token can be viewed as an individual feature that may or
may not be present in a product. The tokenization process is then an
example of feature engineering. Words are indeed reasonably good fea-
tures because they carry a strong signal about a product type, such as
shoes, and its properties, such as a black color. Second, token match-
ing is a way to produce signals about correlations between product
features and query features. Finally, the signals from all tokens are
combined together to produce the final decision on match or mismatch.
This flow is visualized in Figure 4.6.
Figure 4.6: An illustration of a token matching flow that includes tokenization,

scoring, and signal mixing. The product description and its tokens
are denoted as d and ti , respectively. The query is denoted as q.
4.3.2 Boolean Search and Phrase Search
The main shortcoming of basic token matching is an inability to pro-

cess more meaningful queries than just a single term. The first enhance-
ment that helps us to work around this limitation is the Boolean query,
which enables chaining of multiple tokens by using Boolean opera-
tions, namely, AND, OR, and NOT. For example, the following Boolean
query
dress AND red
will match only the second product from example 4.14, whereas the
query
dress AND (red OR black)

196 search
will match both products. Boolean queries do not take into account
positions of tokens in the text and, hence, can be thought of as several
token matching queries combined together.
The second important capability that extends basic token matching is
the phrase query. A phrase query is a query that searches for documents
that contain a sequence of tokens that follow one another, as opposed
to a Boolean query, which searches for documents that contain individ-
ual tokens irrespective of their order and positions in the text. We use
square brackets to denote phrase queries and subqueries. For instance,
the following query will match the first product from example 4.14 but
not the second one:
[black dress]
This result has higher precision and lower recall than the result of the
Boolean query black AND dress, which matches both products. Boolean
and phrase queries together provide very powerful tools to control
relevancy and manage the precision–recall trade-off. A query language
that directly supports Boolean expressions is often a good solution for
expert search where users are willing to learn and use advanced search
functions, but its usage in merchandising search is limited because of
the unintuitive user experience. We will discuss how free-text search
can take advantage of complex Boolean and phrase queries in the later
sections.
4.3.3 Normalization and Stemming
It is not difficult to see that chopping a text into tokens produces re-
sults that are not optimal from a matching standpoint. In natural lan-
guage, words may have different forms and spellings that can be con-
sidered indistinguishable for almost all search intents. Some words do
not carry any meaningful information at all and generate noise signals.
This suggests that we need to perform a normalization of the raw to-
kens to create a cleaner token vocabulary. Such normalized tokens are
typically referred to as terms.
Normalization is a complex process that usually includes multiple
steps to address the different properties and phenomena of natural
language. Let us go through an example that illustrates the key trans-
formations by starting from the following original product description:
Maison Kitsuné Men’s Slim Jeans. These premium

jeans come in a slim fit for a fashionable look.
The first step is to normalize the character set because a search query
can be entered with or without diacritics and this difference typically
does not mean different search intents. Tokenizing the text and con-
verting it to a standard character set, we get the tokens
[Maison] [Kitsune] [Men’s] [Slim] [Jeans] [These]
[premium] [jeans] [come] [in] [a] [slim] [fit]
[for] [a] [fashionable] [look]
The second issue that we face is the presence of lower- and uppercase
characters that are also indistinguishable in most cases. The standard
approach is to transform each token into its lowercase form, which
gives the following result for our example:
[maison] [kitsune] [men’s] [slim] [jeans] [these]
[premium] [jeans] [come] [in] [a] [slim] [fit]
[for] [a] [fashionable] [look]
The third possible step is to exclude high-frequency tokens such as

and, to, the, and will because they are present in most texts and, hence,
do not carry any specific information about an item. Such tokens are
commonly known as stop words. Applying this transformation to the
example, we get
[maison] [kitsune] [men’s] [slim] [jeans] [premium]
[jeans] [come] [slim] [fit] [fashionable] [look]
The exclusion of stop words may have both positive and negative
implications. On the one hand, it can positively influence some match-
ing and ranking methods that we discuss later because meaningless
high-frequency terms can skew some metrics used in relevancy calcu-
lations. On the other hand, the removal of stop words can result in
losses of substantial information and reduce our ability to search for
certain queries. For example, the removal of stop words prevents us
from finding phrases like to be, or not to be or distinguishing new from
not new. Stop words can also destroy the semantic relationships be-
tween the entities, so it becomes impossible to distinguish between an
object on the table and under the table.
The fourth standard normalization technique is stemming. In most
natural languages, words can change their form depending on number
(dress and dresses), tense (look and looked), possession (men and men’s),
and other factors. Stemming is a process of reducing different word
forms to the same root in order to eliminate differences that typically
relate to the same search intent. The problem of stemming is challeng-
ing because of the multiple exceptions and special cases that can be
found in natural languages. There exist multiple stemming methods, ei-
ther rule based or dictionary based, with different strengths and weak-
nesses. One popular family of rule-based stemmers is based on the
so-called Porter stemmer [Porter, 1980]. This represents a few groups
of suffix transformation rules and conditions that check that a word is
198 search
long enough to be reduced (see table 4.1 for an example). By applying

stemming to our sample product description, we obtain the final set
of regular terms, which looks more condensed and focused on the key
features than the original text:
[maison] [kitsun] [men] [slim] [jean] [premium]
[jean] [com] [slim] [fit] [fashion] [look]
Rule Example
...ational − Ñ ate Ñ relate

relational −
...tional − Ñ tion Ñ condition

conditional −
rational −Ñ rational
...ful −Ñ none hopeful −Ñ hope
...ness −Ñ none goodness −Ñ good
...izer −Ñ ize digitizer −Ñ digitize
Table 4.1: An example of the rules used in the Porter stemmer. All rules in this
example require at least one switch from vowel to consonant in front
of the suffix, so the second rule applies to the word conditional but not
to rational.
The same set of normalization algorithms is typically applied to

both the query and documents to map all tokens into the same space
of terms. For example, the query Fashionable will match a product
that contains fashioned in the description because both words will be
mapped to the term fashion.
4.3.4 Ranking and the Vector Space Model
Matching by using Boolean queries and phrase queries enables us to

find a set of items that meet search criteria. However, the number of
matching and potentially relevant items often exceeds the relatively
small number of results that the average merchandising search user
will look through, so the order in which items are presented to the
user becomes critically important. We need to define a building block
that can rank items according to their relevance.
Although ranking cannot improve the global precision–recall prop-
erties of the underlying matching, it can be considered a trick that
improves precision–recall in the sense of local or perceived qualities.
On the one hand, ranking increases the precision of the top results by
boosting relevant items, but, at the same time, it does not remove items
from the search results list, so it provides the same recall.
We can take a first step towards ranking by taking a closer look at

Boolean queries and recognizing their scoring potential. First, we can
notice that both item documents and queries can be represented as
binary vectors, such that each element of a vector indicates whether a
document or a query contains a certain term or not. In other words, an
element that corresponds to a certain term is equal to one if a document
or query contains the term and zero otherwise. If the total number of
distinct terms in all documents in a collection is n, each document or
query is a binary vector with n elements. It is easy to see that a Boolean
query can be expressed by using a dot product between query vector q
and document vector d. Recall that a dot product between two vectors
is given by
ņ
dq q i di (4.15)

i 1
and the Euclidean norm of a vector is defined as

b
kdk d21 d2n (4.16)
Consequently, we can say that a Boolean query that contains multiple

terms chained by using the AND operator is equivalent to the following
condition:
d q ¥ kqk2 (4.17)
because all terms in the query must match corresponding terms in

the document, which results in a dot product that equals the number
of one elements in the query vector. A Boolean query that chains its
terms with OR is equivalent to the condition
dq ¥ 1 (4.18)
because at least one term must match. This interpretation of Boolean

queries reveals a kind of internal scoring that is converted into a match-
ing decision by using a threshold. Equations 4.17 and 4.18 also suggest
that the ratio between the dot product and query norm can be used as
a continuous measure of similarity between the document and query.
We can go even further and ask why we don’t account for the norm of
a document as well. One can argue that a short document that matches
query terms is more relevant than a long document that matches the
same number of terms. This can be justified from a probabilistic stand-
point in the following way. Consider a person who talks about some
topic – if that person uses a lot of relevant words in the first minute
of the speech, it can be an indicator that the speech is really focused
on the relevant topic, whereas the appearance of the same words in
200 search
a one-hour speech can be attributed to a broadness of topics rather

than focus. A metric that normalizes the dot product by norms of both
vectors is a cosine similarity:
qd
cospq, dq
kqk kdk
(4.19)
The cosine similarity, that is, a cosine of the angle between the vec-
tors, is a convenient metric that ranges from zero to one for positively
defined vectors. A cosine similarity of zero means that a document
vector is orthogonal to a query vector in the space of terms, and a simi-
larity value of one means an exact match equivalent to a Boolean query.
Unlike a Boolean query, the cosine similarity does not require opera-
tions to be specified in a query – it treats both query and document
as an unordered collection of terms. Let us illustrate this vector space
model by using an example.
example 4.1
Consider two items that have the following descriptions (for simplicity,
we assume that the descriptions have been tokenized and normalized):
Product 1: dark blue jeans blue denim fabric
Product 2: skinny jeans in bright blue
These two descriptions and the query dark jeans are represented as
binary vectors in table 4.2.
dark blue jeans denim fabric skinny in bright

kk
?5
d1 1 1 1 1 1 0 0 0
?5
d2
q
0
1
1
0
1
1
0
0
0
0
1
0
1
0
1
0
?2
Table 4.2: An example of two documents and one query represented as binary
vectors.
The similarity values between the query and each of the documents
will be
cos pq, d1 q ? ? 0.632
1+1
2 5
(4.20)
cos pq, d2 q ? ? 0.316
1
2 5
Figure 4.7 shows the relationship between the documents and query
in a vector space. Note that the cosine similarity can be evaluated effi-
ciently because only the non-zero dimensions of the query have to be
considered.
N
dark
1.0
d₁ q
)
, d₁
(q
s
co
cos(q, d₂) jeans
1.0
other
terms
d₂
Figure 4.7: An example of the vector space model and cosine similarity for two
documents and one query. The document and query vectors are de-
picted as normalized.
4.3.5 TFIDF Scoring Model
The vector space model with binary vectors has two important short-
comings that negatively impact the relevance of results ranked by us-
ing this method. First, it does not take into account the term frequency
in a document. We can expect that documents that have multiple oc-
currences of query terms are more relevant than documents where
the same terms occur only once. Second, some terms can be more im-
portant than others: rarely used words are often more discriminative
and informative than frequently used words. For example, an apparel
retailer may have word clothing in most product descriptions so match-
ing this term does not signal strong relevance. The stop words that we
discussed earlier are an extreme case of this problem.
The first issue described above can be mitigated by replacing the
zeros and ones in the document vectors with the corresponding term
frequencies in the document. This variant of the vector space model is
often called the bag-of-words model. Term frequency (TF) can be defined as
the number of occurrences of term t in document d, which we denote
202 search
as npt, dq or some non-linear function of this number. One popular

choice is to use the square root of the number of occurrences:
a
tfpt, dq npt, dq (4.21)
For instance, a term that occurs in a document nine times will have
a term frequency of three. The square root function is used to smooth
down scores of documents that have a very high number of term oc-
currences.
The second issue can be addressed by calculating the term frequen-
cies across the entire collection of documents to distinguish rare words
from frequent ones. One possible way to estimate word rarity is to
count the occurrences of a term in all of the documents, similarly to the
method for term frequency but for the entire collection. This approach,
however, is known to produce non-optimal results because a few doc-
uments with multitudinous occurrences of a rare term can skew the
results. A more common method is to count the number of documents
that contain a given term at least once. This metric is known as the
document frequency of a term. The inverse of the document frequency
can then be used as a measure of term rarity. The standard formula for
the inverse document frequency (IDF) for term t is as follows:
idfptq 1
N
dfptq
ln (4.22)
1
in which N is the total number of documents in the collection, and dfptq
is the document frequency of the term. Similarly to term frequency,
a logarithm function is used to smooth down the magnitude of the
coefficient for rare terms.
Term frequency and inverse document frequency are usually com-
bined together so that the elements of a document vector are calcu-
lated as a product of values defined by equations 4.21 and 4.22 for the
corresponding term:
dpiq tfpti , dq idfpti q (4.23)
This is a widely used approach known as the TFIDF model. Sub-
stituting expression 4.23 into the definitions of the dot product and
Euclidean norm, we get the following formulas that can be used to cal-
culate the cosine similarity score for query q and document d under
the TFIDF model:
¸
qd tfpt, dq idfptq tfpt, qq idfptq (4.24)
t in q
d ¸
kqk rtfpt, qq idfptqs2 (4.25)
t in q
d ¸
kdk rtfpt, dq idfptqs2 (4.26)
t in d
The cosine similarity calculated by using formulas 4.24–4.26 is one

of the most commonly used scoring methods. These formulas, how-
ever, are not a rock-solid standard, and many different variations can
be found in real search engine implementations. Let us consider one
particular example that makes the following three adjustments, which
are known to work well in practice:
1. The document norm defined by equation 4.26 normalizes all doc-
ument vectors to a unit length. In many practical applications,
however, shorter documents are often more relevant if they con-
tain the same number of term matches and equal term frequen-
cies. This can be accounted for by replacing the standard docu-
ment norm with a norm that is proportional to the total number
of terms npdq in a document:
d ¸
a
Ld pdq 1 npdq (4.27)
t in d
2. All terms in a query can be considered equally significant and

processed independently, even in the case of duplicates. Conse-
quently, the term frequency tfpt, qq is always equal to one. This
allows us to redefine the query norm as follows:
d ¸
Lq pqq idfptq2 (4.28)
t in q
3. The score of a document in the TFIDF model depends on the

number of words that match the query because missed words
zero out the corresponding terms in the dot product. It can be
argued that missed words should be penalized even more, so
an additional coefficient called a coordination factor can be in-
troduced. Coordination factor cpq, dq is defined as the ratio of the
number of common terms in a query and document to the total
number of terms in the query. For example, the query black skinny
jeans and the document black jeans will have a coordination factor
of two thirds.
Collecting these adjustments together and substituting them into the
definition of cosine similarity, we obtain the following final scoring
formula:
cpq, dq ¸
scorepq, dq tfpt, dq idfptq2
Ld pdq Lq pqq (4.29)
t in q
204 search
The TFIDF scoring model is a fundamental building block that we

will use later to assemble more sophisticated scoring solutions. Nev-
ertheless, it is worth noting that this method has been designed as a
generic solution for the retrieval of relatively large texts, such as jour-
nal articles, and its usage in a merchandising search that deals with
structured data can have pitfalls that sometimes require the use of al-
ternative ranking methods, as we will see later.
example 4.2
We will wrap up the discussion of ranking methods with an example

of TFIDF calculations. Let us take the following product descriptions:
d1 : dark blue jeans blue denim fabric

d2 : skinny jeans in bright blue
Applying formulas 4.21 and 4.22, we obtain the TF and IDF values
summarized in table 4.3. Let us now score these products for the query
skinny jeans. The query norm and document norms can then be calcu-
lated in accordance with formulas 4.27 and 4.28 by using the TF and
IDF values that we just evaluated.
? ?5 2.236
Ld pd1 q 6 2.449, Ld pd2 q (4.30)
b
Lq pqq idfpjeansq2 idfpskinnyq2 1.163 (4.31)
dark blue jeans denim fabric skinny in bright
idfpq 1.00 0.59 0.59 1.00 1.00 1.00 1.00 1.00

p q
tf , 1 1.00 1.41 1.00 1.00 1.00 0.00 0.00 0.00
p q
tf , 2 0.00 1.00 1.00 0.00 0.00 1.00 1.00 1.00
d1 1.00 0.83 0.59 1.00 1.00 0.00 0.00 0.00
d2 0.00 0.59 0.59 0.00 0.00 1.00 1.00 1.00
Table 4.3: An example of TF and IDF calculations for two documents. The last

two lines correspond to the TF IDF vector representations of the doc-
uments.
The coordination factor is 0.50 for the first product and 1.00 for the
second one. By substituting all norms and TF/IDF values into for-
mula 4.29, we get a score of 0.062 for the first product and a much
higher score of 0.520 for the second product, which is in agreement
with the intuitive expectation that the second product is more relevant
for the query that we used.
TFIDF is also dependent on text normalization and stemming. For

instance, it is not difficult to see that the following documents have
equal TFIDF scores for the query dark if no stemming is performed:
d1 : dark darker darkness
d2 : dark darker lightness
d3 : dark light lightness
The first document, however, looks much more relevant in this con-
text. Stemming will map the words dark, darker, and darkness to the
same root dark, which will result in higher scores for the first and sec-
ond documents because of higher term frequency. Moreover, a user
who searches for darkish shoes will get no results without stemming,
which is unlikely to be a good user experience.
N
4.3.6 Scoring with n-grams
We have shown that the vector space model is related to Boolean

queries and TFIDF scoring can be considered as a soft-output
version of a Boolean query that fills the gap between OR-chained and
AND-chained Boolean queries. This line of thinking can be continued,
and a soft version of a phrase query can be created. From the ranking
perspective, the drawback of standard Boolean phrase queries is that
they are too restrictive and require all query terms to match. One
possible relaxation is to match not the entire phrase but shingles, that
is, sequences of several terms. Such sequences are also called n-grams
and can include two terms (bigrams), three terms (trigrams), or more.
Shingles can be viewed as a tokenization technique that is applied to
both document and query, and the resulting n-grams are fed into a
Boolean query or TFIDF scorer. The following example shows how
two products can be tokenized into bigrams:
black cotton polo shirt: [black cotton]
[cotton polo]
[polo shirt]
short sleeve black shirt: [short sleeve]

[sleeve black]
[black shirt]
The TFIDF scorer treats n-grams just like single-word terms and
calculates the cosine similarity in the vector space where each vector
element corresponds to a shingle, and the TFIDF metrics are also cal-
culated for shingles. So these products produce equal TFIDF scores
for the query black shirt if tokenized into single words (unigrams), but
the second product scores higher if bigrams are used because it ex-
plictly contains the black shirt subphrase. It can be argued that scoring
206 search
using bigrams better captures semantic relationships in the text: the

proximity of the words polo and shirt in the first product emphasizes
that polo shirt is the primary property and black is just a refinment, but
the proximity of black and shirt in the second product description in-
dicates that black is the key property. This ability to capture semantic
relationships is especially important to differeniate between compound
terms such as tuxedo coat and sports coat. The use of shingles is a pow-
erful method to increase search precision that is often combined with
the standard single-word scoring, as we will discuss shortly.
4.4 mixing relevance signals
Thus far, we have discussed how to search for items represented as

plain textual descriptions. In merchandising search, as well as most
other search applications, such a plain data format is rarely the case.
We almost always deal with structured source data that characterize
each item by multiple properties:
Name: Levi’s Hooded Military Jacket
Description: Stand collar with drawstring hood
Brand: Levi Strauss
...
Price: 189.90
Category: Women’s Jackets
Items can also have dynamic properties, such as sales data and user
ratings, that also carry important information about their fitness and,
ultimately, relevance. Item property values can be short strings such
as product names, long text snippets such as descriptions or reviews,
numbers, tokens from a discrete set such brand names, or even nested
or hierarchical entities such as product variants or categories. This cre-
ates a diversity of features and signals that are measured on different
scales and may not be directly comparable. We need to find a way to
correlate all of these features with a query and mix the resulting signals
together to produce a relevance score.
One naïve approach to this problem is to blend all of the property val-
ues into one large text and use basic scoring methods to search through
this text. Although this approach is not totally meaningless, it results
in a very smooth and blurry signal that unpredictably scores search
results based on the interplay of term frequencies and text lengths. For
instance, the seemingly simple query black dress shoes can result in a
wild mix of dresses, shoes, black tuxedos, and other items that happen
to have some of the query terms in the description. To manage this
problem, we have to create a method that preserves the focused fea-
tures and signals and provides enough controls to pick the strongest
and most relevant results.
4.4 mixing relevance signals 207
4.4.1 Searching Multiple Fields
A query can be considered as a description of a desired search result

that specifies one or more properties an item should have to be relevant.
For example, the query black levi strauss jeans clearly requests products
of type jeans, color black, and brand levi strauss. On the other hand, each
item is also represented as a set of properties, so we can expect to get
a good result by creating documents with multiple fields where each
field corresponds to an item property, running the query against each
field in isolation to get multiple signals, and then blending the signals
into the final score. This idea is illustrated in Figure 4.8.
Figure 4.8: A basic schema of multifield scoring. F stands for splitting a doc-
ument into fields f1 , f2 , . . . , fn . S is a signal mixing function that
produces the final score.
This approach can produce reasonably good results, but it requires

us to be careful about keeping signals focused and balanced. Problems
with signal focusing can arise in connection with mismatches of the
document fields and concepts in a user query, so the concept that a
user searches for can be scattered across multiple fields. For example, a
user can search for a person by using the first and last name, expecting
that this pair will be interpreted as a single token, but the documents
can store first and last names in different fields, thereby producing
meaningless partial matches. The problem with signal balance appears
because each field exists in its own universe and there is not a common
scale for signal scores (e. g., 0.00 means nonrelevant and 1.00 means
relevant).
208 search
example 4.3
Let us illustrate the problem of signal disbalance by using the example

of a fashion retailer catalog that contains a thousand products compris-
ing jeans and shoes:
Product 1
Name: Men’s 514 Straight-Fit Jeans.
Description: Dark blue jeans. Blue denim fabric.
Brand: Levi Strauss
....
Product 1000
Name: Leather Oxfords.
Description: Elegant blue dress shoes.
Brand: Out Of The Blue
We can expect that it is quite usual for jeans to have the term blue in
the name or description, so its IDF value will be quite low. For instance,
if we have 500 blue jeans out of a thousand products, we get
idfpname:blueq 1 lnp1000{501q 1.69 (4.32)
At the same time, the brand Out Of The Blue that manufactures blue
shoes may be very rare. Let us assume that we have only one product
of this brand and there are no other brand names containing the word
blue, so the IDF value for the term blue in the brand field is
idfpbrand:blueq 1 lnp1000{2q 7.21 (4.33)
Consider the query blue jeans now. The blue shoes will have a very
high score for the brand field and a low score for the description field,
whereas the blue jeans will have a relatively low score for the descrip-
tion field that matches both terms of the query but with low IDFs.
Combining the signals by using a sum or maximum function, we are
likely to get a search results list with the shoes at the top, which does
not match the search intent. The reason is that IDFs depend on term
distributions within a field; therefore, IDFs for different fields are not
comparable.
N
One possible solution for signal disbalance is to adjust signals by us-

ing manually set weights, as shown in Figure 4.8. In the example above,
we can assign a low weight to the brand field to demote its signal and
push the shoes down. This can help in cases of consistent differences
in signal levels or importance (e. g., we know that name matching is
more important than description matching), but this solution is very
brittle. We need to approach the problem of signal equalization in a
more systematic way.
4.4.2 Signal Engineering and Equalization
Multifield search has two aspects that complement each other and typ-
ically need to be solved in parallel – signal engineering and signal
equalization. Signal engineering aims to create clear and focused sig-
nals; meanwhile, signal equalization aims to mix these signals together
to produce the final results. The same relevance problem sometimes
can be solved in different ways, either by tuning the mixing function or
by constructing a better more accurate signal. When we are searching
through multiple fields, the following types of relationship between the
fields and the search intent can be distinguished [Gormley and Tong,
2015]:
• One strong signal. It can be the case that a user searches for a
certain property that ideally should match with one of the fields
and produce a single strong signal. Signals from different fields
do not complement each other but rather compete. For instance,
a user who searches for the brand Out Of The Blue is likely to be
focused on the brand field and does not consider the color blue
to be relevant.
• Strong average signal. We can be interested in taking the average

signal rather than the strongest one if the individual signals are
balanced and related to different aspects of the same search in-
tent. For example, item size and color can be equally important.
• Fragmented features and signals. We can expect to get a clear signal

by scoring individual fields only if a query and field are aligned
and resonate in a meaningful way. It can be the case, however,
that fields contain fragmented pieces of information, such that
the resulting signals do not correlate with relevance. Such frag-
ments can be merged together to obtain a better signal.
These three cases are, however, closely related. Let us go through
the list and discuss signal engineering and equalization techniques for
each case in detail.
4.4.2.1 One Strong Signal

The problem with the Out Of The Blue brand that we discussed in the
previous section arises from inaccurate processing of the brand signal.
We noted that one possible solution is to adjust the weight of the signal,
but another alternative is to refine the brand signal to make it less
ambiguous. We can argue that a brand name is a concept that cannot
be broken down into separate words, so it makes sense to replace the
TFIDF scoring with bigram scoring. This will light up a signal only
210 search
if a recognizable part of the brand name is present in a query. We can

also choose to use a maximum function for signal mixing to select the
strongest signal, so we end up with the scoring pipeline presented in
Figure 4.9.
Figure 4.9: An example of a signal mixing pipeline that focuses on the strongest
signal.
A search result produced by pipeline 4.9 is shown in Figure 4.10. The

items marketed under the Out Of The Blue brand outstrip other items
only if the brand name is clearly articulated in the query; otherwise,
items with relevant descriptions and names will take precedence. Note
that we might need to make more adjustments to make the bigram
scoring work properly, for example, disable stop words.
search results
1 q: blue
2
yes 3 name: blue
description: blue
brand: blue
brand match?
q: out of the blue
no brand: out of the blue
name: blue
description: blue
Figure 4.10: A search result structure for the strongest signal strategy.
The strategy of signal mixing with a maximum function is a power-

ful and popular approach to multifield search, and the use of n-grams
is an efficient signal focusing technique. The creation of focused sig-
nals, however, is a complex process that is not limited to n-grams. It
is not unusual to generate several signals from the same field, such as
parallel unigram and bigram scores that can be mixed together, or to
request merchandisers to attribute items with a completely new prop-
erty that helps to produce a more focused signal than the available
features.
The search result structure in Figure 4.10 is quite basic, and we can
program more complex behavior by relaxing the signal mixing func-
tion. One possible solution is to mix weaker signals with the strongest
one in a controllable way. This can be expressed by using the following
signal mixing formula:
¸
s sm α si (4.34)

i m
in which sm is the maximum (strongest) signal, 0 ¤ α ¤ 1 is a pa-

rameter that controls the weight of all other signals si in the mix, and s
is the final score. It is easy to see that formula 4.34 provides a spectrum
of scoring functions that starts with the strongest signal selection if α is
zero and ends with signal averaging if α is set to one. This approach al-
lows us to achieve result structures with more than two relevance tiers.
For instance, we can prioritize products with a name that matches the
query, but we can also keep brand matching as a second priority, as il-
lustrated in Figure 4.11. This can be implemented by using the scoring
function 4.34 and setting weights such that the product name signal
is amplified and the matching items are elevated to the top tier of the
search results list. Brand matching will be the second-strongest signal
in the mix, so the items in the inner tiers created by name matching
will be ranked based on the brand. A signal mixing pipeline that im-
plements this strategy is shown in Figure 4.12.
search results
1
yes 2
3
brand match? inner tiers
yes no
name match?
no outer tiers
Figure 4.11: A search result structure for weighted signal mixing.

212 search
Figure 4.12: An example of a pipeline for weighted signal mixing.
4.4.2.2 Strong Average Signal

By increasing the magnitude of the secondary signals in formula 4.34,
we arrive at a solution that ranks items based on the average signal.
The sum of signals also produces a kind of tiered search results, al-
though the tiers are based on the number of search criteria met by
an item, rather than on individual fields. A search query can include
several criteria that correspond to different item properties and, con-
sequently, document fields. Summing the signals from different fields,
we achieve a ranking behavior that places items with many resonating
properties at the top of the search results list, followed by items with
fewer properties that resonate with the query. We have already seen
that this approach is vulnerable to signal disbalance, but it can be a suit-
able solution for mixing a group of equalized signals. A common use
for this method is with signals obtained from different versions of the
same property. For example, we can use stemming for a certain field
to improve recall and, at the same time, use the original non-stemmed
version of the same field to boost exact matches. The following product
descriptions contain the words fashion and fashionable
d1 : new popular fashion brand

d2 : stylish and fashionable look
which are reduced to the term fashion by a Porter stemmer. Conse-

quently, both descriptions have identical TFIDF scores for the query
fashionable. We can boost the second document that has a word that ex-
actly matches the query by scoring non-stemmed versions and adding
up the stemmed and non-stemmed scores. Summing of the signals is
a feasible strategy here because the more matches we have, the better
the result.
4.4.2.3 Fragmented Features and Signals

Finally, we have to discuss in detail the case of fragmented features,
which requires its own signal engineering techniques. The problem of
fragmented features and signals arises because the standard multifield
search we discussed above scores each field independently. Although
it may not be obvious at first glance, it can be the case that individual
fields correlate reasonably well with a query to produce a strong av-
erage signal but the overall coverage of query conditions by all fields
remains low. Let us consider the following basic example with two
product documents [Turnbull and Berryman, 2016]:
Product 1
Name: Polo
Brand: Polo
Product 2
Name: Polo
Brand: Lacoste
It is quite intuitive that the second product is more relevant for the
query Lacoste Polo, but TFIDF calculations give a different result. Re-
call that the practical TFIDF scoring formula 4.29 for one-term fields
boils down to the following equation:
query coordination factor
query norm field norm
tfpterm, fieldq idf2field ptermq (4.35)
The query coordination factor is 0.50 for all four fields because only
one of the query terms matches (Polo or Lacoste). The query norm, field
norm, and TF value are also the same for all terms and fields. The IDF
values for Polo and Lacoste are the same for the brand field, but different
from the Polo IDF in the name field. Consequently, the name and brand
fields (pairwise) have the same scores in both documents and the total
document scores are equal as well; it does not matter which function
we use to mix the query signals, sum or maximum. The fundamental
reason is that each field matches exactly one query term (either Polo or
Lacoste) to produce equally strong signals, but the fact that the second
document as a whole covers two terms and the first document covers
only one term is not taken into account. This issue can arise in many
cases where different facets of the same logical property are modeled
as different fields: a person name can be broken down into first and last
names, a delivery address can be split into street name, city name, and
country, and so on. Fragmented signals can lead to frustrating search
results – a document that perfectly matches a query can be present in
the search results list but may have a surprisingly low rank.
One possible way to address the problem is to merge several similar
fields into one, thereby eliminating the problem with fragmented or
214 search
imbalanced signals. This is a valid practical method that can improve

relevance. The downside of this method is that the resulting signal can
be too blurry, in the sense that secondary features become as important
as the primary ones. For example, a dress can have the phrase wear with
any shoes in the description, which can make the dress pop up in the
search results for shoes if the product-type feature is not differentiated
properly.
The second alternative for the fragmented feature problem is based
on the observation that single-term queries are not vulnerable to the
fragmentation described above. Moreover, each term in a query can be
considered as a separate criterion added by the user to narrow down
the search result, so it can be reasonable to score a document for each
query term independently, producing a signal that indicates how well
this particular criterion is satisfied, and then blend all the signals to
produce the final score. This approach is called term-centric scoring to
differentiate it from the field-centric approach that we used earlier. The
term-centric signal pipeline can be thought of as multiple field-centric
pipelines executed for every query term to produce signals that are
finally mixed into the final output, as shown in Figure 4.13. The sig-
nals from different pipelines are summed up because the more terms
that match, the better (the average signal strategy); meanwhile, the
strongest signal strategy can be applied to signals from the different
fields within a term pipeline.
Figure 4.13: Term-centric scoring pipeline. t1 and t2 are query terms; f1 and
f2 are document fields.
Returning to our example with the Polo and Lacoste brands, we can
notice that the term-centric approach will produce more meaningful
results for the query Lacoste Polo. The first document will get a high
score for the query term Polo from both fields and zero scores for the
term Lacoste, so the total score will be
max ttf-idfname pPoloq, tf-idfbrand pPoloqu max t0, 0u (4.36)
in which the TFIDF score for Polo is 1.00 for the name field and
0.35 for the brand field due to the difference in IDF. At the same time,
the second document will get a high score for both query terms:
max ttf-idfname pPoloq, 0u max t0, tf-idfbrand pLacostequ (4.37)
in which the TFIDF score for Lacoste is 1.00 for the brand field.
This result looks more relevant than the result we achieved by using
the field-centric approach.
4.4.3 Designing a Signal Mixing Pipeline
We have seen that the structure of a search result can be derived from
the design of the signal mixing pipeline. We can turn this process up-
side down and attempt to develop a pipeline from a known search
result structure. This problem statement is of great practical value be-
cause it enables us to engineer features and scoring functions to a spec-
ification that describes a desired search result. This is closely related
to both relevance engineering and merchandising controls because the
specification can incorporate domain knowledge about relevance cri-
teria and business objectives. A programmatic system can provide an
interface that facilitates the specification of desired search results and
design of the signal mixing pipelines, as well as experimental evalua-
tion.
Let us go through an example of a relatively complex search result
specification to demonstrate the end-to-end engineering of signals and
scoring functions by using both textual and non-textual features. We
will consider the case of a fashion retailer who builds an online search
service. We will also assume that the user searches within a certain
product category to keep the problem reasonably simple (we will dis-
cuss how to achieve good precision when searching across multiple
categories in one of the next sections). Our starting point is the spec-
ification provided in Figure 4.14 that codifies the following business
rules:
• If a user searches for a certain product by its name or ID, a match-
ing product should be at the top of the search results.
• If a user searches for a certain brand, products of this brand

should be given priority and additionally sorted by newness and
customer rating in order to boost new or highly rated products.
216 search
• Otherwise, the results should be ranked according to average

relevance of product descriptions and other fields.
Figure 4.14: An example of a search result specification that can be used to de-
sign a signal mixing pipeline.
The specification mentions five different signals that should be taken

into account: exact product name or ID match, exact brand match,
product newness, product rating, and base average score. The exact
match signals can be obtained by using n-gram scoring for the cor-
responding name and brand fields. A reasonably good precision can
be implemented by using bigrams, and even stricter matching can be
achieved with trigrams or Boolean phrase matching. Note that we can
even turn off TFIDF calculations for the n-grams and simply count
the number of n-gram matches because we are interested in a binary
outcome (does the product name match or not?), rather than contin-
uous relevance grades. We can assume that newness and ratings are
available as numerical product properties. Newness, for example, can
be measured as the number of days since a product was introduced at
the store, and the average customer rating can be a real number on a
scale from 1 to 5. The base score can be calculated by using one of the
signal engineering methods that we discussed earlier. One of the most
straightforward methods would be to merge all product properties into
one field and calculate a TFIDF score for it.
We can start to build a signal mixing pipeline from the base score
and put more relevance tiers on top of it in accordance with the search
result specification. Each upper tier is created by amplifying the corre-
sponding signal so that all products that exhibit a property required
for this signal are elevated above the lower tiers in the search results
list.
The first tier that we put on top of the base score is the exact brand
match. We have to amplify the brand signal by assigning it a boosting
factor, and we also have to mix it with the rating and newness features
to achieve the desired secondary sorting, as shown in Figure 4.15. We
clearly need to rescale the raw rating and newness values to convert
them into meaningful scoring factors; this can be done in many dif-
ferent ways. A raw customer rating on a scale from 1 to 5 can be too
aggressive as an amplification factor and can be tempered by using a
square root or logarithm function to reduce the gap between low-rated
and high-rated products. For instance, the magnitude of the brand sig-
nal amplified by a raw rating of 5.0 is two times higher than for a rating
of 2.5; however, by taking the square root of the rating, we reduce the
difference down to 1.41.
Figure 4.15: A signal mixing pipeline that corresponds to the specification in

Figure 4.14.
The newness value should be transformed into a factor that gradu-

ally decreases with the age of a product. This can be done by using a
linear function, exponential decay, or Gaussian decay. For example, it
can be a reasonable choice to decrease the scoring factor by 10% every
30 days, which leads us to the exponential decay function
newness factor exppαxq (4.38)
in which x is the value of newness in days and parameter α is de-

termined from the following equation (according to which, the ratio
between any two factors separated by 30 days is 0.9):
expp30 αq 0.9 (4.39)
Combining the brand match signal with the rating and newness fac-
tors, we get a signal that corresponds to the second tier. Finally, the top
tier is crafted from the name match signal mixed with a large constant
factor that elevates the corresponding matching products to the very
top of the search results list.
218 search
4.5 semantic analysis
The search methods that we have discussed so far are all based on to-
ken matching. Although we have used a few techniques, such as stem-
ming, that deal with specific features of natural language, we have es-
sentially reduced the search problem to a mechanical comparison of to-
kens. This approach, sometimes referred to as a syntactic search, works
very well in practice and is used as a core method in most search en-
gine implementations. Syntactic search, however, has limited ability for
modeling the features of natural language that go beyond individual
terms. The meaning of words in natural language is often dependent
on the context created by the preceding and succeeding words and sen-
tences, and several types of such dependencies exist. Most of them are
related to one of the following two categories:
polysemy Polysemy is the association of one word with multiple
meanings. For instance, the word wood can refer to the material
or an area of land filled with trees. Polysemy represents a seri-
ous problem for relevance because a user might have in mind one
meaning of a word (e. g., products made of wood) but the search
engine will return documents that use the same word with a
different meaning (e. g., products that are related to woodlands,
such as forestry equipment). We have already encountered the
problem of complex concepts that are expressed with phrases
that must be treated as a single token, for example, dress shoes.
This issue can be viewed as a particular case of polysemy be-
cause the meaning of the individual words depends on the con-
text, so the meaning of the word dress depends on whether it is
followed by the word shoes. A particular case of polysemy that is
very common in merchandising applications is the usage of vo-
cabulary words in brand and product names. An example of this
problem would be the brand name Blue, which is indistinguish-
able from the color blue in queries like blue jeans.
synonymy Words are considered as synonyms if they can express

nearly the same meaning in a given context, such as candy and
sweet. Synonymy is also a major problem for search relevance
because a basic syntactic search is not able to match relevant doc-
uments that do not contain the given query term but do contain
synonyms. For example, items that contain the term sweet are
likely to be relevant for the query candy, but all of the search
methods we have discussed so far would fail to handle this de-
pendency.
4.5 semantic analysis 219
Some aspects of the polysemy problem can be handled by using

n-grams and the more advanced phrase-matching methods that we
discuss in the next sections, at least to a certain extent. In a similar
way, stemming can be viewed as a technique that handles synonymy
in a very basic form by reducing closely related terms to the same
root. These methods, however, lack the insight into the words’ mean-
ings and relationships that is required to solve polysemy and syn-
onymy problems. These issues require us to develop new techniques
that are focused on contexts and meaning, rather than separate tokens.
This approach is known as semantic search, named after the branch of
linguistics that studies the meaning and relationships of words and
phrases. Some methods of semantic searching are completely indepen-
dent from syntactic searching and compete with it, but many semantics
techniques can be used to extend syntactic search.
Polysemy and synonymy can be viewed as a problem of hidden re-
lationships between words or, alternatively, as a problem of finding
logical concepts that are materialized in words. From the latter stand-
point, polysemy refers to the case of two distinct concepts mapped to
one word, whereas synonymy refers to the opposite case of two dis-
tinct words mapped to one logical concept. The problem of semantic
analysis and search can therefore be viewed as the problem of finding
the right concepts and mappings between words and concepts. This
way of thinking is often referred to as concept search [Giunchiglia et al.,
2009; Hughes, 2015]. This term emphasizes the fact that concepts are
not merely statistical relationships between words but logical entities
that can be defined by using domain knowledge and other considera-
tions.
We will spend the rest of this section discussing the methods of se-
mantic search and analysis that can help with polysemy and synonymy
problems; some of these methods are closely related to recommenda-
tions. We will also use examples from the grocery domain, instead of
apparel, for a change.
4.5.1 Synonyms and Hierarchies
The most basic solution to the synonymy problem is a manually cu-

rated thesaurus, that is, a catalog of words and their synonyms. Once
created, a thesaurus can be used to transform documents and queries
in a way that closely resembles stemming. For example, we can define
the following set of words as synonyms:
candy, sweet, confection

220 search
Our goal is to make these terms identical from the querying stand-
point, so that documents containing the terms sweet and confection are
retrieved for the query candy, and the other way around. This can be
done in a few ways, and each method has advantages and disadvan-
tages.
The first approach is contraction. One of the synonyms in the list is
assigned to be the principal, and all occurrences of other synonyms
are replaced by the principal. For example, we can choose to replace
all occurrences of sweet and confection with candy, both in documents
and queries. Note that a principal does not necessarily have to be a real
word; it can be a special token that never appears in the input texts but
is used as an internal representation of a synonym group. Thus, the
contraction approach works exactly like stemming. Contraction clearly
achieves the goal of making all synonyms identical from the query-
ing standpoint, but the downside is that it collapses all synonyms into
the principal, which makes frequently used synonym terms indistin-
guishable from rarely used ones. This can negatively impact TFIDF
calculations.
The alternative to contraction is expansion. The expansion strategy
replaces each synonym instance with a full list of synonyms:
best candy shop Ñ [best] [candy] [sweet] [confection] [shop]
Expansion can be applied to either documents or queries, but not

both. Document-side expansion may have the same negative impact
on TFIDF scoring as contraction and also increases the size of doc-
uments. Query-side expansion preserves the correct IDF statistics but
makes a query more complex from the computational standpoint.
The expansion technique has one very important application that
goes far beyond basic synonym processing. Although synonyms are
defined as words that have roughly the same meaning, it is often the
case that one word represents a broader logical concept than another.
The relationship between such synonyms becomes asymmetric in the
sense that a broader concept can be considered as a synonym for a
narrower one, but not the other way around. For example, cake can be
used as a synonym for cheesecake, but it would be incorrect to replace
cake with cheesecake in some contexts. Consequently, it can be beneficial
to elaborate the expansion process and replace plain synonym lists
with directed rules, such that occurrences of cheesecake are expanded
into cake and cheesecake but occurrences of cake are not expanded. This
type of expansion is called genre expansion.
Developing this idea further, we can construct a hierarchy of terms
that describes nested classes of concepts, as illustrated in Figure 4.16.
At every level of a hierarchy, terms are expanded from their prede-
cessors and, thereby, become searchable for queries that contain more
generic terms. For instance, an item that contains the term fruitcake will
be included in a search results list for queries with the terms cake and
bakery.
Figure 4.16: Example of a category hierarchy for grocery products.
Similarly to basic expansion, genre expansion can be applied to both

documents and queries, but the methodology is quite different because
of the asymmetry of expansion rules. By applying genre expansion
to documents, we make queries for broad concepts match documents
about specific concepts, but not vice versa. For example, a query for
cake will return cheesecakes and fruitcakes. By applying expansion to
queries, we make queries for specific concepts match documents about
broad concepts, so a query for cheesecake returns cakes. In applying
genre expansion to documents, we also artificially decrease the IDF
of broad concepts because these terms are copied to more documents.
This is not necessarily negative because specific concepts will be scored
higher than broad ones, which is typically reasonable from a relevance
standpoint.
Contraction and expansion are very powerful techniques for mod-
eling semantic networks, and they give us some clues on how known
semantic relationships can be utilized in query processing. At the same
time, they provide no guidance on how these relationships can be in-
ferred. One possible solution is to create lists of synonyms manually.
This is the case in many merchandising search applications because it
enables merchandisers to use synonyms as a control that can express
certain business rules and domain knowledge. On the other hand, it
can be a challenge to manually curate a thesaurus in search applica-
tions with dynamic content, such as marketplaces. Also, it can be dif-
ficult to reveal some types of semantic dependencies without machine
222 search
learning methods. For example, the name of a famous athlete can con-
note to a certain sport, a type of sports equipment, or the brand that
the athlete promotes. Our next goal will be to develop methods that
are able to learn a thesaurus automatically.
4.5.2 Word Embedding
The vector space model states that a document or query can be rep-
resented as a vector in a linear space of terms. In the light of our dis-
cussion of the polysemy and synonymy problems, we know that terms
can be ambiguous and redundant, so it can be the case that a document
representation that uses terms as dimensions is not particularly good
or is at least flawed. Indeed, we have already discussed that polysemy
and synonymy can be viewed as a mismatch between words and con-
cepts, which suggests that words are a convoluted representation that
conceals semantic relationships.
We can attempt to find a better representation by changing the ba-
sis for the document space. Conceptually, we would like to map doc-
uments and queries to vectors of real numbers, so that the ranking
scores can be calculated simply as a dot product between the query
and document representations:
qÑp
dÑv
ķ
(4.40)
scorepq, dq p v pi vi

i 1
in which p and v are the k-element vector representations of a query

and a document and k is the dimensionality of the vector represen-
tation. Note that individual words can also be mapped to such vector
representations because each word can be treated as a single-word doc-
ument. This approach is generally known as word embedding. The term
embedding is used because the number of dimensions k is typically
small relative to the number of distinct words, so a high-dimensional
vector space model representation where each word has its own dimen-
sion is embedded into a low-dimensionality space. Word embedding has
many applications, which depend on how the vector representations
are constructed and used. From the search services perspective, we are
mainly interested in two possibilities. First, embedded representations
can be used for actual query processing, so that document scores for
a given query are calculated as dot products between the representa-
tions. Second, embedded representations of individual words can be
analyzed to create a thesaurus, that is, to match words with their syn-
onyms or related words.
The key problem of word embedding is, of course, how to construct
the new vector representations. Conceptually, we want a vector space
that preserves semantic relationships: words and documents with sim-
ilar or related semantic meanings should be collocated or lie on the
same line, whereas documents that have different semantic meanings
should not be collocated or collinear, even if they contain the same (pol-
ysemic) words. If we could construct such a space, it would be possible
to overcome the limitations of the vector space model. First, it would
be possible to find relevant documents for a query even if a document
and query did not have common terms, that is, to tackle the synonymy
problem. Second, it would be possible to rule out semantically non-
relevent documents even if they nominally contained query terms. In-
tuitively, we can assume that the semantic space can be constructed by
analyzing word co-occurrences, either globally in a document or locally
in a sentence, and identifying groups of related words. The dimensions
of the semantic space can then be defined based on these groups, and,
consequently, the vector representations of individual documents and
words will be defined in terms of affinities to the groups. It turns out
that this simple idea is very challenging to implement, and there exist
a large number of methods that use very different mathematical tech-
niques. We continue this section with a detailed discussion of several
important approaches and concrete models.
Finally, a note about terminology. Word embedding is a relatively
new term, and many semantic analysis methods, including the latent
semantic analysis and probabilistic topic modeling described later in
this section, were not developed specifically for word embedding (in
the sense of equation 4.40), but for different purposes and based on
different considerations. Most of these methods are very generic and
powerful statistical methods used in a wide range of applications from
natural language processing to evolutionary biology. These methods,
however, can also be viewed as word embedding techniques. In this
section, we choose word embedding as the main theme because it is
a convenient way to connect different semantic methods to each other,
at least in the context of merchandising search. The reader, however,
should keep in mind that it is just one possible perspective; seman-
tic analysis methods are not limited to word embedding and search,
and neither is word embedding limited to search applications. Even
in the scope of algorithmic marketing, semantic analysis methods can
be applied to many uses, including automated product attribution, rec-
ommendations, and image search.
224 search
4.5.3 Latent Semantic Analysis
One possible approach to semantic space construction is to analyze

the bag-of-words representations of documents to figure out which
terms frequently co-occur in one document. The intuitive expectation
is that terms that frequently appear together can be synonyms that
correspond to one logical concept. Thus, analysis of co-occurrences can
reveal concepts that are not explicitly observed in the documents as
terms but that exist at a semantic level. These concepts are referred to
as latent concepts.
Let us start with a collection of single-field documents containing
textual product descriptions. Our first step is to prepare a matrix that
contains the term frequencies for each term ti and document dj :
d1 d2 dn
tfpt1 , d1 q tfpt1 , d2 q tfpt1 , dn q

t1
t2 tfpt2 , d1 q tfpt2 , d2 q tfpt2 , dn q
X
(4.41)
.. .. .. ..
. . . .
tm tfptm , d1 q tfptm , d2 q tfptm , dn q
in which n is the number of documents and m is the total number of

distinct terms in the collection. This matrix, known as a term–document
matrix, is a representation of the documents in the term space. We can
calculate the similarity between documents by calculating a dot prod-
uct between the corresponding columns, and the similarity between
terms can also be calculated by taking a dot product between the cor-
responding rows. The term similarity calculated this way already pro-
vides some hints about semantic relationships between terms, in the
sense that terms that frequently appear together can be related to the
same concept. This metric, however, can be too noisy, so we need to
use a more robust statistical method.
Recall that the word embedding paradigm suggests the representa-
tion of each document as a k-dimensional vector. This representation
can also be written as a matrix. Let us define it as the n k matrix Vk ,
in which each row corresponds to a document and each column corre-
sponds to a semantic dimension. The latent semantic analysis method
(LSA) creates this matrix based on the heuristic consideration that it
should be possible to approximately reconstruct the term–document
matrix X from Vk by using a linear transformation [Deerwester et al.,
1990]. In other words, it should be possible to calculate the m k ma-
trix Lk , such that
X Lk VTk (4.42)
More specifically, matrices Lk and Vk should be defined to minimize

the reconstruction error. If the mean squared error is used as a measure,
this principle leads to the following optimization problem:
X Lk VTk

min

(4.43)
Lk ,Vk
Recall that the solution of this problem is given by the singular value
decomposition (SVD) that we discussed in Chapter 2. It is also very im-
portant that the matrices produced by the SVD algorithm are column-
orthonormal, which means that the k concept dimensions (the columns
of matrix Vk ) will be orthogonal to each other. This essentially means
that the original vector space model vectors (the rows of matrix X)
will be decorrelated by collapsing the strongly correlated vectors into
a single principal vector. This follows our intuitive expectation – the
frequently co-occurring terms correspond to highly correlated compo-
nents of the original term vectors (the rows of matrix X), so the decor-
relation is likely to merge co-occurring terms (potentially synonyms)
into one concept vector.
Let us now describe this process more formally. We first consider the
case of full SVD, in which the number of concept dimensions k is not
limited. The SVD algorithm breaks down the matrix into three factors:
X UΣVT
T
σ1 0 v1
. ..

u
1 ur

.
.
..
.

.

..
.

(4.44)
0 σr vn

m r
r r
n r
in which r is the rank of the term–document matrix X. Let us care-

fully examine this decomposition to understand how it can be helpful
in semantic analysis and search.
The columns of matrix U can be interpreted as a new basis for a
document space. Each column can be considered as a latent concept
that can incorporate multiple correlated terms, that is, terms that fre-
quently appear together in the same document. Each row of matrix U
corresponds to a term, so the magnitude of the coefficient uij is the
significance or contribution of term ti to concept uj . It can be the case
that certain terms in a concept have coefficients much higher than those
of the remaining terms – such a pattern indicates that these terms fre-
quently appear together in the same documents and are likely to have
a semantic relationship. The space spanned on concept vectors is often
referred to as the latent semantic space.
226 search
The rows of matrix V correspond to documents and the columns

correspond to concepts. Consequently, each row can be interpreted as
a vector of the coefficients that determine the significance of the cor-
responding concepts in a given document. This matrix is dual to the
original term–document matrix in the sense that each element vij can
be thought of as the frequency of a concept in a document, just like
each element of the term–document matrix represents the frequency
of a term.
The SVD representation enables us to calculate similarities between
queries and documents by using the basis of concepts. First, we can
calculate the cosine similarity between documents by using the rows of
matrix V. As each row is a vector representation of the corresponding
document in the concept space, the similarity for a pair of documents
can be straightforwardly calculated as
kvvik vvj

cos vi , vj (4.45)
i j
Next, we need to convert the query into a vector in the basis of con-
cepts in order to calculate the cosine similarity between the query and
documents. This process is known as query folding. We can rearrange
equation 4.44 to express the document vectors as a function of the
term–document matrix:
V XT UΣ1 (4.46)
A user query can be considered as yet another document that cor-

responds to some vector q of term frequencies, so we can substitute
it into equation 4.46 as a degenerate case of a term–document matrix
with one column:
p qT UΣ1 (4.47)
in which p is the required representation of the query in the basis of

concepts. Once this representation is obtained, we can score documents
against the query by using the cosine similarity in the basis of concepts:
p vi
score pq, di q cos pp, vi q (4.48)
kpk kvi k
Equation 4.48 defines a new scoring method, latent semantic index-
ing (LSI) scoring, which can be used as an alternative to the standard
vector space model and TFIDF approach. The principal advantage of
LSI scoring over term-based methods is the ability to fetch documents
that do not explicitly contain query terms. For example, a concept vec-
tor can include the three key terms candy, sweet, and confection, which
have strong semantic dependencies and are frequently used together.

A document that contains only candy and sweet would still have a large
coefficient for this concept in its vector representation. The same is true
for the query confection. Consequently, the document and query would
have high cosine similarity because of the intersection of this concept,
although they do not have common terms.
The next step is to apply dimensionality reduction, that is, to limit
the number of concept dimensions k r. Although dimensionality re-
duction increases reconstruction error 4.43, it is generally beneficial in
LSA applications because it decreases the noise and leaves only the con-
cept dimensions with the highest energy. Recall that SVD guarantees
that the columns of U are ordered by their significance2 . This implies
that concept u1 corresponds to the most persistent and frequent com-
bination of terms, whereas concept ur corresponds to the least signifi-
cant combination. Consequently, we keep only the strongest concepts
and truncate the leftmost columns of matrices U and V, thereby reduc-
ing the dimensionality of the concept basis and document space. The
number of concepts to preserve is an important parameter of the LSA
method. It is often set empirically by evaluating the precision and re-
call for several possible values and selecting the best one. The optimal
number of concepts is typically much lower than the number of dis-
tinct terms in the collection; 300–500 concepts is a good value, even for
large collections [Bradford, 2008]. With the assumption that the num-
ber of concepts k is set, decomposition 4.44 is transformed as follows:
Xk Uk Σk VTk
T
σ1 0 v1
. ..

u
1 uk

.
.
..
.

.

..
.

0 σk vn
m k
k k
n k
(4.49)
This truncated statement does not reconstruct the original term–

document matrix exactly but produces the approximation Xk . The
documents still correspond to the rows of matrix V, but each vector
has only k elements. In other words, the documents and queries are
mapped into a space with k dimensions, and the similarity metric is
computed in the same space as well.
2 See Chapter 2 for a detailed discussion of the exact meaning of significance in the context
of SVD.
228 search
example 4.4
Latent semantic analysis can be quite difficult to understand without

a numerical example, so we will provide one in the rest of this section.
The example we use is fairly small, but it is composed to highlight
the major features of LSA. It should be kept in mind, though, that
LSA is a machine learning method that requires substantial amounts
of data to work well in practice. We start with a collection of three
small documents about candy stores:
d1 : Chicago Chocolate. Retro candies made with love.

d2 : Chocolate sweets and candies. Collection with mini love hearts.
d3 : Retro sweets from Chicago for chocolate lovers.
Filtering out some stop words and applying basic normalization and
stemming, we get the following term–document matrix:
d1 d2 d3

chicago 1 0 1
chocolate 1 1 1

retro 1 0 1
candy 1 1 0

X
made 1 0 0 (4.50)

love 1 1 1
sweet
0 1 1
collection 0 1 0
mini
0 1 0
heart 0 1 0
By applying SVD and reducing the dimensionality down to two con-

cepts, we get the following factor matrices:
concept 1 concept 2

chicago -0.318 0.424
chocolate
-0.486 0.018

-0.318 0.424
retro
-0.333 -0.148
candy

U2 made
love

-0.166
-0.488
0.257
0.018

(4.51)

sweet -0.320 -0.239

collection -0.168 -0.406

mini -0.168 -0.406
heart -0.168 -0.406

Σ2 3.562 0
(4.52)
0 1.966
concept 1 concept 2

d1 -0.592 0.505
V2 d2 -0.598 -0.798
(4.53)
d3 -0.541 0.329
The first observation we can make is that the columns of matrix U2

highlight some logical themes that we can find in the text. The two
largest coefficients in the first column correspond to the terms choco-
late and love, quite closely followed by the coefficients for sweet and
candy. The largest coefficients in the second column correspond to the
terms Chicago and retro. This is because we have two documents that
consistently use the same set of words to talk about the Retro&Chicago
theme, and all three documents consistently use the same words to talk
about the Chocolate&Love theme.
A second insight can be gained from document matrix V2 . The first
column of the matrix corresponds to the Chocolate&Love concept. All
coefficients in the column have the same sign, so all three documents
are pointed in the same direction along this axis. The second column
corresponds to the Retro&Chicago concept, and the documents are
pointed in different directions because only the first and third docu-
ments mention this theme.
Let us now query the documents by using two queries, Chicago and
candy. The queries correspond to the following term frequency vectors
(the order of terms is the same as in matrix 4.50):

qchicago 1 0 0 0 0 0 0 0 0 0
(4.54)
qcandy 0 0 0 1 0 0 0 0 0 0
Transforming these vectors with formula 4.47 and calculating the

cosine similarities with the document vectors from matrix V2 , we ob-
tain the document scores presented in table 4.4. We can see that only
the first and third documents have high scores for the query Chicago,
which is expected. The second query, candy, is a more interesting case.
All three documents score highly, although the third document has
no occurrences of the term candy. This is because candy is a part of
the Chocolate&Love concept that is clearly present in the third doc-
ument. The LSA method was able to recognize the link between the
query and the document through this concept and rank the document
accordingly.
N
LSA was designed as an alternative to the basic vector space model

search methods, such as standard TFIDF scoring. Empirical study
230 search
Query d1 d2 d3
Chicago 0.891 -0.510 0.806
Candy 0.183 0.969 0.338
Table 4.4: The final document scores for the example of LSA calculations.
shows that it can actually outperform the basic vector space model in
many settings. In addition to that, LSA offers the following advantages:
synonyms A low-dimensional representation is able to capture syn-
onyms and semantic relationships. LSA can also estimate dis-
tances between words to generate a thesaurus that can be used
for synonym expansions in the standard TFIDF scoring, and
there exist specialized LSA-based methods to compute semantic
similarities, such as the correlated occurrence analogue to lexical
semantic (COALS) method [Rohde et al., 2006].
noise reduction Dimensionality reduction can efficiently remove

noise and redundancy from the data.
high recall LSA-based search works for queries and documents

that have no common terms. This allows high recall to be
achieved.
automation LSA relies on unsupervised matrix factorization, and,

thus, the process is completely automatic.
On the other hand, LSA has a number of shortcomings that stem
mainly from its heuristic nature, which neglects the complex statistical
properties of texts:
polysemy LSA has a limited ability to capture polysemy. Although
LSA is able to attach the same word to multiple concepts, thereby
capturing the fact that a word may have different meanings de-
pending on the context, LSA is not able to distinguish different
meanings of the word within a document because all meanings
are averaged into a term frequency in the term–document matrix.
This limitation stems from the nature of the bag-of-words model
and does not allow LSA to recognize more subtle semantic rela-
tionships between words.
completeness The theoretical foundation of LSA is incomplete be-

cause it does not provide any model for documents and terms.
interpretability The concept dimensions created by LSA can be

hard to interpret because of the negative values and absence of a
formal document model.
gaussian assumption One of the key advantages of the principal

component analysis leveraged by LSA is the ability to create un-
correlated concept vectors. The principle of decorrelation is based
on the assumption that the data have a Gaussian distribution, for
which zero correlation between the components implies indepen-
dence. This assumption, however, is not true for count matrices,
such as the term–document matrix.
We will attempt to address some of the LSA limitations in the next
few sections. We first discuss how the heuristic factorization model can
be replaced by a solid probabilistic framework, and we then rethink
the bag-of-words approach to better capture the semantic relationships
between words.
4.5.4 Probabilistic Topic Modeling
Probabilistic topic modeling is a family of semantic analysis meth-

ods that captures the semantic relationships between documents and
words through latent variables called topics. One of the key assump-
tions made in topic modeling is that the documents are generated term
by term with some probabilistic process. This process models the the-
matic structure of a collection of documents by using latent (hidden)
variables that can be interpreted as topics. Each document is typically
represented as a mix of topics, and each topic determines the distribu-
tion of words in the document. The generative process is designed to
reflect only certain statistical properties of the documents, such as the
topic-specific distribution of words, but it does not necessarily gener-
ate a text that looks real or readable. Figure 4.17 illustrates the main
components of a topic model.
Next, the generative process model is fitted by inferring the values
of the latent variables from the data. The data are a collection of doc-
uments in which each document is merely a group of terms, and the
latent parameters, that is, the topics, topic distributions, and relations
between topics and terms, are just abstractions that are never observed
directly but can be estimated. The statistical inference process can be
thought of as the ascent from the bottom of Figure 4.17 to the top. The
fitted model describes the associations between terms and topics (what
are the most distinctive words for a given topic) and the relationships
between topics and documents (what the document is about). Docu-
ments can also be searched by fitting a query into the estimated topic
structure and computing the similarity between the query and the doc-
uments in the latent topic space.
The topic modeling approach is similar to LSA in the sense that it
also uses the notion of latent topics and maps documents to a vector
232 search
Figure 4.17: A conceptual view of probabilistic topic models. Document terms

are generated sequentially, and each term corresponds to some path
in the model graph.
representation in the topic space. At the same time, its mathematical

foundation is very different from that of LSA. This foundation is very
important because it enables not just a single model but a whole fam-
ily of powerful methods and techniques. In the context of algorithmic
marketing, this group of methods is important not only for search but
also for recommendation services because it provides a generic frame-
work for modeling the relationships between different entities, such as
words and documents or users and products. In the next sections, we
will discuss two popular topic models – probabilistic latent semantic
analysis and latent Dirichlet allocation.
4.5.5 Probabilistic Latent Semantic Analysis
Probabilistic latent semantic analysis (pLSA) is one of the most basic

methods of probabilistic topic modeling [Hofmann, 1999]. Although it
approaches the semantic analysis problem from the probabilistic per-
spective, the resulting structure of the model can be viewed as matrix
factorization, which makes pLSA directly comparable to SVD-based
latent semantic analysis. pLSA can be viewed from two different per-
spectives. The first is the latent variable model, a probabilistic model

that uses the notion of latent variables (topics) to explain the relation-
ships between documents and terms. The second is matrix factoriza-
tion, which bridges the probabilistic latent variable model with LSA.
We discuss these two aspects separately in the next sections [Oneata,
1999].
4.5.5.1 Latent Variable Model

The pLSA model is one of the models from the probabilistic topic mod-
eling family. To describe the pLSA model more formally, let us first
define the following three main entities:
documents D td1 , . . . , dn u is the set of n documents.
terms T tt1 , . . . , tm u is the set of m terms (words) that contains all

distinct terms from all documents.
topics Z tz1 , . . . , zk u is the set of k topics, and k is the model pa-

rameter. The notion of topics corresponds to the notion of latent
concepts in LSA.
We explicitly observe pairs of documents and terms pdj , ti q but not
topics. The latent factor model assumes that each document may cor-
respond to multiple topics and the term probabilities within the doc-
ument are determined by the topic. For example, let us imagine two
topics that can be found in a grocery store catalog: dairy and desserts.
Some product descriptions in the catalog will mainly relate to dairy,
some will relate to desserts, and some will correspond to both topics
mixed in a certain proportion. Although the topics are not observed
directly, the distribution of terms in the document that corresponds to
dairy will be determined by the corresponding topic. We can express
this idea more formally by making the assumption that documents are
created by the following generative process:
1. First, draw a document dj from the probability distribution Prpdq
2. For each term ti in document dj :

2.1. Select a topic zl by drawing it from the distribution
Prpz | dj q
2.2. Select a term ti by drawing it from the distribution Prpt | zl q
This process corresponds to the probabilistic model shown in Fig-
ure 4.18. Each document is modeled as a mixture of topics, and the
distribution of tokens within the document is determined by the topics.
The same model is shown in Figure 4.19 in a more compact graphical
notation.
234 search
Figure 4.18: Detailed structure of the pLSA model.
Figure 4.19: Graphical model representation of the pLSA model. The outer box
represents the repeated choice of documents. The inner box repre-
sents the repeated choice of topics and terms within a document
that contains md terms. The shaded circles correspond to the ob-
served variables; the unshaded one denotes the latent variable.
Similarly to LSA, the pLSA model considers each document as a

bag of words. From the probabilistic perspective, this means that the
document–term pairs pd, tq are conditionally independent:
¹
Pr pD, T q Prpd, tq (4.55)
d,t
In addition to that, the pLSA model assumes that terms and docu-
ments are conditionally independent given the topic, that is
Prpt | d, zq Prpt | zq (4.56)
A joint probability model over D T can be expressed as
Prpd, tq PrpdqPrpt | dq (4.57)
for which the conditional probability of the term within the docu-
ment can be expressed as a sum of the probabilities over all topics:
¸ ¸
Prpt | dq Prpt, z | dq Prpt | d, zqPrpz | dq
z z
¸ (4.58)
Prpt | zqPrpz | dq
z
By inserting expression 4.58 into 4.57, we obtain the complete model

specification:
¸ ¸
Prpd, tq PrpdqPrpt | zqPrpz | dq Prpd, zqPrpt | zq
z z
¸ (4.59)
PrpzqPrpt | zqPrpd | zq
z
The next step is to learn the unobserved probabilities and, thus, infer
the latent topics. Given a set of training documents D, the likelihood
function is defined as
¹
L PrpD, T q Prpd, tqnpd,tq (4.60)
d,t
in which npd, tq is the number of times term t occurs in document

d, that is, the term frequency. Simplifying the likelihood function by
taking the logarithm, we obtain the following equation:
¸
log L npd, tq log Prpd, tq
d,t
¸ ¸ (4.61)
npd, tq log PrpzqPrpt | zqPrpd | zq
d,t z
The term probabilities Prpt | zq, document probabilities Prpd | zq,

and topic probabilities Prpzq are parameters of the model that have to
be fitted in a way that maximizes the likelihood. This is equivalent to
solving the following optimization problem:
max log L
¸
subject to Prpt | zq 1
t
¸
Prpd | zq 1 (4.62)
d
¸
Prpzq 1
z
This problem can be tackled by using the expectation–maximization

algorithm, which is the standard approach for maximum likelihood
estimation in latent variable models [Hofmann, 1999]. The challenge,
however, is that we have kpm 1q parameters Prpt | zq for all possible
pairs of terms and topics and kpn 1q parameters Prpd | zq for all pairs
of documents and topics. Note that we have kpm 1q parameters, not
km, because of the probability normalization constraints described by
problem 4.62. Thus, the number of parameters is high, and it grows lin-
early with the size of the document collection. This issue is considered
one of the major drawbacks of the pLSA model.
236 search
Given that the parameters are estimated, the relationships between

the documents, terms, and topics, as well as the semantic meaning of
the topics, can be analyzed by examining the magnitudes of the condi-
tional probabilities. The same parameters can be indexed and stored to
serve as search queries [Park and Ramamohanarao, 2009]. Similarly to
LSA, the similarity between the query and the document can be calcu-
lated in the latent semantic space as the cosine distance or dot product
between two vector representations. In the case of pLSA, the vector rep-
resentations of query q and document d in the latent semantic space
are given by the conditional probabilities Prpq | zq and Prpd | zq, re-
spectively. The similarity measure can then be defined as the following
dot product:
¸
scorepq, dq Prpq | zq Prpd | zq (4.63)
z
Values Prpd | zq are known from the model, but the query representa-
tion Prpq | zq needs to be learned for each query. This can be achieved
by fixing parameters Prpt | zq and Prpzq and fitting model 4.62 with
respect to Prpq | zq. The similarity metric can then be used to score and
rank the documents in the search results list.
4.5.5.2 Matrix Factorization

Although the latent variable approach is very different from LSA (a
probabilistic process instead of an algebraic matrix factorization), the
two methods are closely related. This can be demonstrated by rewrit-
ing the latent variable model in matrix notion. First, recall that LSA
approximates the term–frequency matrix defined by expression 4.41 as
a product of three matrices:
X U Σ VT (4.64)
in which X is an m n matrix of term frequencies for all pairs of

terms and documents, U is an m k matrix of term coordinates in the
concept space, and V is an n k matrix of document coordinates in the
concept space. On the other hand, we have determined that the joint
probability model in pLSA is given by a product of three factors:
¸
Prpd, tq PrpzqPrpt | zqPrpd | zq (4.65)
z
Rewriting this expression in matrix notation, we obtain the pLSA

model in a form that can be directly compared to the LSA factorization:
P L S RT (4.66)
in which L is an m k matrix of all term probabilities Prpt | zq, R

is an n k matrix of all document probabilities Prpd | zq, and S is
a diagonal k k matrix of the prior probabilities of the topics Prpzq.
In other words, pLSA, similarly to LSA, can be viewed as a matrix
factorization algorithm, but this factorization is driven by a different
objective. Whereas LSA is driven by minimization of the approxima-
tion error, pLSA is driven by maximization of the likelihood function
or, alternatively, minimization of the divergence between the observed
distribution and the model.
4.5.5.3 pLSA Properties

The pLSA model offers several important advantages over LSA. First,
the directions in the pLSA space are non-negative and interpretable
as probabilities. The directions in the LSA space do not have formal
interpretation and the values produced by the LSA factorization can
be negative, which also complicates the interpretation.
The second important difference is the handling of the polysemy prob-
lem. LSA is able to map synonyms to the same location in the latent
semantic space, but it is typically not able to distinguish different mean-
ings of the same word depending on the context. By contrast, pLSA
distributes the probability mass of a term over several different topics
that can correspond to the different senses of a word [Hofmann, 1999].
More specifically, if the same term t is observed in two different docu-
ments di and dj , the topic it has the strongest association with in the
context of the first document
argmax Pr pz | di , tq
z
can be different from the topic that this term is associated with in the
context of the second document

argmax Pr z | dj , t
z
Despite these advantages, pLSA generally involves more complex

implementation than LSA. Whereas LSA is based on the determin-
istic SVD factorization, pLSA requires the iterative expectation–
maximization algorithm to estimate the parameters of the model. The
pLSA model also has several structural issues, which we will discuss
and address in the next section.
4.5.6 Latent Dirichlet Allocation
The pLSA model is a major step forward relative to LSA. It establishes

a solid statistical framework that allows one to extend, simplify, or
238 search
combine different models by using probabilistic techniques. The pLSA

model, however, has several shortcomings:
• Each document is represented as a vector of probabilities, not
a generative probabilistic model. These probabilities are the pa-
rameters that need to be estimated from the data. This results in a
high number of parameters that grows linearly with the number
of terms and documents and makes the learning process more
prone to overfitting.
• pLSA does not impose constraints on how documents and terms

are associated with topics. Intuitively, we would expect each doc-
ument to be related to a small number of topics and each topic to
be associated with a small number of terms, but pLSA does not
provide explicit parameters to control this aspect of the model.
These issues can be addressed by creating a model with a more elabo-
rate generative process than the pLSA process that we described earlier.
In this section, we discuss one of the most prominent examples of such
models, called the latent Dirichlet allocation (LDA). The LDA model
can be viewed as a generalization of pLSA, and it is one of the most
popular and widely used probabilistic topic models [Blei et al., 2003].
The LDA model is based on the notion of the Dirichlet distribution, so
the reader can use Appendix A at the back of the book as a reference.
Similarly to pLSA, the LDA model uses the latent variable approach,
which assumes that each document corresponds to a mixture of latent
topics and the document terms are drawn from the distributions asso-
ciated with the topics [Blei et al., 2003]. With the assumption that the
number of latent topics k is predefined, the LDA model is described by
using the following generative process for each document d from the
collection of documents D:
1. Draw the number of terms in the document md from some ran-
dom distribution. The choice of this distribution is not critical for
the model design.
2. Draw a k-dimensional vector of probabilities p from the Dirichlet

distribution Dirpαq, in which α is the model parameter. Each el-
ement of p is interpreted as the probability of the corresponding
topic, so that this vector defines the mixture of topics.
3. For each term in the document:

3.1. Choose a topic zt according to the probability vector p, that
is, Prpzt i | pq pi
3.2. Choose a term t from the multinomial probability distribu-

tion Prpt | zt ; βq conditioned on the topic zt . This distribu-
tion is defined as the model parameter β for each pair of
term and topic.
In comparison with the pLSA process described in section 4.5.5.1,
the key difference is that the LDA model draws topics from a global
parametric distribution, not from the distributions learned for each doc-
ument. The parameters of this model are the k-dimensional Dirichlet
parameter α and the k m matrix of term probabilities β, in which
m is the total number of distinct terms in all documents. Each row of
the matrix β defines the multinomial distributions over the words for
a corresponding topic. These parameters are sampled once for a col-
lection of documents, and, consequently, the number of parameters is
smaller than that with pLSA. The graphical model that corresponds to
the generative process is shown in Figure 4.20.
Figure 4.20: Graphical model representation of the LDA model.
In the context of a single document, the joint distribution of a topic

mixture, all topics, and all terms is given by:
¹
Prpp, z, tq Prpp | αq Pr pt | zt ; βq Pr pzt | pq (4.67)
t
in which the distribution Prpp; αq is defined as Dirpαq and the pa-
rameters α and β are given. Note that Pr pzt | pq is simply the proba-
bility value from p that corresponds to zt . The marginal distribution of
a document can be obtained by integrating over the topic probabilities
and summing over all topics:
» ¹¸
Prpdq Prpp | αq Pr pt | zt ; βq Pr pzt | pq dp (4.68)
t z
The likelihood function for a collection of documents can then be
obtained by taking the product of the document probabilities:
¹» ¹¸
PrpDq Prppd | αq Pr pt | zdt ; βq Pr pzdt | pd q dpd (4.69)
d P
t d z
240 search
The process of fitting this model is complicated because the param-

eters α and β are coupled in the internal sum of equation 4.69. This
problem can be tackled with methods of approximate posterior infer-
ence, such as variational inference and Gibbs sampling [Blei et al., 2003;
Asuncion et al., 2009].
The LDA model addresses the two issues with pLSA that we men-
tioned earlier. First, it reduces the number of parameters by defining
a different generative process that does not use document-specific pa-
rameters. Second, the Dirichlet prior shapes the topic probabilities in a
way that penalizes relationships between the topics and documents.
4.5.7 Word2Vec Model
Word2Vec is a group of models that attempt to overcome the limita-

tions of the bag-of-words-based semantic analysis methods by taking
into account the local context of a word, rather than the entire docu-
ment [Mikolov et al., 2013a,b]. The two main types of Word2Vec models
are the continuous bag-of-words and skip-gram models. The continu-
ous bag-of-words approach is to build a predictive model that estimates
the probability of a word based on one or more words in the surround-
ing context, as shown in Figure 4.21. The words in the sliding context
are interpreted as a bag of words, that is, we account only for distinct
terms and their frequencies, not for their order. The predictive model
can be built in such a way that each word is associated with a vector
of weights set by the model fitting process. These vectors can then be
interpreted as word representations in some latent semantic space, sim-
ilarly to the vectors produced by LSA or topic models, so this vector
representation can be used for search and thesaurus creation. The skip-
gram approach is somewhat opposite to the continuous bag-of-words
model – it takes the target word as an input and predicts the context.
The design of predictive models, however, is very similar for both the
continuous bag-of-words and skip-gram models, and we choose to fo-
cus on the first approach in the rest of the section.
The Word2Vec model uses a shallow neural network to capture se-
mantic relationships and predict a term based on its context. We first
discuss the design of the network under the assumption that there is
only one word in the context and later generalize the result for cases
with multiple words. The neural network used in the Word2Vec model
consists of the input layer, hidden layer, and output layer depicted in
Figure 4.22.
The input of the network is a binary vector that represents a context.
If the total number of terms in the collection is n, the input vector has
n elements, and each element is equal to one if the corresponding term
context target
the best chocolate cakes and candies in the town
Figure 4.21: Example of the continuous bag-of-words model with a context of

two words.
x h s y
t₁ x₁ s₁ y₁
t₂ x₂ h₁ s₂ y₂
W V
n×m m×n
hm
tn xn sn yn
input layer hidden layer output layer
Figure 4.22: The design of a Word2Vec neural network for a single-word context.
is present in the context and zero otherwise. As we are considering the

case of a single-word context, let us denote the only context term as tk .
The input vector has then only one non-zero element xk :
#
ik
xi 1,
(4.70)
0, otherwise
The input is transformed into m intermediate outputs by using the

units of the hidden layer. This transformation is chosen to be linear
(each intermediate output hi is a weighted sum of the inputs xi ), so it
is defined through the weight matrix W:

w1

Wnm ..
.

(4.71)
wn
The intermediate outputs can then be expressed as the product of

the weight matrix and the input vector. According to our assumption,
242 search
the input vector contains only one non-zero element, so the result will
be identical to the corresponding row of the weight matrix:
h WT x wTk (4.72)
The output score is produced by the combination of a linear transfor-

mation and a softmax function. The linear part, similarly to the hidden
layer, is defined by using the weight matrix V:

Vmn v1 vn
(4.73)
This matrix is multiplied with the intermediate outputs to produce

scores for each of n terms:
si vTi h, i 1, . . . , n (4.74)
The signals si are arbitrary values, but we want to interpret them as

predicted probabilities of the corresponding terms given the context. In
other words, we are solving a multiclass classification problem where
the context has to be assigned to one of n classes that corresponds to
the predicted term. As we discussed in Chapter 2, the standard way to
map a vector of arbitrary values to category probabilities is the softmax
function, so the final outputs are defined as follows:
yi
ps i q
Pr pti | tk q °nexpexp (4.75)

j 1s j
The network defined above can be trained by using standard meth-

ods of training for artificial neural networks. We cannot go deeply into
the details of training algorithms here, but it will be useful to briefly
review the main steps to get the idea of how Word2Vec models can be
fitted [Rong, 2014b]. We train the model iteratively by taking sample
pairs of a context and a target word, evaluating the network for a con-
text, comparing the network output with the target, and adjusting the
weights in matrices W and V. Let us assume that, for a given iteration,
the actually observed target term for the context tk is ta . According to
the principle of maximum likelihood, our objective to is to maximize
the predicted probability of the actual term, given the context (because
we ultimately want to maximize the mathematical expectation of this
probability over all contexts):
max Etk ,ta r Pr pta | tk q s (4.76)

Maximization of this probability is equivalent to minimization of the

following loss function:
J log Pr pta | tk q (4.77)
Substituting definition 4.75 for the network output into defini-

tion 4.77 for the loss function, we find
ņ
J log ya sa log exp sj (4.78)

j 1
Our goal is to minimize the loss function with respect to weights w

and v. This can be done by updating the weights with the stochastic
gradient descent based on the prediction errors. Our strategy is to start
from the output side of the network and calculate the weight updates
for matrix V based on the observed prediction errors. Next, we move
one layer backward and calculate the weight updates for matrix W.
This approach is known as backward propagation of errors or simply
backpropagation. At each layer, we need to calculate the gradient of the
loss function with respect to the weights. We do this in two steps – we
first calculate the gradient with respect to the scores, and we then use
the result to calculate the gradient with respect to the weights. So, we
start with the derivative with respect to the scores of the output layer:
BJ Ipj kq B log ņ exp ps q

B sj B s j i 1 i

Ipj kq °nexpexpsj ps q (4.79)
i1 i
y j Ip j k q
ej
in which Ipj kq is the indicator function equal to one if j k and
zero otherwise. We can see that this derivative is simply the prediction
error, so we denote it as ej . By taking the derivative with respect to the
weights of the output layer, we find the gradient for weight optimiza-
tion:
B J B J B sj
Bvij Bsj Bvij ej hi (4.80)
This result means that we should decrease weight vij if the product
ej hi is positive and increase the weight otherwise. The stochastic
gradient descent equation for weights will thus be as follows:
(new)
vj v(old)
j λ ej h, j 1, . . . , n (4.81)
244 search
in which λ is a learning rate parameter. The next step is to repeat

the process for the hidden layer. First, we take the derivative of the loss
function with respect to the intermediate outputs:
BJ ņ BJ Bsj ņ e v ε
Bhi j1 Bsj Bhi j1 j ij i (4.82)
The result denoted as εi can be interpreted as a weighted sum of

prediction errors. We calculate this value for each of m hidden units,
so we obtain an m-dimensional vector of prediction errors:
ε rε1 , . . . , εm s (4.83)
Next, we calculate the gradient with respect to the weights of the

hidden layer:
BJ B J B hi
Bwji Bhi Bwji εi xj (4.84)
We use this result and the stochastic gradient descent to update the
weights of the hidden layer, similarly to our method for equations 4.80
and 4.81 for the weights of the output layer. Taking into account the
fact that all xj values in equation 4.84 are zeros except for xk , we find
that only the k-th row of matrix W needs to be updated:
wk
(new)
w(old)
k λ εT (4.85)
The Word2Vec model can be trained by applying equations 4.81 and

4.85 iteratively for training pairs of context and target words. This pro-
cess, however, is computationally intensive because, in accordance with
equation 4.81, we need to update the weight vectors v for all terms for
each training sample and the number of terms n can be large. This re-
quires the use of optimization techniques, such as hierarchical softmax
and negative sampling, in practical implementations of the Word2Vec
model [Mikolov et al., 2013b; Rong, 2014b].
We can generalize the obtained results for the case of a context with
multiple words in a straightforward way. The input vector for a context
of q words, that is, q non-zero elements, can be thought of as a normal-
ized sum (i. e. , average) of q single-word contexts. This is illustrated
in Figure 4.23, although the actual network design does not change.
This allows us to rewrite the equation for the hidden layer in the
following way:
T
h q1
1 T
W x1 xq wk1 wkq (4.86)
q
x₁₁
x1
x1n
s₁ y₁
x₂₁ h₁ s₂ y₂
x2 + W V
n×m m×n
x2n hm
sn yn
xq1
xq
xqn
Figure 4.23: The Word2Vec model for a context with multiple words.
The equation for the loss function remains the same, although it
represents a different conditional probability:
ņ
J log Pr ta | tk1 , . . . , tkq sa log exp sj (4.87)

j 1
Going through the gradient calculations, we find that all equations

remain unchanged and result in the same pair of weight update for-
mulas:
(new)
vj v(old)
j λ ej h, j 1, . . . , n (4.88)
(new)
wkj w(old)
k j
qλ εT , j 1, . . . , q (4.89)
The only difference is that we update multiple weight vectors w be-

cause we have multiple terms in the context.
example 4.5
After the network is trained, each of n terms in the collection corre-

sponds to a pair of m-dimensional vectors w and v. The power of

the Word2Vec model stems from the fact that these vectors provide
246 search
an insightful representation of words that preserves the semantic rela-

tionships. Let us illustrate this with an example of a Word2Vec model
trained on the following samples, where each sample is a pair of a
context word and a target word [Rong, 2014a]:
drink coffee tea drink

drink juice juice drink
drink tea coffee drink
eat cake pie coffee
eat pie cookie juice
eat cookie cake tea
pie tea cake coffee
We choose to use a network with 8 hidden units to capture the se-

mantic patterns in these samples, so each term is represented as two 8-
dimensional vectors after the model is trained. To visualize the weight
vectors, we can project them to a two-dimensional space by using prin-
cipal component analysis. For instance, the weight vectors of a hidden
layer w are projected onto a plane in Figure 4.24. We can see that the
words are clustered according to their usage patterns and, ultimately,
meaning.
juice
tea coffee
drink
eat
cake
pie
cookie
Figure 4.24: Example of word clustering with a Word2Vec model.
More surprisingly, a Word2Vec model trained on a large collection

of texts produces vectors that enable a sort of algebra in the latent
semantics space. Consider the following examples, in which vpq stands
for a vector representation of a term produced by the Word2Vec model
4.6 search methods for merchandising 247
and minus is used in the sense of a standard element-wise difference

between the vectors:
v psteakq v pbeefq v psaladq v ptomatoq
v psteakq v pbeefq v pbreadq v pflourq (4.90)
v pfranceq v pparisq v pjapanq v ptokyoq
We can see that the vector difference captures the concept of

cooking in the first two cases and the concept of country–capital
relationship in the last case. In other words, by adding the dif-
ference between beef and steak (which can be interpreted as an
act of cooking) to a tomato, we get a salad. This type of semantic
relationship is called word analogy. Note that each concept, such
as cooking or capital, that is, one of the vector differences in the
examples above, is also a vector in the semantic space, so we can
conclude that a certain direction in the space corresponds to cooking,
another direction to zooming in from country to capital, and so on.
N
One of the applications of the Word2Vec methods in merchandising

search is thesaurus generation. For instance, it was used by Dice.com,
a job search engine, to overcome the struggle with synonymy in job
titles and descriptions [Hughes, 2015]. The benefit of this approach is
that it can produce a thesaurus or word clusters that can be used by
the standard syntactic search engine.
We conclude our discussion of Word2Vec with a brief comparison
with the LDA, pLSA, and LSA methods. The key difference is that
Word2Vec uses a local context window to learn semantic dependen-
cies, whereas the topic modeling methods use global document statis-
tics. Both approaches have strengths and weaknesses. For example,
Word2Vec can generally capture word analogy better than topic model-
ing [Pennington et al., 2014]. On the other hand, the vector representa-
tions produced by the Word2Vec model are not sparse and may contain
negative elements, which make them less interpretable relative, for ex-
ample, to the results of LDA, which is able to create sparse vectors that
are interpretable as probabilities. This makes Word2Vec less applicable
in applications related to topic analysis.
4.6 search methods for merchandising
Thus far, we have discussed relatively generic search methods and their
applications in merchandising search. The challenge of merchandising
search, however, goes well beyond the tuning of standard methods
and requires the creation of more specialized search techniques for
248 search
domains like eCommerce. This is explained by a number of reasons

[Khludnev, 2013]:
• Structured entities. Many standard search methods are designed
for documents with a relatively simple structure that includes
one or several textual fields. A merchandising search often deals
with highly structured documents that resemble records in rela-
tional databases, rather than texts. For example, a typical product
item can have hundreds of numerical and categorical attributes,
so that the corresponding document looks like this:
Brand: Tommy Hilfiger
Type: Jeans
Color: Black
Weight: Super Skinny
...
Moreover, it is often the case that the merchandise is grouped

into nested entities or associated with category hierarchies. For
example, a retailer can sell a dinnerware set as a single item,
but this item includes multiple products, and each product, in
turn, may have multiple variants of different colors or sizes. This
requires a search service to operate with nested or interrelated
entities that cannot be correctly represented as plain documents.
• Diversity of merchandise. Many search applications take advantage

of ranking to improve the perceived precision of the search re-
sults. Unfortunately, it turns out that the power of ranking and
its applicability is limited in merchandising search. One of the
main reasons is that search results produced by standard scoring
and signal mixing methods tend to be excessively diverse, which
creates a poor user experience. For example, the query red dress
can match a wide range of products that contain these two terms
in their attributes, including dresses, shoes, and even watches.
The advanced signal engineering and mixing techniques that we
discussed earlier can help to improve the results, but they are
unlikely to provide a robust solution for the problem. Another
reason is that TFIDF scoring and other popular scoring meth-
ods can perform poorly on highly structured documents with
multitudinous categorical fields. This is not surprising because
these methods are designed for natural texts.
• Compound and polysemic terms. Industrial experience demon-

strates that the quality of a merchandising search significantly
depends on proper handling of compound and polysemic terms.
Search queries in merchandising applications often contain
multiword brand names and concepts, such as Calvin Klein
and dress shoes, that clearly communicate the search intent if

processed as phrases but that can be misinterpreted if broken
down into standalone words. Moreover, many brand names
contain common words, which makes proper interpretation
even more challenging. For example, the query pink sweater can
match all products made by the brand Pink Rose, and, conversely,
the query pink rose sweater can match all products of pink and
rose colors although it clearly indiates an intent to find a certain
brand.
The observations above suggest that we should develop search meth-
ods that are focused on precision, exact matching, and attribute struc-
tures, as opposed to scoring. In other words, we should consider search
methods that treat documents more like database records that have to
be queried, rather than free texts that have to be scored. A number of
such techniques have been developed by Macy’s, a leading US retailer,
for their eCommerce search services [Kamotsky and Vargas, 2014; Pe-
ter and Eugene, 2015]. We will spend the rest of this section discussing
these methods.
4.6.1 Combinatorial Phrase Search
Our first goal is to improve the precision of search results, given that
documents have many categorical fields that often contain compound
and polysemic terms. We can make the observation that making users
write structured Boolean queries would be a great solution for this
problem. For example, the free-text query pink rose sweater becomes
much less ambiguous if the user explicitly articulates fields and com-
pound terms:
brand:[pink rose] AND type:[sweater]
This approach can be used in certain applications of merchandising

search if the user is provided with a convenient interface that helps
to specify individual fields. For example, car selling sites often pro-
vide drop-down lists of car makers, car models, and other properties
to enable the user to specify field-level search criteria. This can be a
reasonable solution for business domains with a relatively cohesive as-
sortment, such as automobiles or real estate, but free-text queries are
arguably preferable for domains with diverse merchandise, such as de-
partment stores.
A free-text query does not contain document fields or demarcation of
compound terms, which thereby creates ambiguity. The idea of a com-
binatorial phrase search is to recover some of this information by gen-
erating multiple Boolean queries with different combinations of fields
250 search
and terms from the original free-text query and by searching for docu-
ments containing these artificial queries. A query generation algorithm
is designed to produce relatively restrictive search criteria, so that doc-
uments must strongly correlate with a query to match. This increases
the probability that documents are included in a search results list not
because of accidental matching of separate terms but because the doc-
ument attributes provide a really good coverage of query terms and
phrases. This methodology can be considered as a generalization of
shingling for multifield documents.
The first step of a combinatorial phrase search is to partition the
query into sub-phrases. Let us assume that a query entered by a user
is a sequence of n terms:
q rt 1 t 2 . . . t n s (4.91)
There exist 2n1 possible partitions of this query into sub-phrases

because there are n 1 gaps between the query terms, and we indepen-
dently choose to split or not to split a query at any gap. For example,
there are four possible partitions for a query with three terms (we use
square brackets to denote sub-phrases):
rt1 t2 t3 s
rt1 s rt2 t3 s
rt1 t2 s rt3 s (4.92)
rt1 s rt2 s rt3 s

The second step is to generate a Boolean query for each partition,
such that each sub-phrase in a partition is required to match one of the
fields in a document. If we have m sub-phrases s1 , . . . , sm in a given
partition and a document has k fields f1 , . . . , fk , the Boolean query will
be as follows:
pf1 s1 OR f2 s1 OR . . . OR fk s1 q
AND pf1 s2 OR f2 s2 OR . . . OR fk s2 q
(4.93)
...
AND pf1 sm OR f2 sm OR . . . OR fk sm q
We use the equals sign in query 4.93 to denote an exact match be-
tween the query sub-phrase and the field value; both sides must be
exactly the same, although we can apply normalization, stop words, or
stemming to preprocess both the query and fields. Query 4.93 ensures
that a document provides a reasonably high coverage of a given parti-
tion, in the sense that each sub-phrase must have an exact match with
one of the fields.
Finally, the Boolean queries for all partitions are executed, and the fi-
nal search result set is obtained as a union of the search results from all
Boolean queries. This is equivalent to combining all partition queries
into one big Boolean query with the OR operator. The overall struc-
ture of this query is visualized in Figure 4.25. Our partition generation
algorithm does not try to recognize compound terms in a query and
just mechanically splits it into sub-phrases. Consequently, sub-phrases
are likely to be misaligned with the compound term boundaries. For
example, the query blue calvin klein jeans can be partitioned into the
sub-phrases blue calvin and klein jeans. By combining all of the parti-
tions together, we ensure that at least some of the splits capture the
compound terms correctly.
Figure 4.25: The overall structure of a combinatorial phrase search query.

252 search
example 4.6
We will illustrate the logic of the combinatorial phrase search with an

example. Let us take the query pink rose sweater, which can be parti-
tioned in the following ways:
Partition 1 : rpink rose sweaters

Partition 2 : rpink roses rsweaters
Partition 3 : rpinks rrose sweaters
Partition 4 : rpinks rroses rsweaters
Let us assume that the products in the catalog are represented as
documents with three fields: brand, product type, and color. A combi-
natorial phrase search query assembled for such fields and partitions
will have the structure presented in Figure 4.26. The query can become
very large and computationally challenging as the number of fields
and query terms increases, but we can partially mitigate this by intro-
ducing certain simplifications. For example, we can limit the maximum
length of sub-phrases in a partition because sub-phrases that are too
long are unlikely to represent meaningful compound terms.
The following document will match the combinatorial query because
the brand and type fields cover all of the sub-phrases in the second
partition:
Brand: pink rose

Type: sweater
Color: black
At the same time, the query will not match a product of a sweater
type and pink color until the brand name is rose. Moreover, a combi-
natorial phrase search becomes even more restrictive as the length of
the query grows because all terms need to be covered. This behavior is
different from the standard vector space model that appreciates every
term match and, thereby, decreases in precision as the length of the
query increases.
The disadvantage of combinatorial phrase search is that the number
of partitions grows exponentially with the number of terms in the user
query and so does the number of statements in the resulting Boolean
query. In practice, the complexity of the Boolean query can often be
reduced by excluding some statements based on the field type. For
example, it can be the case that the field color has only a few valid
possible values, so filters like
Color = [sweater]
Color = [pink rose]
can be eliminated by the query generator as meaningless.

N
Figure 4.26: An example of a combinatorial phrase search query.
The combinatorial phrase search can be considered as a technique

for finding documents that provide complete query coverage in terms
of sub-phrases and fields. This method, however, can be connected to
semantic search as well. Although the combinatorial search does not
detect semantic relationships as explicitly as LSA or Word2Vec do, it
attempts to identify and match compound terms that are likely to rep-
resent logical concepts. In other words, it can be viewed as an attempt
to build a semantic search with syntactic search primitives [Giunchiglia
et al., 2009; Khludnev, 2013].
4.6.2 Controlled Precision Reduction
The combinatorial phrase search achieves a high precision of search

results by ruling out all documents that do not fully cover the search
254 search
query. This helps to keep the search results consistent and use the avail-
able display space efficiently. The combinatorial approach, however,
has its own downsides. One of the most significant issues is that strict
matching can return an empty search results list if a query contains
misspelled words or some other unfortunate combination of terms that
cannot be covered by the available documents. This behavior is unde-
sirable because it leaves the user with an empty screen instead of a list
of products, which thereby decreases the probability of a conversion.
This problem can be addressed by taking additional actions if an
empty search result is returned by the basic combinatorial phrase
search algorithm. For example, we can first attempt to run a combi-
natorial search that requires exact field matching and then fall back
to the basic vector space model that allows partial field matches. We
can develop this idea further and create a chain of search methods
with gradually decreasing precision, with each method invoked
sequentially until at least one document (or some other minimum
number of documents) matches. For example, a chain could have the
following structure:
1. Exact match. Search documents with a standard combinatorial
phrase search without normalization or stemming.
2. Normalization, stemming, and spelling correction. If there is no

match, apply normalization and stemming to the field and query
terms and run the combinatorial phrase search again. We can
also apply spelling correction to the query terms.
3. Shingling. If there is no match, search by using shingling instead

of seeking an exact match.
4. Partial match. If there is still no match, attempt to search with

one or more words removed from the search query, so that a
document does not need to provide full coverage of a query but
only partial coverage.
Any search pass can terminate the process and return search results
that have been found. For example, the misspelled query Abibas sneak-
ers is likely to match no documents if the combinatorial phrase search is
used, but the corrected query Adidas sneakers is likely to match enough
documents to stop further relaxation of the search criteria. This tech-
nique, called controlled precision reduction, helps to manage the trade-
off between the high precision of combinatorial phrase search and the
risk of frustrating the user with empty search results.
4.6.3 Nested Entities and Dynamic Grouping
From the merchandising standpoint, the problem of search can be

viewed as the problem of efficient usage of display space. Although
it is very important to present the user with a set of relevant prod-
ucts that meet the search intent, it is also important to demonstrate
the available assortment in the best possible way, given that the user
can browse only a limited amount of search results. Efficient utilization
of the display space is a major challenge for merchandising search be-
cause product catalogs often contain closely related products that are
likely to be perceived by the user as duplicates. For example, it can be
more reasonable to present the user with a set of relevant but different
dresses, rather than multiple variants of the same dress that comes in
several colors and sizes.
The problem of duplicated products and inefficient usage of the dis-
play space arises from the hierarchical relationships between items in
the product catalog. The nature and structure of these relationships
heavily depend on the business domain. The following product hierar-
chy, for example, can often be found in department stores:
• The smallest unit of the merchandise is a product variant, com-
monly referred to as a stock keeping unit (SKU). An example of a
product variant would be Levi’s 511 jeans in a white color and
size of 30W 32L. All physical instances of the same variant are
considered identical.
• A product is a logical entity that includes one or several product

variants. For example, Levi’s 511 jeans is a product that includes
variants of different sizes and colors. A product typically has
a price, with the implication that all of its variants are priced
equally.
• Multiple products can be combined into a product collection. A

collection can be sold as a single unit or a user can be allowed
to select individual items from it. For example, multiple dinner
plates, bowls, and mugs can be sold as a dinnerware set under
one price tag. Alternatively, the user can choose to buy a subset
of specific product variants.
All of the search methods that we have discussed so far assume that
catalog items are modeled as plain documents, so we need to map a
hierarchical structure with product collections, products, and product
variants to a collection of documents. One possible way is to model
each product variant as a separate document, so that each item in a
search results list corresponds to one product variant. Although this
256 search
approach is generally valid and widely used in practice, it is vulner-

able to the issue with duplicates and, consequently, inefficient usage
of the display space. Figure 4.27 illustrates this problem: variant-level
document modeling leads to a search result that is formally relevant
but not very efficient from the merchandising standpoint relative to
product-level modeling, which does a better job of demonstrating the
assortment.
Figure 4.27: Examples of search results lists for the query evening dresses and
different approaches to data modeling.
Product-level document modeling can help to solve the problem

with duplicates, but it introduces its own challenges. A straightforward
approach to product-level modeling is to represent each product as a
single document. This requires all product attributes and product vari-
ant attributes to be merged into a plain list of fields; in other words, a
product has to inherit all of the attributes of its variants. For example,
consider the following two variants of a travel suitcase that belong to
one product:
Brand: Samsonite Brand: Samsonite

Name: Carry-on Hardside Suitcase Name: Carry-on Hardside Suitcase
Color: red Color: black
Size: small Size: large
The two variants can be merged into one product-level document with
the following structure:
Brand: Samsonite
Name: Carry-on Hardside Suitcase
Color: red black
Size: small large
The result looks reasonable because the document is scored well for
queries like red suitcase, small suitcase, and so on. The major problem
with this approach is that it loses the structural information about
nested entities, which makes it impossible to distinguish valid attribute
combinations from invalid ones. The document above is scored well for
the query small red suitcase, and this is correct because one of the vari-
ants really is small and red. However, the same document is scored
equally well for the query small black suitcase. This is not correct be-
cause none of the variants are small and black at the same time, which
makes the product non-relevant for the user’s search intent. This prob-
lem is quite challenging from the implementation standpoint because
it cannot be solved purely in terms of plain documents and requires a
search engine either to explicitly support nested entities or to operate
with variant-level documents internally and then rework the results
to group variants into products. If product filtering is implemented
correctly, product-level results can substantially improve the merchan-
dising efficiency of a search service.
We have found that collapsing product variants into products can be
beneficial, so we can consider the possibility of collapsing products into
product collections as the next step. This question is more complicated
because a user can have different search intents and look for either
products or product collections. For example, a user who searches for
dinnerware is likely to expect collections, whereas a user who searches
for a cup is more likely to expect individual products. This implies
that we need to make a decision dynamically about grouping based
on the query and matched results. This problem can be approached by
introducing heuristic merchandising rules to analyze the structure of
the results and matching attributes and make a grouping decision. For
instance, we can choose to replace individual products by a collection
only if a collection is generally consistent with the query, that is, all or
almost all products in the collection and the collection-level attributes
match the query. Consider the example in Figure 4.28. The query white
cup is likely to match individual products or collections that include
only white cups, but not dinnerware sets with plates, bowls, or cups of
different colors. Consequently, we present a user with a search results
list that contains mainly individual products. On the other hand, the
query white dinnerware is likely to produce a different result. We can
258 search
expect a significant number of dinnerware sets that mainly consist of

white items that match the query term white and are attributed with
the term dinnerware at the collection level. These dinnerware sets are
generally consistent with the query, so we can include them into the
search results list as collections, not as individual products.
type: bowl
query: white cup color: white
type: plate
type: cup type: dinnerware color: white
color: white
type: cup
color: red
type: cup
query: white dinnerware color: white
type: plate
type: dinnerware color: white
search results internal logic type: cup

color: blue
Figure 4.28: An example of dynamic grouping.
4.7 relevance tuning
We have worked through a number of matching, ranking, and signal

mixing methods that provide a wide range of controls for relevance
tuning and merchandising adjustments. The fine-tuning of all of these
controls to make them operate in concert can be a very challenging
task that requires the development of a methodology and optimization
methods. We have already made a step in this direction by introducing
a few search quality metrics, such as precision, recall, and discounted
4.7 relevance tuning 259
cumulative gain. Our next step will be to develop relevance tuning

methods with these metrics as optimization objectives.
Recall that the standard relevance metrics can be used to evaluate the
quality of a single search results list or the average quality of search re-
sults over a set of queries. The optimization objective is to maximize
the overall economic performance of a search service, so we should
optimize the relevance metrics for a set of queries with the largest con-
tribution to the total service profits. In other words, we can informally
express the overall economic performance of a search service as the
sum of contributions from individual search queries:
¸
Revenue Rpqq mpqq (4.94)
P
q Q
in which q is a query, Q is a set of possible queries, Rpqq is the av-

erage revenue of conversions attributed to query q, and mpqq is the
quality metric of the search results, which is assumed to be propor-
tional to the conversion rate. In practice, we are unlikely to be able
to optimize all possible queries, but we can select the most popular
and revenue-generating queries based on the historical data and use
only this set for optimization. The relevance optimization process can
then be organized as a continual review and improvement of the av-
erage performance over this set of queries; this process consists of the
following steps:
1. Service usage statistics are collected and analyzed to determine
the set of queries with the largest contribution to the service per-
formance. We refer to this set as benchmark queries.
2. The relevance metrics are calculated for the benchmark queries

to measure the overall performance of the search service.
3. The search results for each benchmark query are manually ana-
lyzed, and the search algorithms are tuned to improve the rele-
vance metrics.
4. The new scoring configuration is first tested on a subset of real

users and then applied permanently.
The processes above can be continually repeated to receive users’
feedback on changes in scoring algorithms and to keep the set of bench-
mark queries up to date. The relevance metrics calculation and search
algorithms tuning can be viewed as bottlenecks in the programmatic
pipeline because both steps require substantial human involvement to
make relevance judgments and rework scoring formulas. We can at-
tempt to bridge this gap by developing methods that can automatically
260 search
tune scoring formulas and assess the relevance of search results by an-
alyzing user behavior and interactions with the search service. We will
devote the next sections to a discussion of these two topics.
4.7.1 Learning to Rank
The primary goal of a search service is to rank documents according to

their relevance to a given query and context. It is intuitively clear that
this problem is closely related to classification or regression problems –
given a certain query, one needs to accurately predict the relevance grade
or rank of a document. A search results list can then be constructed
by sorting all of the documents according to their predicted grades.
This problem, commonly referred to as learning to rank, has been in-
tensively researched by academics in the field of information retrieval
and by web search companies such as Yahoo and Microsoft. This has
resulted in a large number of research papers, industrial usage reports,
and test data sets that are available for evaluation and comparison of
learning to rank methods. Although the number of learning to rank
algorithms is quite high, many of them use similar feature engineering
techniques and objective functions that can be viewed as a common
framework for learning to rank. Also, some benchmarks, such as the
Learning to Rank Challenge organized by Yahoo in 2010, have demon-
strated that the advantage of the most sophisticated learning to rank
methods compared with the basic ones can be quite limited on real
data [Chapelle and Chang, 2011]. Taking these considerations into ac-
count, we choose to focus on a common framework for learning to rank
and consider one particular algorithm as an example. An extensive cat-
alog of learning to rank algorithms can be found in [Liu, 2009].
We can formally define the problem of learning to rank as follows.
There is a training set that contains Q samples, with each sample be-
ing a pair of a search query and the corresponding search results list.
A results list for query q contains mq documents, and each document
d in the list is attributed with a relevance grade yq,d . We assume that
y is a categorical variable that takes one of K values. For example, the
set of relevance grades can include five values: 1 – perfect, 2 – excellent,
3 – good, 4 – fair, and 5 – bad. If we have defined a function that con-
verts a pair of a query q and a document d into a feature vector xq,d ,
the training set can be naturally represented as a collection of feature
vectors and corresponding training labels:

xq,d , yq,d , q 1, . . . , Q
d 1, . . . , mq
(4.95)
In practice, one can create a training data set by fetching the results
lists for each query by using conventional search methods and setting
relevance grades by using expert judgment. The goal is then to learn a
ranking model that predicts the grade y from the input that consists of
a query and a document.
Similarly to other supervised learning problems, learning to rank
starts with feature engineering. As we have already mentioned, the
relevance grade is predicted for a document in the context of a certain
query, so a feature vector depends on both the document and the query.
More specifically, the following groups of features are typically used
in practice [Chapelle and Chang, 2011; Liu and Qin, 2010]:
document features This type of features contains the statistics
and attributes for the document, including the following:
• The basic document statistics, such as the number of terms.
These statistics can be calculated independently for each field
and for the entire document to produce several groups of
features.
• Product classification labels, such as product type, price category,
and so on.
• Dynamic attributes and web statistics. Examples of such features
include product sales data, user ratings, and newness.
• Web search implementation of learning to rank often includes
web graphs and audience-related features, such as the number
of inbound and outbound links for a web page. Although these
may have limited applicability in merchandising search, such
metrics can be valid candidates if available.
query features These features include various statistics associated

with a query. Similarly to the document features, this group can in-
clude several subcategories:
• The basic query statistics, such as the number of query terms.
• Query usage statistics, such as frequency of query and click-
through rate.
• Attributes derived from the result set associated with the query.
For example, a query can be attributed with a topic, such as
furniture, if most of the results belong to this category.
document–query features Features that depend on both the

query and the document. This is the most important category of
features and can include the following groups:
262 search
• Various statistics calculated for terms that the query and docu-
ment have in common. For example, this can be a sum or vari-
ance of term frequencies or inverted document frequencies for
common terms. These metrics can be calculated for each docu-
ment field, as well as the entire document.
• Standard text matching and similarity metrics, such as the num-
ber of common terms and TFIDF .
• Statistics related to user feedback. This includes different inter-
action probabilities, such as the probability of click (the share of
users who clicked on a given document at least once among all
users who entered a certain query), probability of the last click
(the share of users who ended their search on a given document),
probability to skip (the share of users who click on a document
below the given one), and so on.
The structure of the feature vector is summarized in Figure 4.29. The
total number of features in practical applications can reach several hun-
dreds.
Figure 4.29: Feature engineering for learning to rank.
The next step in the creation of a ranking model is to define the

loss function that will be used as a model training objective. Although
learning to rank is closely related to classification and regression prob-
lems, setting the loss function is not trivial because we are concerned
with the relative order of the documents in the results set, which is not
the same as standard classification or regression errors. One possible
approach is to define the loss function as a relevance measure, such
as discounted cumulative gain (DCG). Unfortunately, DCG is a non-

convex and non-smooth function, which can be a problem for many
supervised learning algorithms that are based on a gradient approach.
Although DCG is often used to evaluate the end quality of learning to
rank algorithms, most methods use different loss functions for training.
These functions typically fall into one of three categories:
pointwise The pointwise approach attempts to predict relevance
grades for each document independently and thereby reduces rank-
ing to a standard regression or classification problem. Consequently,
the overall loss function L0 is defined as a sum of the prediction
errors for individual grades:
¸
L0 L f xq,d , yq,d (4.96)
q,d

in which f xq,d is a predicted relevance grade and Lpq is a classi-
fication or regression loss. For example, a classification loss can be
defined by using an indicator function that equals zero if the predic-
tion is correct and one otherwise:

L f xq,d , yq,d I f xq,d yq,d (4.97)
The classification loss, however, is also non-convex and non-smooth,

so we might need to approximate it with some different function. We
will discuss the possible options later in this section in connection
with concrete learning to rank algorithms. Another alternative is to
use a regression loss, defined as
2
L f xq,d , yq,d yq,d f xq,d (4.98)
It can be shown that the DCG error is upper bounded by the classi-
fication and regression losses, and minimization of the loss functions
thus helps to optimize the DCG [Cossock and Zhang, 2006; Li et al.,
2007]. However, the pointwise approach has a major downside, re-
gardless of the choice of the loss function. The issue is that we are
concerned with the relative order of items in the search results list,
not in qualitative or quantitative estimates of individual grades. For
example, the pointwise approach does not recognize that we will per-
fectly rank a results list of four items with relevance grades (1,2,3,4),
even if the grades are predicted as (2,3,4,5). Consequently, we may
take a different view on the loss function to account for the relative
order of items.
The pointwise approach is used in a number of ranking algorithms,
including McRank [Li et al., 2007] and PRank [Crammer and Singer,
2001].
264 search
pairwise The pairwise approach attempts to overcome the limita-

tions of the pointwise methods by penalizing for pairs of documents
that are ranked in reverse order, instead of penalizing for incorrectly
predicted relevancy grades. Consequently, the overall loss function is
defined as a sum of the pairwise loss functions for all pairs of docu-
ments in the results list with different grades:
¸ m
¸q
L0 L f xq,i , f xq,j (4.99)
¡
q i,j:yq,i yq,j
The pairwise loss function is often defined based on the difference

between the predicted grades, so that documents ranked in reverse
order contribute to losses. For example, the function can be defined
as exponential loss:

L f xq,i , f xq,j exp f xq,j f xq,i (4.100)
The pairwise approach can also be considered as a classification prob-

lem, but, unlike pointwise classification, it aims to classify document
pairs (pairs in which the first document is more relevant than the sec-
ond document versus pairs in which the second document is more
relevant).
Examples of pairwise ranking algorithms include RankNet [Burges
et al., 2005], RankBoost [Freund et al., 2003], and RankSVM [Herbrich
et al., 2000].
listwise In the listwise approach, the loss function is defined based

on the entire results list. In other words, the listwise approach uses
document lists as “instances” in the learning, as opposed to the indi-
vidual documents or document pairs that are used in the pointwise
and pairwise approaches. The loss function has a fairly generic form
that takes a list of pairs of predicted and actual relevance grades:
¸
L0 L f xq,1 , yq,1 , . . . , f xq,mq , yq,mq (4.101)
q
The inner loss function L can be defined as a relevance measure, such

as the DCG or its smooth approximations. Listwise ranking methods
include AdaRank [Xu and Li, 2007] and ListRank [Cao et al., 2007].
Thus far, we have discussed how to prepare the training data and de-
scribed possible options for the loss function. The last step is to choose
a predictive model and train it to minimize the prediction loss. As
learning to rank is closely related to classification, many standard su-
pervised learning methods can be adapted for ranking. In particular,
the industrial experience of Yahoo and Microsoft demonstrated that

decision trees, neural networks, and their ensembles are especially ef-
ficient in practice [Chapelle and Chang, 2011; Burges, 2010]. We con-
clude this section with an overview of the McRank algorithm, which
takes advantage of boosted decision trees to predict relevance grades
[Li et al., 2007].
McRank is a pointwise learning to rank algorithm that reduces the

ranking problem to a multiple classification. As we discussed earlier,
the relevance grades are categorical variables with K classes:
yq,d P t 1, 2, . . . , K u (4.102)
Our goal is to build a classification model that estimates the proba-

bility of each class based on a feature vector:

pq,d,k Pr yq,d k | xq,d , k 1, . . . , K (4.103)
If these probabilities are estimated, the McRank algorithm sorts the

documents according to their expected relevance:
Ķ
rq,d k pq,d,k (4.104)

k 1
The classification model is created in McRank by using a gradient

boosting tree algorithm. As it is a gradient method, a smooth loss func-
tion is required. McRank chooses to use the following smooth version
of the classification error described by expression 4.97 as a loss func-
tion:
¸ Ķ
log pq,d,k I yq,d k (4.105)
q,d k 1
McRank uses a standard gradient boosting tree algorithm that it-
eratively constructs an ensemble of decision trees to minimize loss
function 4.105. The result is a model that estimates the probabilities
described in equation 4.103, which can be used to rank documents in
the search results list.
4.7.2 Learning to Rank from Implicit Feedback
Learning to rank provides powerful capability for automatic relevance

tuning that helps to avoid or simplify manual signal mixing. This ca-
pability is essential for programmatic systems. Learning to rank, how-
ever, relies on expert judgment to set the relevance grades used in
266 search
model training. This step often requires significant human effort and
also limits the system’s ability to self-tune dynamically. We can attempt
to work around this problem by inferring relevance grades automati-
cally based on user interactions with search results. For example, the
results that nobody clicks on are likely to be irrelevant. One possible
way to leverage this information is to incorporate it into feature vectors,
as we already did in the previous section. We could take a step further
and attempt to develop a method that learns the relevance grades from
implicit feedback.
Although it is intuitively clear that users tend to click on relevant
search results and skip irrelevant ones, user behavior can communi-
cate more sophisticated relevance relationships. For example, a user
can enter a search query, browse the results, click through some of
them, reformulate the query, and click through some of the new re-
sults. All queries and documents in such a scenario are related to a
single search intent, so relevance relationships can be established both
within a single search results list and across the queries. In this sec-
tion, we consider a feedback model that captures such relationships
by using several heuristic rules [Radlinski and Joachims, 2005]. This
particular model comes from academic research, although loosely sim-
ilar methods for learning from implicit feedback have been reported
by Yahoo [Zhang et al., 2015].
The model we will consider has two groups of relevance feedback
rules. The first group, illustrated in Figure 4.30, includes two rules that
are applied in the scope of a single search query. The first rule states
that, if a user clicks on some document in the result list, this document
is more relevant than all the documents above with regard to a given
query. This is based on the assumption that a user typically reads the
results from top to bottom. The second rule is based on empirical ev-
idence (including eye-tracking studies) that a user typically considers
at least the top two results in the list before taking an action. Conse-
quently, if a user clicks on the first document in the list, it is considered
more relevant than the second one (with regard to a given query).
The second group of rules is applied to query chains, that is, query
sequences that represent different formulations of the same search in-
tent. This first requires the detection of queries that belong to the same
chain. This problem is not trivial because a user can make multiple
queries with only one search intent but formulate it differently or can
make multiple unrelated queries in a search for completely different
products. The implicit feedback model that we consider approaches
this problem by building an additional classifier that predicts whether
a pair of queries belong to the same chain or not. The model is trained
with manually classified query pairs and uses features, such as time
Figure 4.30: Implicit feedback rules for documents within one results list.
interval between queries, the number of common terms, and the num-
ber of common documents in the corresponding result lists. Once the
queries are grouped into query chains, we can introduce four addi-
tional relevance rules that can be applied to pairs of search results lists.
All of these rules are based on the assumption that queries in the chain
express the same search intent and, hence, can be considered equiva-
lent.
The first two rules in this group are presented in Figure 4.31. They
repeat the two single-query rules that we considered earlier but with
regard to adjacent queries in a query chain. Consider a chain in which
query q1 is followed by query q2 . Rule 3 mirrors rule 1 by stating that
a clicked document in the result list for query q2 is more relevant than
the preceding skipped documents with regard to query q1 because
both queries are related to one search intent. Similarly, rule 4 mirrors
rule 2.
Figure 4.31: Implicit feedback rules for a chain of queries.

268 search
The last two rules are shown in Figure 4.32. These rules establish
relevance relationships between the documents from different results
lists in a query chain. Rule 5 states that documents that are viewed
but not clicked on in the results list for query q1 are less relevant
than the documents that are clicked on in the results set for query
q2 . This relevance relationship is established with regard to the earlier
query. Consistently with rules 1 and 2, documents are considered to be
viewed if they are above the clicked ones or right below the last clicked
document, like document d3 in Figure 4.32. Finally, rule 6 states that
documents clicked on in the later results list are more relevant than
the first two documents in the former list. This rule is based on the
assumption that a user analyzes at least the first two results in the list
before reformulating a query.
Figure 4.32: Implicit feedback rules for a chain of queries (continued).
All six rules are simultaneously evaluated against each query chain
in the training data set to produce relevance relationships in the form
di ¡q dj (4.106)
which means that document di is more relevant that document dj with

regard to query q. These rules can be directly used as training labels
for the pairwise learning to rank algorithms that we discussed earlier.
For example, the authors of the feedback model that we just described
used it in conjunction with the RankSVM algorithm to produce the
final ranking model [Radlinski and Joachims, 2005].
Implicit feedback provides an important signal that can be used for
automated relevance tuning. This signal can be mixed with organic
relevance scores produced by other methods like TFIDF so that flaws
in organic ranking can be corrected by a ranking model learned from
implicit feedback.
4.8 architecture of merchandising search services 269
4.8 architecture of merchandising search services
We conclude our journey through search methods with a review of the

high-level logical architecture for a merchandising search service illus-
trated in Figure 4.33. The goal of this section is to summarize the key
stages of data and query processing that we have discussed previously
without going deeply into the technical and implementation details.
Figure 4.33: The main logical components of a merchandising search service.
A search service can be viewed as a database that contains the cata-

log data and provides search operations on top of it. As user queries
are typically served in real time, a search engine needs to preprocess
and index catalog data to enable efficient querying. This produces a
set of indexes that are used by a query engine to match and score
documents against a query. Consequently, the most basic search ser-
vice consists of two components, an indexer and a query engine. The
indexer typically includes two major data processing steps: mapping
and indexing. The goal of the mapping step is to take various pieces of
incoming data and create documents with well-defined fields, values,
and hierarchical relationships. The mapping step typically includes a
content analyzer, which performs tokenization, stemming, and other
text normalization transformations. The documents created at the map-
ping step are then indexed to produce data structures that enable fast
query processing. The query engine implements the basic search oper-
270 search
ations, such as token matching, Boolean querying, or TFIDF scoring

on top of the indexes.
Many search techniques, such as stemming, shingling, and synonym
expansion, require certain transformations to be applied to both docu-
ment fields and queries. On the document side, these transformations
are often incorporated into the indexer and applied to the input data or
documents before the index is build. This is necessary because indexes
are constructed based on the actual tokens that are present in the doc-
uments, so many transformations cannot be efficiently applied at the
query time if indexes are created based on raw non-transformed data.
On the query side, the transformations are applied at the query time,
so both indexes and queries are cast to the same normal forms. Nor-
malization is not the only type of transformation applied to queries. As
we discussed earlier, some search methods, such as controlled precision
reduction, drastically transform the initial query or generate multiple
intermediate queries. The query transformation logic is often encap-
sulated into a query preprocessor, which decomposes the incoming
query into the basic primitives supported by the core query engine.
The transformation logic on the indexer and query preprocessor sides
must stay in synch so that the same stemming, shingling, or semantic
expansion algorithms are applied on both sides.
The outputs of the query engine are the matching documents and
their relevance scores calculated in accordance with the basic signal
mixing methods. These results can be transformed by the result post-
processor to apply grouping and merchandising rules that extend or
override the basic scoring. For example, the postprocessor can imple-
ment the boost-and-bury rules to elevate promoted products in the
search results.
An efficient merchandising search requires at least two processes
that work behind the scenes and complement the indexing–querying
pipeline we just described. The first one is relevance tuning, which
manages the relevance controls across all components of the indexing–
querying pipeline and makes them operate in concert. This can be a
manual process or a machine learning component that analyzes query
history and user feedback to optimize relevance algorithms. The sec-
ond process is product information management (PIM), which focuses
on the cleaning, preparation, and enrichment of the catalog data loaded
into the search engine. The quality and completeness of the input data
are critical for the quality of the search because the range and ac-
curacy of generated relevance signals directly depend on the range
and accuracy of the product attributes. For example, basic product de-
scriptions can be insufficient to properly serve queries such as gluten-
free soda or long sleeved dress. To make a search engine understand
4.9 summary 271
such queries, products should be meticulously attributed with nutri-

tion facts, sleeve types, and other specific properties. The total num-
ber of product fields indexed by a merchandising search engine can
reach hundreds. It is a significant challenge to obtain and manage this
metadata because different pieces of information can come from man-
ufacturers or third-party data providers or may be created internally.
A marketer can streamline this process by using a product information
management system and specialized tools for metadata generation and
quality checks. For example, some retailers use advanced image recog-
nition tools to derive or validate certain product attributes, such as a
dress type or color based on product images.
4.9 summary
• The purpose of search services is to fetch offerings that are relevant

to the customer’s search intent expressed in a search query or se-
lected filters. Search services solve the problem of product discovery,
which can be viewed as a particular case of targeting.
• The main components of the merchandising search environment in-

clude the user interface where search queries are entered and ranked
catalog items are displayed, the search engine that processes the
queries and ranks the items, and the relevancy tuning process that
optimizes the relevance controls that determine how queries and
items are matched.
• The main objectives of a search service include relevance, flexibility

of merchandising controls, and service quality. Relevance and mer-
chandising controls can be directly linked to the service profits.
• The main relevance metrics include the precision–recall and dis-

counted cumulative gain in the ranked search results.
• Merchandising controls can be used to improve relevance and

achieve additional business objectives, such as promotion of certain
products. Merchandising controls include methods like boost
and bury, filtering, canned results, page redirection, and product
grouping.
• Search service quality metrics include conversion rate, click-through

rate, time on a product detail page, query modification rate, paging
rate, retention rate, and search latency. Some of these metrics can be
used as objectives in automatic relevance tuning.
• A query processing flow can be viewed as a multistage process that

breaks down documents and queries into features, correlates these
272 search
two sets of features to produce relevance signals, and then uses the
signals to make ranking decisions.
• The basic toolkit of search techniques includes text preprocessing

methods (tokenization, stop words, stemming), token matching, and
Boolean search. The most basic scoring techniques are based on the
vector space model, which represents documents as vectors in a lin-
ear space where each dimension corresponds to an individual term.
TFIDF , a popular scoring method, refines the basic vector space
model and scores by using term frequency statistics.
• A real-life search service typically uses documents with multiple

fields, which can be scored separately to produce multiple relevance
signals. These signals can be mixed together with different signal
engineering techniques.
• Term-matching methods are not able to capture semantic relation-

ships, such as synonymy and polysemy. This limitation is addressed
by semantic analysis methods. From a search perspective, most se-
mantic analysis methods can be viewed as word embedding tech-
niques that map words, documents, or queries to vectors of real
numbers with certain semantic-related properties. The key seman-
tic analysis methods include latent semantic analysis, probabilistic
topic models, and contextual word embedding.
• Merchandising search often deals with structured entities and

specific precision–recall requirements that cannot be properly
addressed by generic search methods. Industrial experience shows
that good results can be obtained by using high-precision and
low-recall extensions of Boolean methods.
• Relevance tuning is a process of search quality metrics optimization

with respect to relevance control parameters. This problem is closely
related to classification and regression – given a query, one needs to
predict the rank of a document. The problem, however, differs from
standard classification; thus, specialized learning to rank algorithms
exist. The typical features used by learning to rank methods are doc-
ument statistics, query statistics, relevance signals, and implicit user
feedback.
• The main components of a merchandising search service include the

indexer, core query engine, query preprocessor, query postprocessor,
and relevance tuning modules.
5
R E C O M M E N D AT I O N S
The variety of products and services offered to customers is limited

by a number of factors including distribution and production costs.
A grocery store can carry only a certain number of products that is
limited by the available shelf space, a radio station can squeeze only
a certain number of songs into its daily schedule, and a theater can
put on only a limited number of performances. A seller can increase
capacity and expand the assortment, but the additional revenues cap-
tured by this expansion start to diminish after a certain point because
of the limited total demand. The marginal costs associated with the
expansion, however, may not diminish as rapidly as the revenues or
may not diminish at all, so the marginal costs eventually overrun the
marginal revenue, which makes further assortment expansion uneco-
nomical. Consequently, a seller usually has to focus on relatively pop-
ular products and offer only a limited variety of niche items.
The demand for niche products, however, does exist, which creates
the long tail in the product popularity histogram shown in Figure 5.1. In
practice, the total demand for such niche products can be comparable
to the total demand for popular products [Anderson, 2008]. These two
total demands correspond to areas D1 and D2 under the demand curve
in Figure 5.1. Moreover, the long-tail items can often be high-end prod-
ucts with higher margins than the popular mainstream items, which
makes the contribution of the long-tail products to the total profit even
more significant.
As we already mentioned, many traditional distribution channels,
such as brick-and-mortar stores, movie theaters, and radio stations,
have limited ability to utilize the long-tail demand because of distribu-
tion costs. The advancement of digital channels, however, has changed
the game dramatically. First, new media channels have almost elim-
inated the distribution costs for digital content and enabled the cre-
ation of online services with virtually infinite assortments. An online
video service, for example, can offer an almost unlimited and con-
273
274 recommendations
popularity / demand
Traditional Digital
popular
products
niche products
D₁ D₂
products
sorted by popularity
Figure 5.1: A typical product popularity histogram.
stantly growing variety of videos, including Hollywood movies, tele-

vision series from all over the world, and amateur films. Second, dig-
ital channels have enabled retailers and manufacturers of non-digital
goods to change the structure of their distribution costs and reach out
to more consumers. The customer base of a brick-and-mortar store is
limited to people who live in or visit that area, and so the variety of
demand is similarly limited. An online retailer that operates nation-
wide or worldwide faces much more diverse demand. Therefore, sell-
ers with enormous assortments, such as Amazon, have appeared and
prospered. The impressive increase in the assortment with a strong
emphasis on niche products challenges the old approaches to product
discovery because an average customer can browse only a tiny frac-
tion of available offerings that can be counted in millions. This need
for powerful product discovery services was one of the main drivers
behind the development of recommender systems.
Recommendation services, in contrast to search services, aim to pro-
vide the customer with relevant offerings when a search intent is not
or cannot be clearly expressed. In some cases, the search intent can-
not be explicitly expressed because the desired product properties are
difficult to formalize or codify. For example, a customer who searches
for music might have some sort of search intent determined by their
tastes, but this intent can be difficult to translate into formal criteria.
In other cases, a customer may be unaware of certain product types or
categories or simply forget or doubt their own needs. An amateur pho-
tographer, for instance, might not realize that the type of photography
they are interested in requires special lenses for the best results. Conse-
quently, in contrast to search services that have a query to work with
5.1 environment 275
and can score products according to it, a recommender system needs

to guess the purchasing intent based on indirect information such as
product ratings and customer purchase histories. This information can
then be used to calculate different similarity metrics that can be used
as an alternative to the query–product similarity scores computed by
search services. More specifically, a recommender system can leverage
the following similarities:
user similarities The purchasing intent of a given customer can
be inferred from the past behavior of similar customers. This
approach resembles the look-alike modeling that we discussed
previously.
product similarities Interactions with the merchandise and past

purchases can be used to determine the most relevant product
groups and categories for a given customer, so that more similar
products can be recommended.
context similarities The accuracy of recommendations can be in-

creased by using not only customer and product attributes but
also contextual information that carries additional signals about
the purchasing intent. For instance, a fashion retailer can recom-
mend very different products for the same customer depending
on the season.
Algorithmically, recommendation methods have a lot in common
with search techniques and also take advantage of the predictive meth-
ods that we used earlier for targeted promotions. In the rest of this
chapter, we will systematically describe recommender systems by start-
ing with the environment and economic goals and then diving deeper
into various recommendation methods.
5.1 environment
The basic settings for recommendation services are similar to those for
search services. Similarly to search services, the primary purpose of a
recommender system is to provide a customer with a ranked list of rec-
ommended items. These recommendations can be delivered through
different marketing channels. We will assume that the recommenda-
tions are requested by a channel in real time, which is typically the
case for websites and mobile applications, although some channels
such as email can have more relaxed requirements and allow recom-
mendations to be calculated offline. The basic inputs of a recommender
system, depicted in Figure 5.2, are, however, different from those of a
search service and include the following:
276 recommendations
• The vast majority of recommendation methods make the assump-

tion that customer ratings are available for catalog items. The rat-
ings can be explicitly provided by customers or derived from be-
havioral data, such as purchases and online browsing histories.
Each rating value represents the feedback of a certain customer
on a certain item measured on some scale. Note that a customer
can rate any item in the catalog, not only the recommended ones.
In other words, ratings capture customer feedback on catalog
items, not recommendations. We will discuss ratings and their
properties in greater detail in the next section.
Marketing Channels
request
Context / Criteria
Recommendations Ratings
Item 1 Item 1
Recommender
System Item 2
provide
...
...
Item k Item n
implicit / explicit ratings
load load
Content Data User Data
Figure 5.2: The main concepts of the recommendation service environment.
• Certain recommendation methods rely on content and catalog

data to calculate similarities between the items based on their at-
tributes. Similarly to search engines, this requires the integration
of the recommender systems with the source of the catalog data,
such as the product information management system.
• Some recommender systems can take advantage of additional

user data, such as online order histories or store transactions.
This information needs to be distinguished from the basic rat-
ing data: ratings describe individual user–item interactions, but
5.1 environment 277
external data such as order history can provide additional details

on how interactions are grouped into orders and so on.
• Both recommendation requests and customer ratings can be com-

plemented with contextual information, such as time, location,
or marketing channel. A recommender system can use the con-
textual data to improve the relevance of recommendations. For
example, rating time stamps can be used to account for seasonal
and temporal trends in customer tastes. A recommendation re-
quest or context can also include explicit criteria or customer pref-
erences that can be used to refine recommendations. For instance,
skincare product recommendations can be adjusted based on the
skin type (normal, dry, oily, etc.) specified in the customer’s pref-
erences.
Different families of recommendation methods can use different sub-
sets of data, and their strengths and weaknesses are largely determined
by the range of data that they can take advantage of.
5.1.1 Properties of Customer Ratings
Customer ratings are often considered the most important source of

information for making recommendations, so we need to take a closer
look at how ratings are defined and what the typical properties of
rating data are.
Ratings are typically represented as a matrix where each row cor-
responds to a user and each column corresponds to an item. Let us
denote the rating matrix as R rij , so that rij is a rating given by
user i for item j. In a recommender system that tracks m users and
contains a catalog of n items, R is an m n matrix. In practice, the
rating matrix is almost always incomplete – ratings are known only
for a subset of user–item pairs and the remaining elements are miss-
ing (unspecified). The known ratings are usually numerical values that
can be defined differently depending on the business domain, market-
ing channel, and source of data. We will distinguish the following two
cases:
ordinal The interface of a recommender system often allows the
user to express their preferences by choosing ratings from a dis-
crete set of numbers (e. g., 1, 2, or 3 stars) or a continuous range
(e. g., from -5 to +5). It is also typical to capture ratings as cat-
egorical variables with two (e. g., like or dislike) or more (e. g.,
bad, good, or excellent) categories and then map them to discrete
numerical values. Implicit feedback can also be expressed as or-
dinal values, but such values usually indicate confidence rather
278 recommendations
than preferences. For example, an implicit rating can indicate how

frequently a user purchases a certain item or how much time a
customer spent on a product details page on the website.
unary In many cases, a rating matrix does not capture the level of
affinity between a user and item but merely registers the fact of
an interaction. For example, many interfaces have only one Like
button, so a user can either like an item or provide no feedback at
all. Another typical example of unary ratings is implicit feedback
that registers interactions between users and items but does not
capture details, such as the number of purchases, although it can
be argued that the simple quantitative properties of this feedback
are important [Hu et al., 2008]. The elements of an unary rating
matrix can take only two values – either one or unspecified.
The rating values in matrix R are often attributed with contextual
information, such as the date and time that the rating was set, the
marketing channel used by the customer to provide the rating, and so
on. This information can be used by a recommender system to figure
out which rating values are most relevant for a given context.
An important observation about rating values is that a matrix with
explicit ordinal ratings also contains implicit feedback. The reason is that
users typically tend to rate the items they like and avoid the items they
find unattractive. For example, a user can totally avoid music of cer-
tain genres or products from certain categories. Consequently, which
items are rated is important, in addition to how they are rated. In other
words, the distribution of ratings for random items is likely to be dif-
ferent from the distribution of ratings for items selected by a user. This
means that a recommender system, strictly speaking, should not rely
on the assumption that the distribution of observed ratings is represen-
tative of the distribution of missing ones [Devooght et al., 2015]. As we
will see later, some advanced recommendation methods take this con-
sideration into account and infer the implicit feedback from the rating
matrix. In a more general case, a recommender algorithm can involve
two separate rating matrices for both explicit and implicit feedback.
The second important property of a rating matrix is sparsity. A rat-
ing matrix is inherently sparse because any single user interacts with
only a tiny fraction of the available items, so that each row of the ma-
trix contains only a few known ratings and all other values are miss-
ing. Moreover, the distribution of known ratings typically exhibits the
long-tail property that we have discussed earlier. This means that a
disproportionally large number of the known ratings correspond to a
few of the most popular items, whereas niche product ratings are es-
pecially scarce. This property can be illustrated by a well-known data
set published by Netflix, a video streaming service, which contains

movie ratings for about 500,000 Netflix subscribers – it turns out that
about 33% of ratings involve only the 1.7% of most popular items [Cre-
monesi et al., 2010]. The long-tail property is a challenge for the design
and evaluation of recommender systems because both recommenda-
tion algorithms and recommendation quality metrics tend to have a
bias toward popular items, which reduces the quality of recommenda-
tions for niche products.
The key business objectives of recommendation services are closely

related to the objectives of merchandising search. The main considera-
tions that we discussed earlier in the context of merchandising search,
such as relevance and merchandising controls, are generally applica-
ble to recommender systems. The major difference, however, is that
the search intent is not explicitly expressed and may not even exist.
This requires an extension of the basic objective of providing relevant
results because the notion of relevance becomes increasingly shaky as
the search intent loses its sharpness. Consequently, the standard set of
objectives for a recommender system is often defined as follows:
relevance The recommendations suggested to a user should be rel-
evant, in the sense that a user has a high propensity to purchase
the recommended items and rate them highly.
novelty A recommender system does not fulfill an explicit search re-

quest but rather advises users on available options. Consequently,
the recommender system should provide users with options that
are not already known to them; otherwise, the recommendations
can be perceived as trivial and useless. A typical example of this
problem is the recommendation of popular similar items that the
user is likely to be aware of. For instance, a user who bought one
of the Harry Potter books might be provided with recommenda-
tions to buy more books from this series, instead of other books
in the same genre.
serendipity Recommendations can help a user to discover products

that are unexpected and surprising, as well as novel. For exam-
ple, a user who buys books about machine learning might be pro-
vided with recommendations to buy more books on this topic –
although some of the recommended items might be novel to the
user, we probably cannot consider them as serendipitous. On the
other hand, a recommender system can try to guess the business
domain that the user is interested in and recommend a book on
280 recommendations
domain-specific analytical methods, such as customer analytics

or trading models, which can be serendipitous. Even better exam-
ples of serendipitous recommendations are those for items from
completely different categories. For example, a user who buys
Harry Potter books can be recommended to visit an amusement
park with a corresponding attraction or users who study Euro-
pean medieval epics such as Beowulf and The Song of Roland can
be recommended to buy related opera tickets. Serendipitous rec-
ommendations not only increase usability and conversion rates
but can also help to establish a new long lasting theme in the
relationship with a client.
diversity Finally, a list of recommendations suggested to a user

should be diverse to increase the chance of a conversion. A list of
recommendations that consists of very similar items, even if they
are relevant, novel, and serendipitous, might not be optimal.
Similarly to search services, the overall profit of a recommender sys-
tem can be defined in terms of product margins and quantities sold:
¸
Profit Quantity soldproduct Marginproduct
(5.1)
products
The relevance, novelty, serendipity, and diversity objectives are fo-

cused on increasing conversion rates and, consequently, quantity sold.
The recommendations produced in accordance with these objectives
can be re-ranked in the pursuit of additional merchandising goals, such
as the promotion of high-margin or seasonal products, with the aim of
improving the margin part of the equation.
5.3 quality evaluation
Our next step is to design quantitative metrics that can be used to eval-
uate the quality of recommender systems with regard to the objectives
defined in the previous section.
The quality of search results can generally be evaluated by using
expert judgement to score the relevance of items in the context of a
given query, but this approach has limited applicability for recommen-
dations because the context typically includes user profile data and,
hence, is unique for each user. This makes it challenging or impossi-
ble to manually score the quality of recommendations for every pos-
sible context. On the other hand, the rating matrix already contains
expert judgements provided by users in their own personalized con-
texts. Consequently, the recommendation problem can be viewed as a
5.3 quality evaluation 281
rating prediction problem (how to recommend items with the highest

predicted ratings for a given user), and the quality of the recommen-
dations can be measured by comparing the predicted ratings with the
actual ratings known from the rating matrix. From this perspective, the
recommendation problem is very close to classification or regression.
Recall that a classification/regression problem can be defined by us-
ing a matrix in which each row represents a data point and columns are
either features or responses. A prediction model is trained on data with
both features and responses, and it is then used to predict responses
based on features. This is illustrated in example 5.2, where data points
1–3 are used for training and the actual prediction is done for points
4–6. Furthermore, in order to train and tune the model, data points
with known responses are typically split into training, validation, and
testing data sets. The model is initially built by using the training set.
The responses of the validation set are then predicted and compared
to the actual values, and the model quality is evaluated. Based on the
evaluation results, the model can be re-built with different parameters
by using the training data and evaluated again. The testing data set is
used to finally assess the quality of the model at the very end of the
process.
Feature 1 Feature 2 Feature 3 Response

Data point 1 x11 x12 x13 y1
Data point 2 x21 x22 x23 y2

Data point 3
x31 x32 x33 y3
(5.2)
Data point 4
x41 x42 x43

Data point 5 x51 x52 x53
Data point 6 x61 x62 x63
The key difference for the rating prediction problem is that there are
no features and responses in the rating matrix. Known and unknown
ratings are mixed together without any particular structure, as illus-
trated in example 5.3, and the goal is to train the model with known
ratings to predict unknown ones. This task of filling in the missing en-
tries of a partially observed matrix is known as the matrix completion
problem.
Item 1 Item 2 Item 3 Item 4

User 1 r11 r13
User 2
r22 r24
(5.3)
User 3 r32 r33
User 4 r41 r43 r44
As with the standard classification problem, we need to split the

available data into training, validation, and testing sets in order to
282 recommendations
train the model and evaluate its quality. In the case of classification,
this is done on a row-by-row basis. For example, the available data in
matrix 5.2 can be split into three sets by assigning the first row to the
training set, the second row to the validation set, and the third row
to the test set. This approach does not work well for matrix comple-
tion because it implies that the model is trained on one set of users
and evaluated on another, which is not really the case. Instead, the rat-
ing matrix is typically sampled on an element-by-element basis. This
means that a certain fraction of known ratings is removed from the
original rating matrix to leave a training matrix, and the removed rat-
ings are placed into validation or testing sets, which are later used to
evaluate the quality of prediction.
By interpreting the recommendation problem as a rating prediction
problem, we can define several quality metrics that can be linked to the
business objectives. We will spend the next sections developing these
metrics.
5.3.1 Prediction Accuracy
The accuracy of rating predictions can be considered a measure of rel-

evance because it quantifies how well a recommender system can pre-
dict the user utility as estimated by the users themselves. To measure
the accuracy, we can choose from a wide range of metrics used in ma-
chine learning and information retrieval, including the search quality
metrics that we discussed earlier.
The first family of metrics we will consider are prediction accuracy
metrics that are widely used for the evaluation of classification and
regression methods. Let us denote the set of observed ratings ruj P R
as R and the test subset used for accuracy evaluation as T R. For each
rating in T , a recommendation algorithm produces an estimate pruj , so
the prediction error can be defined as follows:
euj pruj ruj (5.4)
The overall quality of the rating predictions can then be obtained by

averaging the pointwise prediction errors. There are several ways to
define this average metric. The first option is the mean squared error
(MSE) defined as
¸
MSE
1 2
|T | pu,jq P T euj (5.5)
The MSE metric is not always convenient because it operates with

squared error values, which cannot be directly compared to the orig-
inal rating values. We can fix this by defining the root mean squared
error (RMSE), which is measured in the same units as the original rat-
ings:
?
RMSE MSE (5.6)
Furthermore, the RMSE can be normalized to the range p0, 1q to ob-

tain the normalized RMSE (NRMSE):
NRMSE
RMSE
rmax rmin
(5.7)
The RMSE and its variations are widely used in practice for recom-
mender system evaluation as a result of their simplicity. However, the
RMSE and similar pointwise accuracy metrics have several important
drawbacks:
• As we discussed earlier, ratings typically follow a long-tail distri-
bution, which means that ratings are dense for the popular items
and sparse for those items from the long tail. This makes rating
prediction more challenging for the long-tail items relative to the
popular items and results in different prediction accuracies for
these two item groups. The RMSE does not differentiate between
these two groups and simply takes the average, so poor accuracy
for the long-tail items can be counterbalanced by high accuracy
for the popular items. To measure and control this trade-off, we
can calculate the RMSE separately for different item groups or
add item-specific weights into equation 5.5 to account for item
margins or other considerations.
• The goal of a recommender system is to predict how a user will

rate an item in the future based on the historical data. User tastes
and interests may change over time, so the system should be able
to recognize such temporal trends to make accurate predictions
about future behavior. The RMSE does not directly account for
this aspect of recommendations. The problem, however, can be
addressed with the proper design of a test set T . In order to test
the ability to predict future ratings, we can select test set T from
ratings R, not at random but in such a way that the training set
contains the older ratings and test set T contains the most re-
cent ones. This approach is somewhat inconsistent with the stan-
dard model evaluation methodology because training and test
sets constructed this way have different distributions, but it is a
valid practical technique that was used, for example, in the Net-
flix Prize, an open competition for the best collaborative filtering
algorithm held by Netflix, an online video streaming service, in
2006–2009 [Aggarwal, 2016].
284 recommendations
The consideration above is especially apparent in the case of rat-

ings obtained from implicit feedback. For example, if ratings are
defined based on purchasing events, the prediction of future rat-
ings effectively means the prediction of future purchases.
• A recommender system provides a user with a ranked list of

recommendations that is typically limited to the top K items. The
RMSE does not account for ranking and equally penalizes predic-
tion errors for items at both the top and bottom of the list. It can
be argued that algorithms with a very small difference in RMSE
can have a big difference in their lists of the top K items [Koren,
2007].
5.3.2 Ranking Accuracy
To measure the quality of the top K recommendations, we can leverage

the large set of methods and metrics that we developed for search
services. First, we should note that the concepts of precision and recall
are directly applicable to the top K recommendations problem. If Iu is
the subset of items in test set T that is positively rated (e. g., purchased)
by user u and Yu pKq is the list of top K items recommended to that
user, we can define the precision and recall metrics as functions of K:
precisionpKq
|Yu pKq X Iu |
|Yu pKq| (5.8)
recallpKq
|Yu pKq X Iu |
|Iu | (5.9)
For any given K, the quality of a recommender algorithm can be

measured in terms of these two metrics – precision is the percentage
of relevant recommendations in the list, and recall is the percentage of
items consumed from the set of available relevant items. Two recom-
mender algorithms can be compared in terms of precision and recall
averaged by users, just like search methods can be compared in terms
of these metrics. The number of recommendations in the list, how-
ever, is a critical parameter that influences the precision–recall trade-
off achieved by an algorithm. Short recommendation lists tend to miss
relevant items, whereas long lists tend to have a high percentage of
irrelevant ones. This trade-off can be visualized by using a precision–
recall curve, which we also discussed earlier in the context of search
services. The curve depicts the precision and recall values for different
values of K and makes it possible to see the range of precision–recall
trade-offs that the recommendation algorithm can achieve.
The disadvantage of a precision–recall curve is that it does not pro-

vide a single numerical metric that summarizes the quality of a recom-
mendation method. Fortunately, we have already introduced a range
of ranking quality metrics that can be adapted to provide just such a
summary view. For example, we can adapt the discounted cumulative
gain (DCG) for recommendations by using known ratings as relevance
grades. Recall that we defined the DCG for a list of K items as follows:
2 gi 1
Ķ
DCG
log2 pi 1q
(5.10)

i 1
in which gi is the relevance grade of the i-th item in the list. If test set
T contains ratings provided by m users, we can define the overall DCG
as the average of the DCGs for the recommendation lists for individual
users:
¸ 2rui 1
m̧
DCG
1
log2 pRui 1q (5.11)
m
P
¤
u 1 i Iu
Rui K
in which Iu is the subset of items in test set T positively rated by user

u, Rui is the rank of item i in the list of recommendations for user u,
and rui is the rating from set T provided by user u for item i, which is
used as an approximation of the relevance grade gi from equation 5.10.
Note that the inner sum in equation 5.11 simply iterates over the top K
recommendations with known test ratings for a given user. The other
standard information retrieval metrics, such as the normalized DCG
(NDCG) and mean average precision (MAP), can be reformulated in a
similar way.
5.3.3 Novelty
Recommendations are considered novel if the user is not aware of the

recommended items at the moment that the recommendation is pro-
vided. This information is not directly available in the rating matrix, so
it should either be gathered through real-life testing and user surveys
or somehow inferred from the rating matrix. Since real-life testing and
surveys are generally time and resource consuming, we can attempt
to design a novelty metric based on the rating matrix, which requires
certain assumptions to be made. One possible approach is to train a
recommendation algorithm with older ratings and evaluate it by using
more recent ratings, as illustrated in Figure 5.3.
The assumption is that a recommendation algorithm that tends to
predict items ranked or purchased immediately after the time bound-
ary tT of the training set provides lower novelty than an algorithm
286 recommendations
Figure 5.3: Evaluating the novelty of recommendations.
that tends to predict items ranked or purchased in the more distant

future because the immediately purchased items are more likely to
be known to the user. Consequently, the novelty metric can use time-
weighted scores to boost farsighted accurate predictions and penalize
shortsighted ones.
5.3.4 Serendipity
Serendipity is a measure of the extent to which recommendations are

both attractive and surprising to the user [Herlocker et al., 2004]. The
evaluation of serendipity is even more challenging than the evaluation
of novelty because this property is highly subjective and feedback infor-
mation typically provides no hint of the level of serendipity. It is, how-
ever, possible to develop heuristic serendipity measurement method-
ologies. One possible approach is to compare the recommendations
produced by an algorithm under evaluation with the recommendations
produced by some basic algorithm that is known to suggest trivial and
non-serendipitous items [Ge et al., 2010]. If the set of items recom-
mended to the user by the algorithm under evaluation is denoted as
Yu and the set recommended by the baseline algorithm is denoted as
0 , the measure of serendipity can be defined as
Yu
m̧ ¸
serendipity P z Yu0
1
I i
mK
Yu (5.12)
P Iu
u 1i
in which m is the number of users, Iu is the set of items in the test

set positively rated by the user, K is the number of recommendations
in the list, and Ipq is the indicator function that equals true if the item
belongs to the set Yu but not to Yu 0 . In other words, this serendipity
measure scores a recommender system based on the fraction of non-

trivial and relevant items in the recommendation list.
5.3.5 Diversity
Diversity is the ability of a recommender system to produce a list of

recommendations that consists of dissimilar items. High diversity is
generally preferable because it increases the chance that at least some

items in the list will be relevant for the user. High diversity can also be
preferable from the merchandising perspective because it contributes
towards cross-selling and wide catalog coverage.
We can leverage the similarity metrics developed earlier for search
services to measure the diversity. For example, we can calculate cosine
distances between item descriptions for all pairs of items in the list of
recommendations and estimate the diversity as the reciprocal of the
average distance.
5.3.6 Coverage
The goal of a recommender system is to predict the missing ratings

in the rating matrix. As we will discuss later, many recommendation
algorithms rely on item-to-item or user-to-user similarities computed
on the basis of the rating matrix, so it can be a challenge to predict
ratings for items or users who do not have many ratings in common
with other items and users. This makes it important to measure the
coverage provided by a recommender system, that is, the percentage
of users or items over which the system can make recommendations.
In certain cases, this percentage can be estimated based on the require-
ments imposed by a recommendation algorithm. For example, an algo-
rithm might require a user to have at least five ratings to be eligible
for recommendations. In a general case, a recommender system can
always predict a rating for any pair of user and item by simply provid-
ing a default or random value. This means that we might be interested
in tracking the reliability of predicted ratings (the probability that the
estimated value is accurate) and assessing the trade-off between the
coverage and accuracy by excluding users or items with the least reli-
able rating estimates from the accuracy evaluation.
An alternative view on coverage is the so-called catalog coverage [Ge
et al., 2010]. Catalog coverage is defined as the percentage of items
that are actually recommended to users. The problem is that a recom-
mender system might be able to estimate ratings for a wide range of
items but the top K recommendation lists presented to users might still
include almost the same recommendations, which actually equates to
poor coverage from the merchandising standpoint. We can define the
catalog coverage metric as the percentage of items that appear in at
least one recommendation list:

1 ¤
m
catalog coverage

Yu (5.13)
n
u 1
288 recommendations
in which n is the total number of items in the catalog. Equation 5.13

uses the union of recommendation lists over all users in the system to
estimate the number of covered items. An alternative approach is to
count the number of distinct items recommended over a large number
of real user sessions.
5.3.7 The Role of Experimentation
The metrics described above help to measure the quality of recommen-

dations from several important standpoints. The ultimate goal of a rec-
ommender system, however, is to uplift the revenue and conversion
rates. Although the discussed metrics provide a solid quality assess-
ment framework, they do not have a strong link to financial perfor-
mance indicators. This link can be established through practical exper-
imentation, multivariate testing, and uplift measurements.
5.4 overview of recommendation methods
Thus far, we have described the environment and data sources that a
recommender system is integrated with, its business objectives, and the
metrics that can be used to evaluate the quality of recommendations.
This provides a reasonably solid foundation for the design of recom-
mendation algorithms. This task can be approached from several dif-
ferent perspectives, and there are several families of recommendation
methods that differ in the data sources leveraged to make recommen-
dations (rating matrix, catalog data, or contextual information) and
the type of rating prediction model. Although we will methodically go
through all major categories of recommendation algorithms in the rest
of this chapter, it will be worthwhile to briefly review the classification
of recommendation methods and make a few general comments before
we dive deeper into the details of individual methods.
Recommendation methods can be categorized in a number of ways,
depending on the perspective taken. From the algorithmic and informa-
tion retrieval perspectives, recommendation methods are categorized
primarily by the type of predictive model and its inputs. The corre-
sponding hierarchy is shown in Figure 5.4. Historically, the two main
families of recommendation methods are content-based filtering and
collaborative filtering. Content-based filtering primarily relies on con-
tent data, such as textual descriptions of items, and collaborative filter-
ing primarily relies on patterns in the rating matrix. Both approaches
can use either formal predictive models or heuristic algorithms that
typically search for a neighborhood of similar users or items. In addi-
tion to these core methods, there is a wide range of solutions that can
5.4 overview of recommendation methods 289
combine the multiple core algorithms into hybrid models, extend them
to account for contextual data and secondary optimization objectives,
or make recommendations in settings where the core methods are not
optimal because, for example, of a lack of data for personalization. We
will thoroughly analyze each of these approaches in the following sec-
tions.
Recommendation methods
core methods additional methods
Content-Based
Hybrid
Filtering
Neighborhood-based Switching
Model-based Blending
Feature Augmentation
Collaborative
Filtering
Contextual
Neighborhood-based
Pre-filtering
User-based
Post-filtering
Item-based
Contextual modeling
Combined
Model-based Non-personalized
Regression
Most-popular
Latent factors
Neighborhood-based
Multi-objective
Figure 5.4: Classification of recommendation methods from the algorithmic per-

spective.
The hierarchy of recommendation methods will look different if we

focus on the usage scenarios rather than algorithmic and implementa-
tion details. One possible way to visualize this classification is shown
in Figure 5.5. We basically categorize all usage cases in two dimen-
sions: level of personalization, from non-personalized to segmented to
one-to-one personalized, and usage of the contextual information, from
290 recommendations
context unaware to context aware. The core recommendations methods

that we mentioned earlier, content-based and collaborative filtering, are
mainly focused on the personalized and context-unaware corner of the
rectangle and make recommendations based on the interaction histo-
ries of individual users, that is, the items that the user rated, browsed,
or purchased in the past. In the user interface, such recommendations
are often displayed in sections such as You might also like, Inspired by
your browsing history, or simply Buy it again. The recommendations can
be made even more personalized by taking into account contextual in-
formation such as the user location, day of the week, or time of day. A
restaurant recommendation system that suggests Restaurants near your
based on both a user’s interaction history and location is an exam-
ple of this class of recommenders. An alternative approach is taken
by non-personalized methods that rely on global statistics and item
properties, rather than personal profiles. In the user interface, these
recommendations are often presented in sections such as Most popular,
Trending, or New releases. Note that personalized and non-personalized
recommendations can be blended together in a number of ways. For
example, personalized recommendations selected based on the user’s
interaction history can be sorted by popularity, or, alternatively, the
most popular items can be selected within a category of products pre-
ferred by the user. Finally, non-personalized recommendations can also
be contextualized by user location or marketing channel attributes. For
instance, a product detail page can include the recommendation sec-
tions Frequently bought together and More like this, which are created in
the context of the currently browsed item and constrained accordingly.
Figure 5.5: Some typical recommendation usage cases and the corresponding
categories of recommendation methods.
5.5 content-based filtering 291
5.5 content-based filtering
The first family of recommendation methods that we will consider re-

lies mainly on the catalog data (content) and uses only a small fraction
of the information available in the rating matrix. This is the reason
why this group of methods is referred to as content filtering. The main
idea of content filtering is quite straightforward: take items that the
user positively rated in the past and recommend other items similar to
these examples, as shown in Figure 5.6. The important constraint, how-
ever, is that the measure of similarity is based on the item content 1 and
does not include behavioral data, such as information about items that
are frequently purchased or rated together by other users. This effec-
tively means that a content-based recommender system uses only one
row of the rating matrix – the profile of the user for whom the recom-
mendations are prepared. This limited usage of the rating information
is typically counterbalanced by a similarity function that uses a wide
range of carefully engineered item features. The recommendations are
then ranked according to their similarity scores and, optionally, the
rating values of corresponding items in the profile. For example, as-
suming that item 1 has the highest rating in the example shown in
Figure 5.6, that is, ru1 ¡ ru2 and ru1 ¡ ru3 , a candidate item similar
to item 1 can be ranked higher than candidate items equally similar to
items 2 or 3.
items
R
users profile u ru1 ru2 ru3
items the user likes item 1 item 2 item 3
content
candidate item similarity
score
Figure 5.6: Similarity-based approach to content filtering.
The above approach to content filtering is, however, somewhat lim-

ited because it inherently relies on a nearest neighbor model for rating
1 See Chapter 4 for a detailed discussion of such measures. One of the most common exam-

ples would be the TF IDF distance between textual descriptions discussed in Section 4.3.5.
292 recommendations
predictions – a candidate item is scored based on the average pairwise

similarity to the items in the user profile. A more generic and flexi-
ble interpretation of the content filtering problem is to consider it as
a machine learning task in which each item should be rated by us-
ing a regression or classification model trained on the user profile. In
other words, for each user, we create a dedicated profile model that can
predict whether the user will like a given item or not. The model is
trained by using items from the profile: each item rated by the user
is converted into a feature vector by a content analyzer and the corre-
sponding rating is used as a training label. Each candidate item is then
also converted into a feature vector and evaluated by using the profile
model, as illustrated in Figure 5.7. Finally, the list of recommendations
is created by ranking the candidate items according to the predicted
ratings and selecting the top ones. Note that the similarity-based ap-
proach depicted in Figure 5.6 is a particular case of the more generic
schema presented in Figure 5.7 derived under the assumption that the
profile model uses the nearest neighbor classifier.
Content-based
recommender profile
system Content items Profile
analyzer
item 1 - ru1
item item 2 - ru2
features ...
user
Profile ratings
candidate learner
item
feedback
item predicted
features ratings
Content Profile Recommendations
analyzer model
Figure 5.7: Rating prediction approach to content filtering.
Content filtering has both advantages and disadvantages in compar-

ison to other recommendation methods. The main strengths of content-
based recommender systems can be summarized as follows:
• User independence. Content filtering uses only the ratings provided
by the user for whom recommendations are prepared. This can be
a significant advantage when the total number of system users is
small or the total number of collected ratings is small, as can be the
case when a new recommender system is launched and the amount
of historical data is limited. This problem is generally known as
a cold-start problem. The second advantage of user independence is

the ability to select recommendations for users with unusual tastes,
which may not work well in systems that rely on user similarities
and, hence, tend to follow average tastes.
• New and rare items. A particular case of the cold-start problem is the
recommendation of new or rare items that have a few or no ratings.
A recommendation algorithm that heavily relies on ratings may not
be able to recommend such items, which negatively impacts the cov-
erage of the catalog. Content filtering is not sensitive to this problem
because it relies on content similarity. This capability is especially
important in the context of the long-tail property that we discussed
earlier – the catalog often contains many rare items that receive few
ratings, even over a long period. The same issue appears in domains
with a rapidly changing assortment, such as apparel stores, where it
can be difficult to accumulate enough statistics about items.
• Cross-category recommendations. It can be difficult or impossible to do

certain types of recommendations without taking the content into
account. For example, it can be challenging to recommend furniture
to people who bought clothes based only on purchasing patterns
because the number of users who buy from both categories simulta-
neously is likely to be small [Ghani and Fano, 2002]. As we will see
later, content filtering can be an appropriate solution for this kind of
task if the content features are engineered properly.
• Usage of catalog data. Content filtering is able to leverage catalog data,

which is an important source of information for recommendations.
This contrasts with some other families of recommendation algo-
rithms that ignore this data.
• Interpretability. Recommendations provided by content-based sys-

tems are often interpretable and explainable to the user because
item scores are based on certain content features. For example, a
movie recommender system can provide an explanation that an
action movie is recommended because a user highly rated action
movies in the past. Other families of recommendation methods can
produce results that are much more difficult to explain or interpret.
On the other hand, content filtering has a number of disadvantages
that are often the flip sides of its advantages:
• Feature engineering. It is clear that feature engineering plays a crit-
ically important role in content filtering, and the quality of rating
predictions heavily depends on the quality and completeness of cat-
alog data and feature design. Feature engineering for catalog items
294 recommendations
can be a challenging problem, even for textual product descriptions,

because of polysemy, synonymy, and other issues that we have dis-
cussed previously in the context of search services. The problem can
become far more challenging for images, movies, or music where
one needs either to manually attribute each item with content tags,
such as a music genre, or to leverage advanced deep learning meth-
ods. Content management and feature engineering are indeed the
key challenges in practical applications of content filtering, and we
will spend a significant amount of time later in this section dis-
cussing these aspects.
• New users. Although content filtering helps to solve the cold-start

problem for new items, it is not able to provide recommendations
for new users with empty profiles, which is the second variant of
the cold-start problem.
• Trivial recommendations. One of the most important drawbacks of con-

tent filtering is the tendency to produce trivial recommendations,
that is, recommendations that are neither novel nor serendipitous.
This property is a direct result of content-based scoring, which fa-
vors closely related items, such as books from the same series.
We will continue this section with a more thorough analysis of
content-based recommendation methods. We will first consider two
specific examples of profile models and will then discuss a few
advanced feature engineering methods developed for recommender
systems in the retail domain.
5.5.1 Nearest Neighbor Approach
As we have already mentioned, one can build a content-based recom-

mender system by using the k nearest neighbor (kNN) algorithm as a
profile model. Let us flesh out the details of how this approach can be
implemented. First, let us denote the set of items rated by user u as Iu .
We can also assume that each item j is represented as a document dj
with one or more attributes or fields, similarly to the situation we had
for search services. Consequently, the set of items Iu corresponds to
the set of documents Du . For each candidate item i, we can calculate
similarity metrics between its document representation di and each
of the documents in Du . Let us denote k documents in Du with the
highest similarity to di as
! )
dui ui
1 , . . . , dk Du (5.14)
These documents are the k nearest neighbors of di with regard to

some similarity measure. The user rating for item i can then be pre-
dicted as the average rating of nearest neighbors:
ķ
p
rui k1 ru dui
t (5.15)

t 1
in which ru pdq is the rating of an item that corresponds to docu-

ment d. This estimate can be further refined by weighting the ratings
according to the similarity score:
ķ
p
rui k1 ru dui
t sim di , dui
j (5.16)

t 1
The similarity measure is typically calculated by using the tech-

niques that we developed for search services. One popular approach
is to use the basic vector space model: the textual fields of the item
documents are preprocessed by using stemming and stop words,
each document field is represented as a vector of terms, the distance
between the corresponding fields is calculated according to the
TFIDF model, and then the scores for different fields are combined
into the final similarity score by using some signal mixing function.
This is exactly the same as document scoring in search services; the
only difference is that the TFIDF similarity score is calculated for a
pair of documents, rather than a document and query.
The second popular option is to use the latent topic model instead
of the basic vector space model: each document field is represented
as a vector in the space of latent topics by using latent semantic anal-
ysis (LSA) or latent Dirichlet allocation (LDA), the distance between
the corresponding fields is calculated as a cosine distance between the
corresponding vectors in the latent topics space, and the scores for dif-
ferent fields are combined together into the final score. Once again, this
approach is almost identical to the LSA and LDA search methods that
we discussed earlier. LDA-based scoring is sometimes considered to
be superior to LSA and basic TFIDF models [Falk, 2017]. It has been
successfully used in some major industrial systems, such as The New
York Times’s article recommendation engine [Spangher, 2015]. On the
other hand, some reports show that LSA can outperform LDA in cer-
tain applications, such as movie recommender systems [Bergamaschi
et al., 2014]. Clearly, these results strongly depend on the used data
sets and quality evaluation methodologies.
296 recommendations
5.5.2 Naive Bayes Classifier
The second approach to content filtering that we will consider was

developed for book recommendations [Mooney and Roy, 1999]. In con-
trast to the nearest neighbor regression, this method does not rely on
heuristic similarity metrics but uses the Naive Bayes classifier, a stan-
dard algorithm for text classification, as a profile model to predict the
ratings.
First, let us assume that each item in the catalog has multiple textual
attributes. For example, a book may have attributes such as title, au-
thors, synopsis, published reviews, customer comments, related titles,
and related authors. We choose to apply tokenization, stemming, and
stop words to the attributes and then model each attribute as a bag of
words, that is, a vector in which each element corresponds to a word
and the value is the number of word occurrences in the attribute text.
Consequently, each item is represented as a document with multiple
fields, and each field is a bag-of-words model of the corresponding
attribute.
Next, we need to create a profile model. Recall that the ultimate goal
of a content-based recommender is to rank catalog items for each user
in order of preference. We can approach this problem by building a
binary classifier that estimates two probabilities: the probability that
an item would be rated positively by a user and the probability that
it would be rated negatively. The ratio between these two probabilities
indicates whether an item would be rated positively rather than nega-
tively, and, consequently, we can use it as a ranking score for the rec-
ommended items. Let us assume that a user rates items on a discrete
scale from 1 to rmax , and all ratings below rmax {2 are interpreted
as negative, whereas ratings above rmax {2 are interpreted as positive.
For example, with a rating scale from 1 to 10 stars, ratings of 1–5 are
considered negative (dislike) and ratings of 6–10 are positive (like).
Recall that the main idea of the Naive Bayes text classifier is to es-
timate the probability of a document d being in a certain class c by
using conditional probabilities of document words w occurring in a
document of class c, under the assumption that these conditional prob-
abilities are independent. This approach can be expressed by using the
following formula:

¹
Pr cj | d Pr Pr wi | cj
cj
Pr pdq
(5.17)
P
wi d
in which cj is the document class, which in our case is either nega-

tive c0 or positive c1 , Pr cj is the empirical probability of class cj in
the training data (the fraction of documents that belong to the class),

and Pr wi | cj is the empirical conditional probability of word wi
(the fraction of documents of class cj that contain this word). This ba-
sic Bayes rule needs to be extended to accommodate the multiple fields
that we have in each item document. By assuming that each item doc-
ument has F fields and each field fqm is a text snippet that contains
fqm words, we can rewrite formula 5.17 for the posterior class proba-
bility as follows:

Pr cj ¹F ¹
Pr cj , d Pr pdq
Pr wi | cj , fm (5.18)

m 1 wi P fm
The ranking score of the item can then be estimated as
Pr pc1 | dq
scorepdq
Pr pc0 | dq
(5.19)
and items in the recommendation list can be sorted accordingly, from

the highest to lowest scores.
Our next step is to estimate the probabilities in formula 5.18 based
on the user profile, that is, items rated by the user. As we discussed
earlier, user ratings are set on a scale from 1 to rmax . If a user has
rated Q items, let us map each rating to two auxillary variables, for the
positive and negative classes, respectively:
αq1 r rq 1 1 , q 1, . . . , Q (5.20)

max
αq0 1 αq1 , q 1, . . . , Q (5.21)
in which rq is the original rating in the user profile. Note that we

consistently omit the user subscript u in all equations because the al-
gorithm uses only the profile of the active user. The class probability
can then be estimated as follows:
¸
Q
Pr cj Q1 αqj , j 0, 1 (5.22)

q 1
The conditional probabilities of words should be estimated for each

field of an item document separately. If the number of times word wi
occurs in the field m of document q is denoted as nqm pwi q, we can
estimate the conditional probability of the word as
¸ nqm pwi q
Q

Pr wi | cj , field m αqj , m 1, . . . , F (5.23)

q 1
Ljm
298 recommendations
in which Ljm is the total weighted length of the texts in field m for
class j:
¸
Q

Ljm αqj fqm , m 1, . . . , F (5.24)

q 1
The length of a field is defined as the number of words in its bag-

of-words representation. These estimates enable us to evaluate the pos-
terior document class probabilities from equation 5.18 and, finally, to
score the items. Note that the probabilities Pr pdq can be ignored be-
cause they cancel each other out in scoring formula 5.19.
example 5.1
Let us consider a numerical example to better understand how the

Naive Bayes classifier can produce recommendations and what the lim-
itations of this approach are. Consider an online bookstore where each
book in the catalog is represented by a document with two fields: title
and synopsis. Let us take a user profile with two books rated on a scale
from 1 to 10 and create a profile model from it. The original profile is
as follows:
Book 1
Title: Machine learning for predictive data analytics
Synopsis: Detailed treatment of data analytics applications
including price prediction and customer behavior
Rating: 8
Book 2
Title: Machine learning for healthcare and life science
Synopsis: Case studies specific to the challenges of
working with healthcare data
Rating: 3
We first convert the textual fields into bags of words and remove
stop words to obtain the following:
title1 : (machine, learning, predictive, data, analytics)

synopsis1 : (detailed, treatment, data, analytics, applications,
including, price, prediction, customer, behavior)
title2 : (machine, learning, healthcare, life, science)
synopsis2 : (case, studies, specific, challenges, working,
healthcare, data)
Next, we calculate the class proximity values according to formu-

las 5.20 and 5.21.
α11 8 9 1 97 (5.25)
α10 1 α11 29 (5.26)
α21 3 9 1 29 (5.27)
α20 1 α21 79 (5.28)
We use these values to estimate the class probabilities according to

expression 5.22. As the user liked the first book (the rating is 8 out of 10)
and disliked the second one (the rating is 3 out of 10), the probabilities
are equal:
Pr pc0 q pα10 α20 q

1 1
2 2
(5.29)
Pr pc1 q pα11 α21 q
1 1
2 2
Calculating the weighted lengths of fields according to formula 5.24,
we get
L0, title α10 | title1 | α20 | title2 | 29 5 7

9
5 5
L1, title α11 | title1 | α21 | title2 |5

(5.30)
L0, synopsis α10 synopsis1 α20 synopsis2
23
3

L1, synopsis α11 synopsis1 α21 synopsis2
28
3
Finally, we can estimate the conditional probabilities of words by
using expression 5.23. As an illustration, let us estimate the conditional
probability of the word price in the field synopsis, given the negative
class:
Pr pprice | c 0, field synopsisq
n1, synopsis ppriceq n2, synopsis ppriceq

α10 L0, synopsis
α20
L0, synopsis (5.31)
92 23/3
1 7
0
9 23/3
692
300 recommendations
By estimating the probabilities for all combinations of words, classes,

and fields, we obtain the results shown in table 5.1. This table is effec-
tively the profile model that would be calculated by a recommender
system in advance, stored, and then used to rank recommendations in
accordance with formulas 5.18 and 5.19.
Title Synopsis
c 0 c 1 c 0 c 1
analytics 0.044 0.160 0.029 0.083
applications 0.000 0.000 0.029 0.083
behavior 0.000 0.000 0.029 0.083
case 0.000 0.000 0.100 0.024
challenges 0.000 0.000 0.100 0.024
customer 0.000 0.000 0.029 0.083
data 0.044 0.160 0.130 0.110
detailed 0.000 0.000 0.029 0.083
healthcare 0.160 0.044 0.100 0.024
including 0.000 0.000 0.029 0.083
learning 0.200 0.200 0.000 0.000
life 0.160 0.044 0.000 0.000
machine 0.200 0.200 0.000 0.000
prediction 0.000 0.000 0.029 0.083
predictive 0.044 0.160 0.000 0.000
price 0.000 0.000 0.029 0.083
science 0.160 0.044 0.000 0.000
specific 0.000 0.000 0.100 0.024
studies 0.000 0.000 0.100 0.024
treatment 0.000 0.000 0.029 0.083
working 0.000 0.000 0.100 0.024
Table 5.1: Example of a Naive Bayes profile model.
Table 5.1 provides a few useful insights into the logic of the Naive
Bayes recommender. First, we can see that the words that occur in both
positively and negatively rated items cancel each other out. For ex-
ample, both books have the words machine learning in their titles, so
each of these words has an equal probability value for the positive
and negative classes, and these values will cancel each other out in ra-
tio 5.19 used for scoring. Second, we can see that words present in the
attributes of the negatively scored book (e. g., healthcare) are interpreted
as negative signals, in the sense that their probability values for the neg-
ative class are higher than those for the positive class. Similarly, some
other words (e. g., behavior) are interpreted as positive signals. In prac-
tice, this interpretation may or may not be correct. In our example, the
user has disliked the second book about machine learning for health-
care. We do not really know the reason – it may be that this particular
book is not well written or the healthcare domain is not relevant for
the user. If it is assumed that the user has rated the book after purchas-
ing and reading it, the first explanation is probably more likely than
the second one because the user was almost certainly aware that the
book was about the healthcare domain and chose it deliberately. The
Naive Bayes model, however, interprets the word healthcare as a nega-
tive signal, so all books with this word in the title will be scored lower.
This limited ability to differentiate between the quality and relevance
of content is one of the major shortcomings of content-based filtering.
As we will see later, collaborative filtering takes a different approach
to the problem and places more emphasis on item quality signals.
N
5.5.3 Feature Engineering for Content Filtering
The main idea of content filtering is to create a regression or classifica-

tion model that scores the item content. It is clear that this approach
requires careful feature engineering because the quality of classifica-
tion heavily depends on the available item attributes and the quality
of their modeling. Trivial attributes often lead to trivial or meaning-
less recommendations, whereas thoughtfully designed features can en-
able a recommender system to accurately predict user decisions. Let
us illustrate this problem with an example from the apparel domain
[Ghani and Fano, 2002; Ghani et al., 2006]. Consider a user who pur-
chased and rated several garments, such as dresses, blouses, or coats.
We can expect that an average product information system can provide
some basic information about each of these items, such as a product
category, price, and color. A basic recommender system that calculates
similarities between items by using such attributes is likely to recom-
mend items from the same categories, in the same price range, and
of the same color. Although this approach does not necessarily pro-
duce bad recommendations, it is flawed in at least two ways. First,
customer choice is heavily influenced and driven by personality, atti-
tudes, and lifestyle. Customers are likely to think about garments in
terms of style and functionality, choosing between casual and formal,
sporty and businesslike, conservative and flamboyant. Both users and
garments can be described in terms of such psychographic features, and
recommendations can be made based on a user’s affinity to certain
styles and attitudes. A recommender system that uses only the basic
attributes, such as product category and price, is typically not able to
recognize these latent affinities. Second, certain types of recommen-
dations, such as cross-category recommendations, are fundamentally
difficult to do if nothing but basic features are available. One of the
reasons is that the layout of product attributes can be different in differ-
302 recommendations
ent categories, so it can even be difficult to define the similarity metric

of a profile model that can compare and score items across categories.
For example, a large department store may sell apparel, kitchenware,
and furniture. It can, however, be challenging to recommend furniture
based on purchases in the clothing department because garments and
furniture have different attributes and different attribute semantics. For
example, the size attribute has a totally different meaning for a dress
relative to that for a bed. Again, psychographic features can be a so-
lution because users who buy conservative garments can be recom-
mended to consider furniture of a conservative style and so on.
One observation that we can make is that textual product attributes,
such as product name, description, and customer reviews, often carry
an implicit signal about the psychographic characteristics of a product.
Merchandisers who create product names and descriptions, for exam-
ple, deliberately choose certain words such as stylish, sexy, or luxury to
make a product appealing to certain audiences. This fact can be lever-
aged to measure product affinity to certain psychographic properties
and to define the corresponding product features. These features can
then be used to train and evaluate a profile model. More specifically,
we can use the following method to extract implicit psychographic fea-
tures [Ghani and Fano, 2002]:
• We first define the set of product features to be extracted by using
domain knowledge. Table 5.2 provides examples of such features
for the apparel domain.
Feature Name Feature Values
The most appropriate age group for

Age group the product: juniors, teens, mature,
etc.
A typical product usage scenario:
Functionality evening wear, sportswear, business
casual, business formal, etc.
Degree of formality From informal to very formal
From very conservative clothes,
Degree of conservativeness
such as gray suits, to flashy clothes
Degree of sportiness From sloppy or formal to athletic
Degree of trendiness From timeless classic to fast fashion
From unknown or unappealing
Degree of brand appeal
brands to highly appealing brands
Table 5.2: Example of psychographic features for the apparel domain.

5.6 introduction to collaborative filtering 303
• Next, a subset of items is manually labeled by domain experts

according to the features defined in the previous step. This set
is used to train classification models that predict psychographic
labels based on textual product descriptions and other standard
attributes, such as brand name and size. For example, one can use
a Naive Bayes classifier to identify words in product descriptions
that indicate a high degree of formality or trendiness.
• The classification models are used to label the remaining items.

This enables merchandisers to tag even very large and frequently
changing catalogs with a limited amount of manual effort.
It is worth noting that some recommender systems completely rely
on manually created item attributes. One of the best known examples
is Pandora Internet Radio, a music streaming and recommendation
service. Pandora uses professional music analysts to manually attribute
each song in their catalog with 450 features, such as Child or Child-
like Vocal and Melodic Articulation Clean-to-Dirty. This analysis, known
as the Music Genome Project, requires significant effort because the
catalog contains hundreds of thousands of songs and the classification
of a single song takes about 20 minutes [Walker, 2009]. This metadata,
however, is Pandora’s core asset and a major competitive advantage in
a market of music discovery services.
5.6 introduction to collaborative filtering
Content-based filtering attempts to approximate user tastes and judge-

ments by a similarity measure between the contents of catalog items.
The fundamental shortcoming of this approach is that human tastes
cannot be easily expressed in terms of basic product attributes, so man-
ual product tagging and advanced feature engineering is often needed
to achieve good results. On the other hand, the rating matrix captures
a lot of information about user tastes and judgements. Indeed, each
known rating can be interpreted as a manually set product attribute,
and, consequently, the collection of ratings and other feedback data
from users can be viewed as a crowdsourcing approach to product
tagging with psychographic features. Content filtering does not fully
leverage this valuable information because recommendations are cre-
ated by using a single profile model. This line of thinking leads us to
a different family of recommendation methods known as collaborative
filtering.
The term collaborative filtering was coined by developers of Tapestry,
a recommender system for news and articles created at Xerox PARC in
1992 [Goldberg et al., 1992; Terry, 1993]. In the context of Tapestry, col-
304 recommendations
laborative filtering meant that users were able to provide feedback on

news emails and prioritize incoming emails based on the feedback pro-
vided by other users. It was essentially a functional feature for email
filtering, not a recommendation algorithm. The idea of prioritizing rec-
ommendations based on the feedback provided by other users, how-
ever, had a lot of traction and resulted in the development of new
recommendation methods that leveraged this approach and the wide
adoption of these methods in industrial recommender systems, includ-
ing the major ones developed by Amazon and Netflix. The meaning of
collaborative filtering has also transformed and become more focused
on making rating predictions based on information available in the
rating matrix. Collaborative filtering in this newer, narrower sense is
a pure matrix completion problem. Consequently, collaborative filter-
ing methods are essentially matrix completion algorithms that use the
rating matrix as the only input. Under the hood, collaborative filter-
ing uses a predictive model to capture interactions between users and
items known from the rating matrix and to predict a rating for a given
pair of user and item based on how similar users rated similar items
in the past.
The key advantage of collaborative methods is that they are capa-
ble of making recommendations based only on the patterns and sim-
ilarities available in the rating matrix, without any additional infor-
mation about catalog items. This makes these methods much more
universal than content-based filtering, which requires domain-specific
knowledge, data, and feature engineering effort. Even more impor-
tantly, collaborative filtering implicitly accounts for the psychographic
profiles of users and items because the ratings capture human tastes
and judgements. This helps to produce non-trivial recommendations.
On the other hand, collaborative filtering has a number of weaknesses:
• Rating sparsity. A collaborative recommender system requires a
sufficient number of known and trustworthy ratings. If the rating
matrix is too sparse, it can be difficult or impossible to build a
reliable rating prediction model.
• New users and items. Collaborative filtering predicts ratings for a

given user or item based on the known ratings for this user or
item. This means that collaborative filtering does not work well
for new users and items or for users and items with very few
known ratings. Collaborative filtering is, therefore, more vulner-
able to the cold-start problem than content-based filtering, which
leverages content information in addition to the rating data.
• Popularity bias. Content filtering makes recommendations based

on typical patterns in the rating matrix, so it is inherently bi-
ased towards popular items and standard choices. This limits the
ability to produce non-trivial recommendations and recommen-
dations for users with unusual tastes.
• Product standardization. Although collaborative filtering is able to

recognize that some items are frequently bought together, it es-
sentially treats each item as an opaque independent entity. This
can create certain modeling challenges for products with a com-
plex internal structure, such as garment variants in different sizes,
customizable products, or products that are upgraded over time.
• Domain knowledge. As we mentioned earlier, one of the key ad-

vantages of collaborative filtering is its ability to work with an
abstract rating matrix without any assumption about the nature
of items and their attributes. Although this is generally true, col-
laborative filtering methods may need to make certain domain-
specific assumptions. For example, a collaborative recommender
system may or may not assume that customer tastes change over
time and, consequently, may or may not account for rating re-
cency.
Collaborative filtering algorithms are usually categorized into
two subgroups: neighborhood-based and model-based methods.
Neighborhood-based (also known as memory-based) methods predict
unknown ratings for a given user or item by using the nearest neigh-
bor approach, that is, by finding the most similar users or items and
averaging known ratings from their records. Model-based methods
go beyond the nearest neighbor approach and use other, usually
more sophisticated, predictive models. Although the nearest neighbor
algorithm can also be considered as a sort of predictive model and,
hence, the boundary between the two categories is a bit blurry, it
makes sense to separate them because of the high practical importance
of neighborhood-based methods. We will analyze both approaches in
detail in later sections.
5.6.1 Baseline Estimates
Most practically used collaborative filtering models are able to capture

relatively complex interactions between users and items by recogniz-
ing sophisticated patterns in the rating matrix. Before discussing these
models, it is important to note that observed ratings typically follow
a few simple but strong patterns that can be captured by using a rel-
atively simple model. This basic model can produce baseline rating
estimates that can be used as building blocks in more advanced collab-
orative filtering methods.
306 recommendations
A typical rating matrix exhibits strong user and item biases – some
users systematically give higher (or lower) ratings than other users,
and some items systematically receive higher ratings than other items
[Koren, 2009; Ekstrand et al., 2011]. This can be explained by the fact
that some users can be more or less critical than others, and items, of
course, differ in their quality. We can account for these systematic user
and item effects by defining the baseline estimate for an unknown user
rating rui as
bui µ bu bi (5.32)
in which µ is the overall average rating in the rating matrix R, bu

is the observed deviation of user u from the average, and bi is the
observed deviation of item i from the average. In practice, the user and
item biases have strong effects, and, consequently, the baseline estimate
defined by equation 5.32 has substantial predictive power. Although
this model captures only the average user and item effects, it can help
to absorb the biases and isolate the signal that represents user–item
interactions that can be captured by more specialized models.
The biases µ, bu , and bi can be estimated one after the other as the
average residual errors of the previous estimate. This means that we
first compute µ, and then estimate item biases bi as follows:
¸
bi | U1 | prui µq (5.33)
i iPU
i
in which Ui is the set of users who rated item i. Then, the user biases
are estimated as
¸
bu | I1 | prui µ bi q (5.34)
u uPI
u
in which Iu are items rated by user u. The estimates calculated with

formulas 5.33 and 5.34 can be unstable in the case of a sparse rating
matrix where only a few known ratings are available for a user or item.
The stability of the estimates can be improved by adding regularization
parameters λ1 and λ2 as follows:
¸
bi | U 1| prui µq
i λ1
i P Ui
¸ (5.35)
bu | I |1 prui µ bi q
u λ2
u P Iu
The regularization parameters decrease the magnitudes of biases bi
and bu when the user or item has few ratings, so the baseline estimate
described by equation 5.32 becomes closer to the global mean and less
dependent on the unreliable bias estimates.
The bias parameters can be estimated more accurately by solving the
following least squares problem [Koren, 2009]:

¸ ¸ ¸
min prui µ bi bu q λ b2u b2i (5.36)
bi , bu
i,u PR u i
in which R is the training set of known ratings and λ is a regular-

ization parameter. This is a straightforward optimization problem that
can be solved by using standard methods, such as stochastic gradient
descent. The advantage of this approach is that expression 5.36 can
be easily modified and extended to include additional constraints and
variables, such as temporal effects.
example 5.2
To illustrate the baseline estimates, as well as other collaborative fil-

tering methods, we will need a sample rating matrix. We have chosen

to use an example with movie ratings that became very popular after
the Netflix Prize. It is worth noting, however, that all of the collabora-
tive filtering methods considered in this chapter are domain agnostic,
in the sense that they do not have any dependency on the nature of
items, so the movie names are provided solely for the purpose of con-
venient reading and can be replaced by products from other domains
like groceries or apparel.
Our example is a rating matrix with six movies and six users, as
shown in table 5.3. The ratings are given on a 5-star scale where 1 is
the lowest possible rating and 5 is the highest. The matrix contains 28
known and 8 missing ratings, so it is very dense in comparison with
real-life rating matrices, where more than 99% of possible ratings can
be missing. One can easily notice a few patterns in the example matrix.
First, we can see that the first three users apparently like drama movies
(Forrest Gump, Titanic, and The Godfather) more than action movies (Bat-
man, The Matrix, and Alien). The last three users, in contrast, apparently
prefer action to drama. Next, we note that user 3 generously gives high
ratings to most movies, whereas user 2 seems to be more critical. We
would expect that a good collaborative filtering model would be able
to recognize such patterns and make adequate predictions.
308 recommendations
Forrest The The

Titanic Batman Alien
Gump Godfather Matrix
User 1 5 4 — 1 2 1
User 2 4 — 3 1 1 2
User 3 — 5 5 — 3 3
User 4 2 — 1 4 5 4
User 5 2 2 2 — 4 —
User 6 1 2 1 — 5 4
Table 5.3: Example of a rating matrix for a movie recommendation service.
Let us now calculate the baseline estimates for the missing ratings.
Calculating the global average and bias values by using formulas 5.33
and 5.34, we get
µ 2.82
p0.02 0.42 0.42 0.82 0.51 0.02q
bi (5.37)
bu p0.23 0.46 1.05 0.53 0.44 0.31q
Although quite trivial, we should note that these coefficients cor-

rectly capture the fact that user 3 tends to give high ratings (bias +1.05)
and the Batman movie has generally low ratings (bias –0.82). Substitut-
ing this result into baseline estimation formula 5.32, we obtain the final
rating predictions shown in the table 5.4. Note that this result gener-
ally does not match our intuitive expectations about affinities between
users and movie genres.
Forrest The The

User 1 5 4 [ 2.16 ] 1 2 1
User 2 4 [ 2.78 ] 3 1 1 2
User 3 [ 3.85 ] 5 5 [ 3.05 ] 3 3
User 4 2 [ 3.78 ] 1 4 5 4
User 5 2 2 2 [ 1.55 ] 4 [ 2.35 ]
User 6 1 2 1 [ 1.68 ] 5 4
Table 5.4: Example of baseline rating estimates.
N
5.7 neighborhood-based collaborative filtering 309
5.7 neighborhood-based collaborative filtering
The neighborhood-based approach to collaborative filtering relies on

a similarity measure between users or items defined in terms of the
ratings that two users or two items have in common. These two cases,
user-based and item-based similarity, are distinct but have a lot in com-
mon.
Let us first consider the user-based approach illustrated in Figure 5.8.
Recall that the goal of a recommender system is to predict the ratings
that a certain user would give to different catalog items and then to
create a list of recommendations by selecting and ranking items with
the highest predicted ratings. If we assume that the user has already
rated some items in the catalog, so that the corresponding row of the
rating matrix contains some known values, we can try to find more
users who rated the same items and did it with a similar sentiment,
in the sense that these users mainly like the items positively rated by
the given user and dislike the negatively rated ones. The key idea of
neighborhood-based collaborative filtering is that such users are likely
to have the same tastes and preferences as the given user, so their past
ratings can be used to predict the given user’s future ratings. Conse-
quently, the system can recommend items that have not been rated by
the given user but have been positively rated by at least some neigh-
borhood users. The predicted ratings for these items can be obtained
by averaging the ratings provided by the neighborhood users.
items
R
most
users similar
users
known ratings
Figure 5.8: User-based collaborative filtering.
The item-based approach shown in Figure 5.9 is structurally similar

to what we just described, but users (rows) are replaced with items
(columns). In order to predict ratings for a given item, we first find
items that are similar to the given one, in the sense that they are sim-
310 recommendations
ilarly rated by the same users. Next, the rating that a given user will
give to this item is estimated based on the ratings that the user gave
to other items in the neighborhood. Once again, the key assumption is
that a user who positively rated a few items in the past will probably
like items that are rated similarly to these past choices by many other
users.
items
R
users
known
ratings
most similar items

Figure 5.9: Item-based collaborative filtering.
Both user-based and item-based collaborative filtering require the

definition of a similarity measure, between either users or items, and
some rating averaging method. Although the user-based and item-
based approaches are structurally similar, they can use different mea-
sures and there are many different variants of similarity and rating
averaging formulas for each approach. We delve deeper into these de-
tails in the next sections.
5.7.1 User-based Collaborative Filtering
The two main steps of neighborhood-based collaborative filtering are

the selection of users or items that should be included in the neigh-
borhood and the rating prediction by averaging neighbors’ ratings. In
the case of the user-based approach, this means that we need to define
two key functions: a similarity measure for users and a rating aver-
aging formula. There are many different variants of these functions
described in the academic literature and industrial reports, so we will
first describe one the most basic and well-known options and will then
discuss possible variations and improvements.
To define the similarity measure, let us consider two users u and v

who have rated items Iu and Iv , respectively. The set of items rated by
both users is
Iuv Iu X Iv (5.38)
The similarity measure can then be defined based on this itemset.

The most common way is to use the Pearson correlation coefficient,
defined as follows [Herlocker et al., 1999]:
°
simpu, vq b°
P Iuv prui b
i µu q prvi µv q
2 °
i PIuv prui µu q i PIuv prvi µv q
2 (5.39)
in which µu and µv are the average user ratings:

¸
µu | I1 | rui (5.40)
uv i P I
uv
Note that formula 5.40 computes the average user rating over the
set of common items Iuv , as required by the definition of the Pearson
correlation coefficient. Hence, this value is not constant for a given user
u but is unique for each pair of users. In practice, however, it is quite
common to use the global average rating for user u computed over all
items Iu rated by this user [Aggarwal, 2016].
The similarity measure allows us to identify k users who are most
similar to the target user u. The size of the neighborhood k is a param-
eter of the recommendation algorithm. As our goal is to predict the
rating for user u and item i by averaging the ratings given for this item
by other users, we select not simply the top k most similar users but
the top k most similar users who have rated item i. Let us denote this
set of peers as Skui . This set can include less than k users if the rating
matrix does not contain enough ratings for item i or enough peers of
user u with commonly rated items. The rating can then be estimated
as a similarity-weighted average of the peer ratings:
°
p µu v P°Skui simpu, vq prvi µv q
rui
v P Skui
| simpu, vq | (5.41)
Formula 5.41 exploits the idea of separating the user biases from the
interaction signal, which we discussed earlier in the section devoted
to baseline estimates. The global user averages µu and µv are consid-
ered biases and are initially subtracted from the raw ratings, then the
interaction signal is estimated as a product of similarity measures and
mean-centered ratings, and finally the user bias µu is added back to
account for the preferences of the target user.
312 recommendations
Let us now discuss several alternatives of formulas 5.39 and 5.41.

Most of these variations are heuristic adjustments that can help to im-
prove the accuracy of estimates and computational stability in practical
applications [Su and Khoshgoftaar, 2009; Breese et al., 1998]. First, we
consider several choices for the similarity measure:
basic similarity functions. The Pearson correlation coefficient
is known to be a good option for the similarity measure, but
one can choose to use other metrics including cosine similarity,
Spearman rank correlation coefficient, and mean-squared differ-
ence. For example, the cosine similarity between two users can
be defined as follows:
°
rui rvi
P Iuv b
simpu, vq
i
b° ° (5.42)
i PIuv rui i PIuv rvi
2 2
These alternatives are generally considered inferior to the Pear-

son correlation and, as we will see shortly, it can be more effi-
cient to calculate similarity coefficients by using regression anal-
ysis, rather than heuristically selecting and tuning similarity func-
tions.
discounted similarity. The similarity measure is calculated

based only on the items rated by both users. The reliability of
this estimate depends on the number of items the users have in
common, so it is often beneficial to adjust the similarity measure
according to the item support (the number of ratings that two
items have in common) [Koren, 2008]:
sim 1 pu, vq
| Iuv |
| Iuv | λ simpu, vq (5.43)
By increasing the regularization parameter λ, we can shrink the

unreliable similarity coefficients with low support.
inverse user frequency. The Pearson correlation coefficient, as

well as many other standard similarity metrics, treats all items
in the set Iuv equally. One can argue that this approach is not
optimal because some items can be more indicative than others.
For example, if two users positively rate some rare or niche item,
it is probably a stronger indicator of their similarity than if they
both like a very popular commodity item. This idea is inspired
by the TFIDF text similarity metric, in which each word is
weighted proportionally to its inverse frequency. More formally,
we can define the inverse user frequency (IUF) for item i as

wi log | Um | (5.44)
i
in which m is the total number of users and | Ui | is the number

of users who rated item i. This weight can then be inserted into
the Pearson correlation formula as follows:
°
b° P Iuv wi prui
i µu q prvi µv q
b°
i PIuv wi prui µu q i PIuv wi prvi µv q
2 2 (5.45)
default ratings. The similarity measure for users u and v is typi-

cally calculated over the intersection of rated items Iu X Iv . Given
a sparse rating matrix, this intersection is typically small, which
decreases the reliability of the estimates. An alternative is to cal-
culate the similarity over the union of rated items Iu Y Iv , rather
than the intersection, by inserting some default (neutral) rating
for items that have only been rated by one of the users.
Rating prediction function 5.41 mixes together known mean-
centered ratings by using the similarity scores as weights. There
exist several alternative options for this function that differ in rating
centering and weighting logic:
standard score (z-score). In statistics, the standard score (also
known as the z-score) of a data point x is the deviation of this
point from the mean measured in standard deviations:
xµ
zpxq (5.46)
σ
The standard score can be viewed as a normalization of the
mean-centered values by standard deviation. We can use stan-
dard scores as an alternative to mean-centering in the rating
prediction formula. First, we calculate the standard score of a
rating in the context of a given user as
rui µu
z prui q (5.47)
σu
in which σu is the standard deviation of the known user ratings:
d°
P Iu prui µu q
2
σu i
| Iu | 1
(5.48)
The rating prediction function can then be redefined with the

standard scores instead of mean-centered ratings:
°
p µu σu
v P Skui simpu, vq z prvi q
°
rui
v P Skui
| simpu, vq | (5.49)
314 recommendations
Similarly to mean-centering, we first transform the original rat-

ings by applying formula 5.47 and then perform an inverse trans-
formation by multiplying the result with the standard deviation
of user ratings σu and adding back the mean µu . The standard
score approach boosts ratings from users with low rating vari-
ance and decreases the weight of ratings from users with high
variance.
baseline-centering. Another alternative to mean-centering is

baseline-centering, which uses the baselines estimates given by
formula 5.32, instead of the average user ratings [Koren, 2008].
The baseline-centered prediction formula can be defined as
follows:
°
P° simpu, vq prvi bui q
p
rui bui v Sk
ui
| simpu, vq | (5.50)
v P Sk
ui
amplification. The standard rating prediction function uses sim-

ilarity scores as weights. We can choose to apply some nonlin-
ear transformation to these raw scores to amplify certain ratings.
For example, the following transformation boosts high similarity
scores for the amplification parameter ρ ¡ 1:

sim 1 pu, vq simpu, vq simρ1 pu, vq

(5.51)
neighborhood selection. The quality of recommendations gen-

erally depends on the number of users k included in the neigh-
borhood. Some studies show that the accuracy of rating predic-
tions can monotonically increase as the size of the neighborhood
increases, given an advanced, properly designed, and well-tuned
rating prediction model [Koren, 2008]. The incremental improve-
ment delivered by the neighborhood expansion gradually dimin-
ishes, of course, and becomes negligible after a certain point.
Some other studies, however, show that neighborhoods that are
too large can negatively impact the accuracy of recommenda-
tions made by basic neighborhood-based methods because of
noise added by neighbors with low similarity [Herlocker et al.,
1999; Bellogín et al., 2014]. The size of the neighborhood k can be
limited by either an empirically determined constant value or a
similarity threshold that causes neighbors with small similarity
scores to be filtered out.
example 5.3
Let us briefly illustrate the user-based approach with a numerical ex-

ample by using the data in table 5.3. The first step is to estimate the

pairwise similarities between the users according to formula 5.39. This
generates the following similarity matrix:
User 1 User 2 User 3 User 4 User 5 User 6

User 1 1.00 0.87 0.94 -0.79 -0.59 -0.78
User 2 0.87 1.00 0.87 -0.84 -0.81 -0.88

User 3
0.94 0.87 1.00 -0.93 -0.87 -0.91 (5.52)

User 4
-0.79 -0.84 -0.93 1.00 0.85 0.95

User 5 -0.59 -0.81 -0.87 0.85 1.00 0.94
User 6 -0.78 -0.88 -0.91 0.95 0.94 1.00
We can see that the first three users are positively correlated to each
other and negatively correlated with the last three. The similarity ma-
trix allows one to look up a neighborhood of the top k most similar
users for a given target user and mix their ratings to make a prediction.
For example, let us predict the missing rating for user 1 and The Godfa-
ther movie by assuming the neighborhood size k 2. The most similar
neighbors are users 3 and 2 who rated The Godfather as 5 and 3, respec-
tively. Applying rating prediction formula 5.41, we get the following
estimate:
simp1, 3q pr33 µ3 q simp1, 2q pr23 µ2 q
r13 µ1
p
| simp1, 3q | | simp1, 2q |
0.94 p5 4.00q 0.87 p3 2.20q
(5.53)
2.60 0.94 0.87
3.50
By repeating this process for all missing ratings, we obtain the results
shown in table 5.5. Note that these estimates look more intuitive and
accurate than the baseline estimates in table 5.4.
N
In practice, user-based recommendation methods can face scalabil-

ity challenges as the number of system users approaches tens and hun-
dreds of millions. If the neighborhood for the target user is determined
at the time that the recommendations are requested, similarity metrics
need to be calculated between the target user and all other users in
the system online. If the neighborhoods are computed in advance, the
amount of computations will be a quadratic function of the number
of users. In addition, the target user profile might not be available in
advance (e. g., the browsing history within the current web session).
One possible way to work around this limitation is to switch from user
similarities to item similarities, as we will discuss in next section.
316 recommendations
Forrest The The

User 1 5 4 [ 3.50 ] 1 2 1
User 2 4 [ 3.40 ] 3 1 1 2
User 3 [ 6.11 ] 5 5 [ 2.59 ] 3 3
User 4 2 [ 2.64 ] 1 4 5 4
User 5 2 2 2 [ 3.62 ] 4 [ 3.61 ]
User 6 1 2 1 [ 3.76 ] 5 4
Table 5.5: Example of ratings predicted with the user-based collaborative filter-
ing algorithm.
5.7.2 Item-based Collaborative Filtering
The main idea of the item-based approach is to recommend items that

are similar to items that have been positively rated by the target user
by computing a similarity measure between the items based on the
known ratings from other users. This approach is similar to content-
based recommendations in the sense that recommendations are made
based on the similarity between items, although the nature of the sim-
ilarity measure is totally different. At the same time, this approach is
structurally similar to user-based collaborative filtering because both
methods are based on the notion of neighborhoods and, consequently,
use the same algorithmic framework [Linden et al., 2003; Sarwar et al.,
2001].
In order to predict the rating that user u would give to item i, the
item-based recommender system first determines the neighborhood of
item i, that is, the set of k most similar items. To calculate the similarity
measure between two items i and j, let us denote the set of users who
rated item i as Ui , the set of users who rated item j as Uj , and the
users who rated both items as
Uij Ui X Uj (5.54)
The similarity can then be measured as the Pearson correlation coef-

ficient between the vectors of common ratings:
°
simpi, jq b°
u P Uij prui bµi q ruj µj
° 2 (5.55)
P
u Uij prui µi q2 u UijP ruj µj
in which µi and µj are the average ratings for items i and j, respec-
tively. This formula is the same as the Pearson correlation for users in
equation 5.39; the only difference is that users (rows) are replaced by
items (columns). All items rated by user u can then be sorted by their
similarity to given item i, and the top k most similar items can be se-
lected from this list. Let us denote this neighborhood of item i as Qk
ui .
Note that the neighborhood includes only the items rated by the target
user u and not the most similar items in the catalog in general, so set
Qkui converges to Iu as k increases. The rating can then be predicted as
a weighted average of ratings of the top k most similar items by using
mean-centered ratings as inputs:
°
p µi j P °Qkui simpi, jq ruj µj
rui
j P Qkui
| simpi, jq | (5.56)
Similarly to the user-based approach, formulas 5.55 and 5.56 are

just the basic options, which can be adjusted and improved with the
various techniques we discussed earlier, such as discounted similar-
ity and weight amplification. Most of these techniques are applicable
both to user-based and item-based methods. For example, one can use
baseline-centered input ratings instead of mean-centered ones to im-
prove the accuracy of predictions [Koren, 2008]:
°
PQ k simpi, jq ruj buj
p
rui bui j
°ui
| simpi, jq | (5.57)
j P Qk
ui
5.7.3 Comparison of User-based and Item-based Methods
The item-based approach was proposed years after the first user-based
methods appeared, but it has quickly gained popularity because of bet-
ter scalability and computational efficiency [Linden et al., 2003; Koren
and Bell, 2011]. One of the key advantages is that the total number of
items m in the system is often small enough to precalculate and store
the m m item similarity matrix, so the top k recommendations can be
quickly looked up for a given user profile. This enables a more scalable
architecture for the recommender system: the heavy computations re-
quired to create the similarity matrix are done in the background, and
the recommendation service uses this matrix to make recommenda-
tions in real time. Although the same strategy can be applied to user-
based methods, it can be very expensive or completely impractical in
recommender systems with a high number of users. Finally, some stud-
ies found that item-based methods consistently outperform user-based
approaches in terms of prediction accuracy for certain important data
sets, such as Netflix data [Bell and Koren, 2007].
At the same time, it should be noted that user-based approaches are
able to capture certain relationships that might not be recognized by
item-based methods [Koren and Bell, 2011]. Recall that the item-based
318 recommendations
approach predicts rating rui based on the ratings that user u gave to
the items similar to i. This prediction is unlikely to be accurate if none
of the items rated by the user is similar to i. On the other hand, it may
still be the case that the user-based approach will identify users similar
to u who rated i, so the rating can be reliably predicted. As we will see
later, some advanced recommendation methods combine item-based
and user-based models to take advantage of both methods.
The ratio between the number of users and items is one of the key
considerations in the choice of approach. In many retail applications,
the item-based approach is preferable because the number of items is
smaller than the number of users. The number of items, however, can
exceed the number of users in some other domains. For example, an
article recommender system for researchers can benefit from the user-
based solution because the total number of all research articles ever
published reaches many hundreds of millions, whereas the research
community that uses the system is relatively smaller [Jack et al., 2016].
5.7.4 Neighborhood Methods as a Regression Problem
The neighborhood methods that we considered in the previous sections

rely on a heuristic rating prediction function that estimates unknown
ratings as weighted averages of known ones. To make the statement
about weighted averages more explicit, let us note that the user-based
and item-based rating prediction functions 5.41 and 5.56 essentially
have the following forms:
¸
p
rui wuv rvi (user-based)
(5.58)
v P Skui
¸
p
rui wij ruj (item-based)
(5.59)
j P Qkui
in which wuv and wij are the weights proportional to the user-based
and item-based similarities, respectively. In other words, weights w are
interpolation weights. This consideration quite naturally leads to the
question of how the optimal weights can be determined by means of re-
gression analysis instead of the heuristic similarity-based weights. The
regression analysis can be applied to both user-based and item-based
models, as well as hybrid methods that combine these two models, so
we will start with the arguably more practical item-based approach
and then discuss alternative options [Bell and Koren, 2007].
5.7.4.1 Item-based Regression

The item-based methods predict the ratings for item i by averaging rat-
ings from similar items according to expression 5.59. Input ratings ruj
can be taken directly from the original rating matrix or the matrix can
be preprocessed to center the ratings by subtracting the global average,
item average, or baseline predictions. In the case of centered input rat-
ings, the output rating prui is also centered, so the global average, item
average, or baseline prediction needs to be added back in at the end.
In order to solve a regression problem for the rating interpolation
weights, let us first consider a hypothetical case in which the rating
matrix is so dense that all users except u have rated both item i and all
its neighbors Qkui , as shown in Figure 5.10.
items
R j i
users u
rvi
wij
Qui
Figure 5.10: Item-based nearest neighbor regression.
The optimal interpolation weights for item i can then be determined

by solving the following least squares problem (for each item sepa-
rately):
¸ 2
min rvi prvi (5.60)
w

v u
Inserting the rating prediction function 5.59 into this, we get

2
¸ ¸
min rvi wij rvj (5.61)
w

v u j P Qk
vi
320 recommendations
Rearranging the terms, we can rewrite this problem in vector form

as
min rT r 2bT w wT Aw (5.62)
w
in which A is a k k matrix defined as

¸
Ajh rvj rvh (5.63)
v u
b is a k-dimensional vector defined as
¸
bj rvj rvi (5.64)

v u
and rT r is a constant term with respect to w:

¸
rT r r2vi (5.65)

v u
If we take the gradient of quadratic form 5.62 with respect to w and

equate it to zero, we get the following linear system of equations:
Aw b (5.66)
In a more realistic case with a sparse rating matrix, we would expect

that only a few users had rated both item i and all its neighbors Qkui .
Consequently, the estimates of the corresponding elements of A and
b can be more or less reliable, depending on the number of known
ratings. We can account for this by using the discounted similarity
technique that we discussed earlier and shrinking the estimates by the
corresponding support:
¸
Ajh U1

rvj rvh
jh v P Ujh
¸ (5.67)
bj U1

rvj rvi
ij v P Uij
in which Uij is the set of users who rated both item i and item j. It
is worth noting that one can compute and store all possible elements
of matrix A in advance, that is, compute the m m item correlation
matrix according to expression 5.63 for all values of 1 ¤ j, k ¤ m and
then use those values to quickly assemble the k k matrix A and vector
b for a given item and target user.
One possible way to compute the optimal weights is to solve
equation 5.66 numerically by inverting matrix A, but this is not the
only possible option. An alternative approach is to directly solve
problem 5.62 by using the gradient descent or some other generic
optimization method. The advantage of this approach is the ability

to add additional constraints and variables. For example, it has been
reported that prediction accuracy can be slightly improved when
weights w are constrained to be non-negative [Bell and Koren, 2007]:
min wT Aw 2bT w
w
(5.68)
subject to w¥0
Even better results can be obtained by adding more variables into

basic rating prediction formula 5.59 and jointly optimizing them. For
example, it has been shown that the following extension of the rating
prediction formula is a good practical choice for baseline estimates
[Koren and Bell, 2011]:
¸
p
rui µ bu bi wij ruj buj cij
(5.69)
j P Qkui
in which µ is the global rating average, buj are the baseline estimates,
and bu , bi , wij , and cij are the variables to be optimized. This expres-
sion can be inserted into least squares problem 5.60 and optimized with
respect to bu , bi , wij , and cij by using the gradient descent method.
In this case, one does not necessarily need to restrict the neighborhood
to the top k items and can use the entire Iu set, instead of Qk ui .
example 5.4
We conclude the discussion of item-based regression with a numerical

example that uses the movie rating matrix from table 5.3. We choose

to work with mean-centered ratings, so we first preprocess the input
matrix by subtracting the column average (i. e. , the average item rating)
from each element, which gives the result presented in table 5.6.
Forrest The The

Mean 2.80 3.25 2.40 2.00 3.33 2.80

User 1 2.20 0.75 — -1.00 -1.33 -1.80
User 2 1.20 — 0.60 -1.00 -2.33 -0.80
User 3 — 1.75 2.60 — -0.33 0.20
User 4 -0.80 — -1.40 2.00 1.66 1.20
User 5 -0.80 -1.25 -0.40 — 0.66 —
User 6 -1.80 -1.25 -1.40 — 1.66 1.20
Table 5.6: Example of ratings centered by using item averages.

322 recommendations
Next, we precompute the item correlation matrix A according to ex-

pression 5.63 for all values of 1 ¤ j, k ¤ m by using the ratings from
table 5.6 as inputs:

10.80 4.90 4.68 -5.00 -10.60 -8.04
4.90 6.75 6.80 -0.75 -4.50 -2.50

4.68 -3.32
A 6.80 11.20 -3.40 -7.20
-5.00 (5.70)
-0.75 -3.40 6.00 7.00 5.00

-10.6 -4.50 -7.20 7.00 13.33 8.20
-8.04 -2.50 -3.32 5.00 8.20 6.80
This precomputed matrix can be used to quickly assemble the sys-

tem of equations described by expression 5.66 for a given target user
and item. For example, with a neighborhood size of k 2, we pre-
dict the rating that user 1 would give to The Godfather by averaging the
ratings from Titanic and Forrest Gump, which are the two closest neigh-
bors of The Godfather in terms of Pearson similarity. Consequently, we
use the correlation values for these three movies (highlighted in bold
in matrix 5.70) to construct the following equation for the interpolation
weights:

6.75 4.90 w32
6.80
(5.71)
4.90 10.80 w31 4.68
By solving this equation, we get the weights w32 1.033 for Titanic
and w31 0.035 for Forrest Gump. The rating can then be predicted
as
p
r13 µ3 w32 r12 w31 r11
2.40 1.033 0.75 0.035 2.20 3.09
(5.72)
in which µ3 is the average rating for The Godfather movie and the in-
put ratings r12 and r11 are taken from table 5.6. Repeating this process
for all unknown ratings, we get the final results presented in table 5.7.
N
5.7.4.2 User-based Regression
The regression analysis framework we have just developed for item-
based methods can be applied to user-based models quite straightfor-
wardly. The input to the process is the rating matrix centered by user
averages (row average is subtracted from each element in the row) or
baseline predictions. The user-based variant of the least squares prob-
lem 5.60 can then be defined as
¸ 2
min ruj pruj (5.73)
w

j i
Forrest The The

User 1 5 4 [ 3.09 ] 1 2 1
User 2 4 [ 3.83 ] 3 1 1 2
User 3 [ 4.02 ] 5 5 [ 1.98 ] 3 3
User 4 2 [ 2.34 ] 1 4 5 4
User 5 2 2 2 [ 2.28 ] 4 [ 3.15 ]
User 6 1 2 1 [ 2.94 ] 5 4
Table 5.7: Example of ratings predicted by using item-based nearest neighbor

regression.
This problem needs to be solved for each target user u. Inserting the
rating prediction formula 5.58 into the previous equation, we get
2
¸ ¸
min ruj wuv rvj
(5.74)
w
j i v P Sk
uj
Optimal weights wuv can be determined by using the same meth-

ods as those used for the item-based approach – one can either solve a
linear system of equations or use generic optimization methods to op-
timize weights based on cost function 5.74. In comparison to the item-
based approach, user-based regression inherits all of the advantages
and disadvantages of user-based methods that we discussed earlier. In
particular, user-based methods are more challenging from a computa-
tional standpoint if there are many more users than items because one
needs to precompute the n n user correlation matrix, instead of the
m m item correlation matrix.
5.7.4.3 Fusing Item-based and User-based Models

One of the key advantages of the regression approach is the ability
to extend the rating prediction function with new terms and variables
that can be jointly optimized. We have already seen an example of such
an extension in expression 5.69 where we added new variables to the
basic item-based model to capture user and item biases. We can extend
this solution even further and combine the item-based and user-based
models into one rating prediction function:

p
rui µ bu bi
¸ pitemq
wij ruj buj

cij
j P Qkui
¸
wuv
puserq pr b q (5.75)
vi vi
v P Skui
324 recommendations
pitemq
in which wij
puserq
and wuv are two different sets of weights to be
learned. This model essentially sums a centered version of user-based
function 5.58 with item-based function 5.69 [Aggarwal, 2016; Koren
and Bell, 2011]. The rating function can then be inserted into the least
squares problem for the prediction error and optimized with respect to
all bias variables and weights. As the weights are learned from the data,
the user and item neighborhoods do not necessarily need to be limited
by the top k items, and sets Iu and Ui can be used instead of Qk ui and
Skui , respectively. However, if the sets are limited by a finite value of k,
the computational complexity can be reduced, at the expense of model
accuracy.
The combined model is able to learn both item–item and user–user
relationships (see section 5.7.3 for details) and, consequently, to
combine the strengths of the two approaches. It has been shown
that combined models can outperform individual user-based and
item-based models on industrial data sets [Koren and Bell, 2011].
It is important to note that the regression framework can be used
not only to combine user-based and item-based solutions but also
to integrate neighborhood-based methods with completely different
models, including some that we will discuss in the next section.
5.8 model-based collaborative filtering
From the machine learning perspective, the neighborhood-based

approach to collaborative filtering is a very narrow view of the
problem because it focuses on the k nearest neighbor estimates and
does not leverage other machine learning methods. Consequently, a
neighborhood-based recommender system inherits some fundamental
limitations of the k nearest neighbors approach. First, the performance
of neighborhood methods can decrease on sparse data where items
or users have very few common ratings, so that the recommendations
can be made based on neighbors that are not really similar to the
target user or item. Next, the k nearest neighbors algorithm relies
on pairwise instance comparison and defers the computation of the
recommendations until it is requested, which makes it challenging to
split the computation into offline and online phases.
An alternative approach is to build a rating prediction model by
using more advanced methods of supervised and unsupervised ma-
chine learning. Collaborative filtering is essentially a matrix comple-
tion task, so many standard classification and regression methods can
be adopted to it. This approach, known as model-based collaborative
filtering, generally offers a few advantages over the neighborhood-
based methods:
5.8 model-based collaborative filtering 325
accuracy. Some machine learning methods, such as the Naive Bayes

classifier, are based on a solid theoretical framework that en-
ables more accurate rating predictions than the heuristic simi-
larity measures used by neighborhood-based recommender sys-
tems.
stability. Dimensionality reduction methods can transform a sparse

rating matrix into a more condensed representation, which im-
proves the stability of rating predictions on the incomplete data.
scalability. Machine learning methods often consist of model train-

ing and model evaluation phases that help to separate offline
computations from online recommendation requests, thereby im-
proving system scalability.
Some model-based methods can deliver all of these improvements,
whereas others achieve only some of these goals. In the rest of this sec-
tion, we will consider several important methods that can outperform
neighborhood-based systems or can be combined with neighborhood-
based algorithms to create hybrid solutions.
5.8.1 Adapting Regression Models to Rating Prediction
In a general case, classification and regression models can be adapted

to the rating prediction problem by treating known ratings as features
and missing ratings as response variables. Let us first consider a hy-
pothetical case where only one rating in the matrix is missing and all
other ratings are known. Similarly to the neighborhood approach, we
have two symmetrical alternatives, depending on how the columns and
rows of the rating matrix are interpreted. The first option is to treat the
columns of the rating matrix as features and the rows as data samples.
A classification model is trained for each item i by separately consider-
ing the i-th column as the response and other columns as features, so
that the rating for a given item is predicted based on the other item rat-
ings, as illustrated in Figure 5.11. This approach is structurally similar
to the item-based neighborhood methods. The second alternative is to
treat the rows of the rating matrix as features and the columns as data
samples, so that a classification model is created for each user and the
ratings for the target user are predicted based on the ratings from their
peers. This can be viewed as a user-oriented approach to the problem.
In practice, however, the rating matrix is often very sparse, so one
cannot assume that all ratings in the training instances are known. This
is a serious issue that can substantially impact the quality of rating
predictions, depending on how the missing values are handled. There
are several possible ways to address this challenge:
326 recommendations
items
x1 ... xi - 1 y xi + 1 ... xm
u rui = fi(xu)
rui
users training
samples fi(x)
R i
Figure 5.11: Adapting a regression or classification model to rating prediction in

an item-oriented way. Known ratings of the target user u are inter-
preted as features x1 , . . . , xm , and the rating to be predicted is in-
terpreted as response variable y. Regression or classification model
pq
fi x is trained for a given item i with other users as training in-
stances, and it is evaluated on feature vector xu , which corresponds
to the target user.
• Some classification methods can be directly adapted to handle

the missing values. The Naive Bayes classifier described in the
next section is an example of such a solution.
• In certain cases, missing ratings can be filled with zeros. This pri-
marily applies to unary rating matrices where each element indi-
cates whether a user interacted with a given item or not [Aggar-
wal, 2016]. This approach, however, cannot be universally used
for every rating type because the insertion of default (zero) rat-
ing values results in prediction bias.
• One of the most generic approaches to the problem is the iterative

one [Xia et al., 2006; Su et al., 2008]. The missing values can be
initialized with some basic estimates, such as row or column aver-
ages. This gives a complete rating matrix that can be used to train
classification models. The originally missing ratings can then be
estimated by using the obtained classifiers, and the correspond-
ing elements in the rating matrix can be updated with these new
values. This gives a second complete rating matrix that can used
to retrain the models, so the process can be repeated iteratively
until convergence.
Finally, it is worth noting that the techniques described above can
be mixed with each other, as well as with neighborhood-based and
content-based recommendation methods. For example, one can use the
Naive Bayes collaborative filtering algorithm described in the next sec-

tion to initialize the missing ratings and then compute Pearson simi-
larities between users or items based on this complete rating matrix to
make the actual recommendations [Su et al., 2008].
5.8.2 Naive Bayes Collaborative Filtering
The Naive Bayes collaborative filtering algorithm attempts to predict

a missing rating by estimating the probabilities of all possible rating
values (e. g., 1, 2, 3, 4 and 5 stars) and selecting the most probable op-
tion [Miyahara and Pazzani, 2000; Su and Khoshgoftaar, 2006]. As we
mentioned in the previous section, the Naive Bayes classifier can be
constructed in either user-centric or item-centric ways. We choose to
focus on the item-centric approach because it is arguably more impor-
tant for practical usage, for the reasons we discussed in section 5.7.3.
The user-centric solution can be constructed in almost exactly the same
way by interchanging users and items, that is, the rows and columns
of the rating matrix.
According to the item-centric approach, we build a Naive Bayes clas-
sifier for a given item i in order to predict rating rui . The rating esti-
mate is then obtained by evaluating the model with the set of known
ratings for user u, which we denote as Iu , as the input. If the ratings are
categorical variables that take values from K possible classes c1 , . . . , cK ,
the prediction problem is to find the most probable rating class given
the observed ratings:
rui argmax Pr prui ck | Iu q (5.76)

ck
To evaluate the probability of a certain rating class, given the ob-

served ratings, we first apply the Bayes rule to decompose this proba-
bility:
Pr prui ck | Iu q Pr pck q PrPrpIpuI |qrui ck q (5.77)

u
in which Pr pck q is the prior probability of rating class ck and

Pr pIu | rui ck q is the likelihood of observing the known rating of
user u given that this user rated item i as ck . The probability of the
observed ratings Pr pIu q in the denominator can be ignored because
it is constant for all classes and, consequently, does not influence the
choice of the most probable class. The next step is to apply the Naive
Bayes assumption to estimate the likelihood of the observed ratings.
According to this assumption, all observed ratings are considered
328 recommendations
conditionally independent, so the likelihood can be broken down into

the product of individual rating probabilities:
¹
Pr pIu | rui ck q Pr ruj | rui ck (5.78)
j P Iu
Gathering all of these intermediate results together, we obtain the
final expression for the rating prediction:
¹
rui argmax Pr pck q Pr ruj | rui ck (5.79)
ck j P Iu
The last task is to estimate the probabilities in equation 5.79 from the
data. In the context of item i, the prior probability of rating class ck is
estimated as the fraction of the ratings for item i that are equal to ck :
°
Pr pck q
v P Ui Iprvi ck q
| Ui | (5.80)
in which Ui is the set of users who rated item i, and Ipxq is the
indicator function that equals one if the argument is true and zero
otherwise. The conditional probability that user u rates item j as ruj ,
given that this user had previously rated item i as ck , can be estimated
as follows:
°
Pr ruj | rui ck
v P Ui°Iprvj ruj and rvi ck q
v P Ui Iprvi ck q
(5.81)
The numerator in expression 5.81 is equal to the number of users

who have rated item j similarly to user u and, at the same time, rated
item i as ck . The denominator is simply the number of users who rated
item i as ck . Consider Figure 5.12 as an illustration. Assuming that we
are assessing the probability of the hypothesis that rui is 3 stars, that
we have three users who rated item i as 3 stars, and that there is one
user among them who rated item j in the same way as user u (a rating
of 5 stars), the likelihood of observing the known ratings for item j,
given that the hypothesis is correct, is

Pr ruj | rui 3 32 (5.82)
In practice, formula 5.81 is often adjusted by using the Laplace esti-

mator technique to avoid zero counters and to smooth the estimates. If
| C | is the total number of rating classes, the smoothed version of the
likelihood estimator is as follows:
°
Pr ruj | rui ck
v P U°i Iprvj ruj and rvi ck q 1
v P Ui Iprvi ck q |C| (5.83)
i j
R
3 5
ck = 3 5 u
2 3
3 4
Figure 5.12: Example of likelihood estimation in the Naive Bayes recommender.
example 5.5
Let us now consider a complete numerical example by using the stan-

dard movie rating matrix from table 5.3. We have five rating classes,

from 1 to 5 stars, and the input rating matrix is as follows:

5 4 — 1 2 1
4 — 3 1 1 2

R
— 5 5 — 3 3
(5.84)
2 — 1 4 5 4
2 2 2 — 4 —
1 2 1 — 5 4
Let us take the missing rating r13 as an example and go through
the calculation needed to predict it with the item-oriented Naive Bayes
algorithm. The first step is to estimate the prior class probabilities ac-
cording to expression 5.80. This gives the following vector of probabil-
ities for the classes from 1 to 5, based on their frequencies in the third
column of the rating matrix:
c1 c2 c3 c4 c5
(5.85)
p q
Pr ck 2/5 1/5 1/5 0 1/5
Next, we estimate the conditional probabilities according to expres-
sion 5.83 with respect to the target item i 3 for all classes ck and all
items j 3. This results in the following matrix of probabilities:
c1 c2 c3 c4 c5

j 1 1/7 1/6 1/6 1/5 1/6
j2 1/7 1/6 1/6 1/5 1/6

j3
— — — — —
(5.86)
j4
1/7 1/6 1/3 1/5 1/6

j5 1/7 1/6 1/6 1/5 1/6
j6 1/7 1/6 1/6 1/5 1/6
330 recommendations
The above matrix can be calculated on demand for a given target

item, or we can precompute all m m | C | values for all possible
combinations of target items i, peer items j, and classes ck . By multi-
plying the values in matrices 5.85 and 5.86 column-wise in accordance
with expression 5.79, we obtain the following class probabilities:
c1 c2 c3 c4 c5

p | q
Pr r13 ck I1 2/84035 1/38880 1/19440 0 1/38880
(5.87)
This means that a rating of 3 is the best estimate for r13 . If we re-
peat the process for all missing ratings, we get the results presented in
table 5.8.
Forrest The The

User 1 5 4 [3] 1 2 1
User 2 4 [4] 3 1 1 2
User 3 [2] 5 5 [1] 3 3
User 4 2 [2] 1 4 5 4
User 5 2 2 2 [4] 4 [4]
User 6 1 2 1 [4] 5 4
Table 5.8: Example of ratings predicted by using the item-based Naive Bayes
collaborative filtering algorithm.
As a side note, let us also show how the Naive Bayes approach can be
connected to neighborhood-based collaborative filtering. We can make
their structural similarity more apparent by replacing the product in
equation 5.78 by a sum of logarithms and inserting it into class proba-
bility formula 5.77:
Pr prui ck | Iu q Pr pck q ¸
sk pi, jq
Pr pI q (5.88)
u j P Iu
in which

sk pi, jq log Pr ruj | rui ck (5.89)
Note that sk pi, jq is estimated by a pairwise comparison of ratings

for items i and j, so this value can be interpreted as a kind of simi-
larity measure between the two items. This result can be compared to
the item-based nearest neighbor formula 5.56 from which ratings are
predicted with a heuristic item similarity metric. This more accurate

foundation gives some advantage to the Naive Bayes approach relative
to the basic neighborhood methods, and it can substantially outper-
form them on some data sets [Miyahara and Pazzani, 2000].
5.8.3 Latent Factor Models
The collaborative filtering algorithms that we have discussed so far

do most of their calculations by using individual elements of the rat-
ing matrix as inputs. Neighborhood methods estimate the missing rat-
ings directly from the known values in the rating matrix. Model-based
methods add an abstraction layer on top of the rating matrix by creat-
ing a predictive model that captures certain patterns of user–item inter-
actions, but model training still heavily depends on the rating matrix
properties. Consequently, these collaborative filtering methods gener-
ally face the following challenges:
• The rating matrix may contain many millions of users, millions
of items, and billions of known ratings, thereby creating major
computational and scalability challenges.
• The rating matrix is typically very sparse (in practice, about 99%
of ratings can be missing). This impacts the computational sta-
bility of the recommendation algorithms and leads to unreliable
estimates in cases when a user or item has no really similar neigh-
bors. This problem is often aggravated by the fact that most of the
basic algorithms are either user-oriented or item-oriented, which
limits their ability to capture all types of similarities and interac-
tions available in the rating matrix.
• The data in the rating matrix are usually highly correlated be-
cause of similarities between users and items. This means that
the signals available in the rating matrix are not only sparse but
also redundant, which contributes to the scalability problem.
The above considerations indicate that a raw rating matrix can be
a non-optimal representation of the rating signals and we should con-
sider some alternative representations that are more suitable for collab-
orative filtering purposes. To explore this idea, let us go back to square
one and reflect a little bit on the nature of recommender services. Fun-
damentally, a recommender service can be viewed as an algorithm that
predicts ratings based on some measure of affinity between a user and
item:
p
rui affinity pu, iq (5.90)
332 recommendations
One possible way to define this affinity measure is to take the latent
factors approach and map both users and items to points in some k-
dimensional space, so that each user and each item is represented as a
k-dimensional vector:
u ÞÑ pu ppu1 , . . . , puk q (5.91)

i ÞÑ qi pqi1 , . . . , qik q
The vectors should be constructed in such a way that the correspond-
ing dimensions of p and q are comparable to each other in a consis-
tent way. In other words, each dimension can be viewed as a feature
or concept, so that puj is the measure of proximity between user u
and concept j, and qij is, symmetrically, the measure of proximity be-
tween item i and concept j. In practice, these dimensions are often
interpretable as genres, styles, or other attributes that are applicable
to both users and items. The affinity between the user and item and,
consequently, the rating can then be defined as a product of the corre-
sponding vectors:
ķ
p
rui pu qTi pus qis (5.92)

s 1
As each rating is decomposed into a product of two vectors that be-

long to the concept space that is not directly observed in the original
rating matrix, p and q are called latent factors. The success of this ab-
stract approach, of course, totally depends on exactly how the latent
factors are defined and constructed. To answer this question, let us first
note that expression 5.92 can be rewritten in matrix form as follows2
p P QT
R (5.93)
in which P is an n k matrix assembled from vectors p, and Q is

an m k matrix assembled from vectors q, as illustrated in Figure 5.13.
The main objective of a collaborative filtering system is usually set
as minimization of the rating prediction error, and this allows us to
straightforwardly define the optimization problem with respect to the
latent factor matrices:
min R R R P QT
2 2
p

(5.94)

P, Q
With the assumption that the number of latent dimensions k is fixed

and k ¤ n and k ¤ m, optimization problem 5.94 is an instance of the
2 In mathematics literature, it is very common to denote such factors as U and V. We use P
and Q notation here to avoid clashes with the user index u commonly used throughout
the book.
m k
pu m
T
n R = n P Q k
+
qi
Figure 5.13: The latent factor approach to collaborative filtering.
low-rank approximation problem that we discussed in Chapter 2. To

demonstrate an approach to the solution, let us ignore for a moment
the sparsity of the rating matrix and assume it to be complete. In this
case, the optimization problem has an analytic solution in terms of
the singular value decomposition (SVD) of the rating matrix. More
specifically, the rating matrix can be broken down into a product of
three matrices by using the standard SVD algorithm:
R UΣVT (5.95)
in which U is an n n column-orthonormal matrix, Σ is an n m

diagonal matrix, and V is an m m column-orthonormal matrix. The
optimal solution of problem 5.94 can then be obtained in terms of these
factors truncated to the k most significant dimensions:
p U Σ VT
R (5.96)
k k k
Consequently, the latent factors that are optimal from the prediction
accuracy standpoint can be obtained by means of the SVD as follows:
P Uk Σ k
Q Vk
(5.97)
This SVD-based latent factor model helps to solve the collaborative

filtering challenges outlined at the beginning of this section. First, it
replaces a large n m rating matrix with n k and m k factor ma-
trices, which are typically much smaller because the optimal number
of latent dimensions k is typically quite low in practice. For example,
it has been reported that a rating matrix with 500,000 users and 17,000
items can be approximated reasonably well by using just 40 dimen-
sions [Funk, 2016]. Next, the SVD decorrelates the rating matrix: the la-
tent factor matrices defined by expression 5.97 are column-orthogonal,
which means that the latent dimensions are decorrelated. If k ! n, m,
which is typically true in practice, the SVD also addresses the sparsity
334 recommendations
problem because the signal present in the original rating matrix is ef-
ficiently condensed (recall that we select the top k dimensions with
the highest signal energy) and the latent factor matrices are not sparse.
Figure 5.14 illustrates this property. The user-based neighborhood al-
gorithm (5.14a) convolves the sparse rating vector for a given item with
the sparse similarity vector for a given user to produce the rating es-
timate. In contrast, the latent factor model (5.14b) estimates the rating
by convolving the two vectors of reduced dimensionality and higher
energy density.
i sim(u,v)
m k k
R P Q
n n qi
pu
rui rui
(a) Neighborhood model (b) Latent factor model
Figure 5.14: Signal energy distribution in the user-based neighborhood and la-
tent factor models
Although the approach we just described looks like a neat solution

for the latent factor problem, it actually has a major flaw because of
the assumption that the rating matrix is complete. If the rating matrix
is sparse, which is almost always the case, the standard SVD algorithm
cannot be straightforwardly applied because it cannot handle the miss-
ing (undefined) elements. The simplest solution to this problem is to fill
the missing ratings with some default value, but this can cause a major
prediction bias. It is also inefficient from the computational standpoint
because the computational complexity of this solution is equal to the
complexity of the SVD for a full n m matrix, although it is desirable
to have a method with a complexity proportional to the number of

known ratings. These issues can be addressed by the alternative factor-
ization methods described in the following sections.
5.8.3.1 Unconstrained Factorization

The standard SVD algorithm is an analytic solution of the low-rank
approximation problem. This problem, however, can be viewed as an
optimization task, and generic optimization methods can be applied to
it as well. One of the most basic approaches is to use the gradient de-
scent method to iteratively refine the latent factor values. The starting
point is to define the cost function J as the residual prediction error:
J R PQT
2
min (5.98)

P, Q
Note that we do not impose any constraints, such as orthogonality,

on the latent factor matrices at this point. Calculating the gradient of
the cost function with respect to the latent factors, we get the following
result:
B J 2 R PQT Q 2EQ
BP
(5.99)
B J 2PT R PQT 2PT E
B QT
in which E is the residual error matrix:
E R PQT (5.100)
The gradient descent algorithm minimizes the cost function by mov-

ing in the negative gradient direction at each step. Consequently, we
can find the latent factors that minimize the squared rating prediction
error by updating matrices P and Q iteratively in accordance with the
following expressions until convergence:
PÐP α EQ
(5.101)
Q T
ÐQT
α PT E
in which α is the learning rate. The disadvantage of the gradient de-

scent approach is that each iteration requires the computation of the
entire residual error matrix and the updating of all latent factor val-
ues at once. An alternative approach that is arguably more suitable for
large matrices is stochastic gradient descent [Funk, 2016]. The stochas-
tic gradient descent algorithm exploits the fact that the total prediction
error J is the sum of the prediction errors for individual elements of
336 recommendations
the rating matrix, so it is possible to approximate the overall gradient

of J by a gradient at a single data point and update the latent factors
element by element. A complete implementation of this idea is shown
in algorithm 5.1.
input training set R (sampled from the original rating matrix)
output matrices P and Q
initialize pud
p0q random with mean µ 1 ¤ u ¤ n, 1¤d¤k
0
initialize qid
p0q random with mean µ 1 ¤ i ¤ m, 1¤d¤k
0
for concept dimension d 1, 2, . . . , k do

repeat
for each rating rui in the training set do
ḑ
p
rui pus qis

s 1
e rui prui
pud Ð pud α e qid
qid Ð qid α e pud
SSEpt 1 q Ð SSEpt 1q e2 (sum of squared errors)

end

SSEpt 1q SSEptq ε

until (convergence condition)
end
Algorithm 5.1: Unconstrained matrix factorization with the stochastic gradient

descent algorithm. t is the iteration counter for the middle loop,
µ0 µ{
ak|µ|.
ε is the convergence threshold, µ is the mean known rating, and
The first stage of the algorithm is to initialize the latent factor ma-
trices. Selection of these initial values is not really important, but we
choose to evenly distribute the energy of known ratings among the
randomly generated latent factors. The algorithm then optimizes the
concept dimensions one by one. For each dimension, it repeatedly goes
through all ratings in the training set, predicts each rating by using the
current latent factor values, estimates the prediction error, and adjusts
the factor values in accordance with expressions 5.101. A given dimen-
sion is considered done once the convergence condition is met and the
algorithm switches to the next dimension.
Algorithm 5.1 helps to overcome the limitations of the standard
SVD. It optimizes the latent factors by cycling through the individ-
ual data points and, consequently, avoids the issues with the missing
ratings and algebraic operations over ginormous matrices. The itera-
tive element-by-element approach also makes the stochastic gradient
descent more convenient for practical applications than the gradient
descent, which updates entire matrices by using expressions 5.101.
example 5.6
The latent factor approach is essentially a group of representation

learning methods that is able to reveal the patterns that are implicitly

present in the rating matrix and render them explicitly as concepts.
It is sometimes possible to interpret certain concepts, especially
high-energy ones, in a meaningful and insightful way, although this
does not mean that all concepts always have a clear semantic meaning.
For example, the matrix factorization algorithm applied to the movie
ratings database can produce factors that roughly correspond to
psychographic dimensions, such as romance, comedy, horror, and so
on. We illustrate this phenomenon with a small numerical example
that uses the rating matrix from table 5.3 as an input:

5 4 — 1 2 1
4 — 3 1 1 2

R
— 5 5 — 3 3
(5.102)
2 — 1 4 5 4
2 2 2 — 4 —
1 2 1 — 5 4
We first subtract the global mean of µ 2.82 from all elements to
center the matrix and then execute algorithm 5.1 with k 3 latent
dimensions and the learning rate α 0.01, to obtain the following two
factors:

1.40 0.30 0.95
1.16 0.31 0.60

1.03 0.90 0.34 0.96 0.82 0.43
0.94 1.26 0.83 (5.103)

P Q
1.53 0.12 0.71
1.26
0.80
0.34

0.66 1.29 0.30 0.37
0.16 0.07 1.18 0.90 0.60
1.47 0.38 0.05 0.83 0.37 0.44
338 recommendations
Each row in these matrices corresponds to either a user or a movie,

and all 12 row vectors are visualized in Figure 5.15. Note that the el-
ements in the first column (first concept vector) have the largest mag-
nitudes, and the magnitudes in the subsequent columns gradually de-
crease. This is because the first concept vector captures as much signal
energy as it is possible to capture with a single dimension, the second
concept vector captures only a part of the residual energy, and so on.
Next, note that the first concept can be semantically interpreted as the
drama–action axis, where the positive direction corresponds to the ac-
tion genre and the negative direction corresponds to the drama genre.
The rating data in this example are highly correlated, so one can clearly
see that the first three users and first three movies have large negative
values in the first concept vector (drama movies and users who like
such movies), whereas the last three users and movies have large pos-
itive values in the same column (action movies and users who prefer
this genre). The second dimension in this particular case corresponds
mainly to the user or item bias, which can also be interpreted as a
psychographic attribute (Is the user generous or critical? Is the movie
popular or not?). The remaining concepts can be considered as noise.
The Godfather
drama
user 1 -1
Forrest
Gump
user 3 0 d₁ action
0.5 Titanic 1
d₃ -0.5 user 2 The
Matrix
user 5 Alien
1
user 4
0 d₂ user 6
Batman
Figure 5.15: Visualization of latent factors 5.103. Dimensions d1 , d2 , and d3

correspond to the first, second, and third columns of the matrices,
respectively. Users and movies are shown as filled and empty cir-
cles, respectively.
The obtained factor matrices are not perfectly column-orthogonal,

but they are leaning toward orthogonality because it follows from the
optimality of the SVD solution. We can see this by examining products
PT P and QT Q, which are close to diagonal matrices:

8.28 0.19 0.62
PT P 0.19 3.54 0.05
0.62 0.05 1.47
(5.104)
7.60 0.28 0.63
QT Q 0.28 2.31 0.49
0.63 0.49 1.92
Matrices 5.103 are essentially a predictive model that can be used
to estimate both known and missing ratings. The estimates can be pro-
duced by multiplying the two factors and adding back the global mean:
p PQT
R µ

5.11 4.00 [ 4.01 ] 0.75 2.00 1.35
3.94 [ 2.93 ] 3.19 1.11 1.00 1.49

2.67 (5.105)
[ 1.86
4.31 ] 5.03
[ 1.61 ]
5.19
0.94
[ 2.12 ]
4.30
3.03
5.01 3.71

1.99 2.15 1.88 [ 3.87 ] 3.95 [ 3.51 ]
1.26 1.69 1.20 [ 4.81 ] 4.94 4.17
The results reproduce the known ratings quite accurately and predict
the missing ratings in a way that corresponds to our intuitive expecta-
tions. The accuracy of the estimates can be increased or decreased by
changing the number of dimensions, and the optimal number of di-
mensions can be determined in practice by cross-validation and the
selection of a reasonable trade-off between computational complexity
and accuracy.
N
5.8.3.2 Constrained Factorization
The standard SVD algorithm gives an optimal solution of the low-rank
approximation problem, and the factors P and Q produced by the SVD
are column-orthogonal. The stochastic gradient descent algorithm 5.1
approximates this optimal solution. If the input rating matrix is com-
plete, algorithm 5.1 converges to the same column-orthogonal outputs
as the SVD. The only difference is that the diagonal scaling matrix
present in the SVD is rolled into the two factors. However, if the input
rating matrix is not complete, the outputs produced by the algorithm
are not necessarily orthogonal. This means that the concepts remain
correlated in the statistical and geometric senses. Although this does
340 recommendations
not necessarily impact the quality of the rating prediction, it makes

the results less interpretable because of the residual correlations. We
can pose the question of how additional constraints, such as orthogo-
nality or non-negativity can be imposed on the latent factors to better
define the concept space. Fortunately, the stochastic gradient descent
algorithm can be modified to support such additional constraints and
provide substantial flexibility and control over the factorization pro-
cess.
Let us consider the problem of factorization with orthogonality con-
straints. Similarly to the unconstrained optimization, the objective is
to find the latent factors that minimize the squared prediction error,
but the additional constraint is that the basis of the concepts has to be
orthogonal. This translates into the following constrained optimization
problem:
R P QT
2
min

P, Q
(5.106)
subject to PT P is a diagonal matrix
QT Q is a diagonal matrix
The gradient descent method can be adapted to constrained opti-

mization problems by using a technique called projected gradient de-
scent. The idea of this technique is to apply the constraints at each
iteration of gradient descent, so that the updated variable is projected
back to the set of feasible solutions. More formally, gradient descent
minimizes the cost function by iteratively moving in the negative gra-
dient direction:
min f pxq
x
(5.107)
learning rule: xpt 1q xptq α ∇f xptq
The projected gradient descent generalizes this approach to con-

strained optimization. At each step, we first move in the negative gra-
dient direction and then adjust the solution to stay within the feasible
set:
min f pxq
x
subject to xPC
(5.108)
learning rule: z pt 1 q xptq α ∇f xptq
xpt q argmin pt q x

1 1
z
xPC
In the case of orthogonality constraints, the feasible set for concept

d is all vectors that are orthogonal to the previously computed concept
vectors. This means that we can map the solution updated by gradient
descent to the feasible set by subtracting its projections onto the previ-
ously computed concepts. For instance, with the assumption that the
first user concept vector p1 (the first column of matrix P) is determined
during the first iteration of the outer loop of algorithm 5.1, a candidate
solution for the second concept vector p2 (the second column of matrix
P) can be orthogonalized as
p2 p2 proj pp2 , p1 q (5.109)
in which proj pa, bq is the vector projection of a onto b defined as

ab
proj pa, bq
bb
b (5.110)
At the next step, the third user concept vector can be orthogonalized
by subtracting its projections onto the previous two
p3 p3 proj pp3 , p1 q proj pp3 , p2 q (5.111)
and so on. This process is essentially an iterative version of the

Gram–Schmidt process, a basic procedure in linear algebra that takes
an arbitrary set of linearly independent vectors and constructs a set of
orthogonal vectors from it. The same approach can be used for the item
concept vectors, that is, for the columns of matrix Q. Inserting these
orthogonalization operations into algorithm 5.1, we get algorithm 5.2,
which uses exactly the same inner loop to update the elements of the
latent factors but has an additional projection step to orthogonalize the
basis of the concept vectors.
Finally, we conclude this section with a remark about the different

types of constraints that can be imposed on the latent factors, other
than orthogonality. Algorithm 5.2 produces strictly orthogonal latent
factors, even in the case of an incomplete rating matrix. This improves
the interpretability of the results to a certain extent, but the relation-
ships between users, items, and concepts can still be difficult to inter-
pret because of the interplay of positive and negative factor values. We
can try to address this issue by changing the constraints from orthogo-
nality to non-negativity:
min R P QT
2

P, Q
P¥0
(5.112)
subject to
Q¥0
342 recommendations
input training set R (sampled from the original rating matrix)
output matrices P and Q
initialize pud
p0q random with mean µ 1 ¤ u ¤ n, 1¤d¤k
0
initialize qid
p0q random with mean µ 1 ¤ i ¤ m, 1¤d¤k
0
for concept dimension d 1, 2, . . . , k do

repeat
for each rating rui in the training set do
update elements of pd and qd (see algorithm 5.1)
end

d¸1
pd Ð pd proj ppd , ps q (projection)

s 1
1
d¸
qd Ð qd proj pqd , qs q (projection)

s 1
until convergence condition

end
Algorithm 5.2: Matrix factorization with orthogonality constraints by using the

stochastic gradient descent algorithm.
This optimization problem, known as non-negative matrix factoriza-

tion, can also be solved by using a variant of the gradient descent al-
gorithm [Zhang et al., 1996; Lee and Seung, 2001]. The advantage of
non-negative factorization is better interpretability of the results be-
cause each factor element indicates proximity to a concept and each
user or item can be represented as an additive linear combination of
concepts.
5.8.3.3 Advanced Latent Factor Models

The factorization methods we discussed in the previous sections pro-
vide a solid framework for the creation of latent factor models. These al-
gorithms, however, are very basic and leave a lot of room for additional
improvements and extensions. Such extended models, sometimes very
sophisticated, can deliver a substantial improvement in the quality of
the top k recommendations, although the incremental improvement in
terms of prediction accuracy (i. e. , the mean squared error) can be very
limited [Koren, 2008]. In this section, we review several advanced mod-

els that can be viewed as practical implementations of the latent factors
approach. These methods are mainly based on ideas that we have al-
ready discussed in connection with other recommendation algorithms,
but they have been adapted to the latent factors approach.
regularization and biases As we discussed earlier in sec-

tion 5.6.1, the user and item biases are important baseline estimators
that can help to capture and remove the average user and item effects.
Because both baseline estimates and latent factors can be determined
by using gradient descent, we can combine the two models into one
and jointly optimize both bias and latent factor variables. The rating
prediction formula for this model will be defined as follows:
p
rui µ bi bu pu qTi (5.113)
in which µ is the global mean, bi is the item bias, bu is the user bias,
and the last term corresponds to the latent factors part of the model.
Adding a regularization term that helps to avoid overfitting on sparse
data, we translate this model into the following optimization problem:
¸ 2
min rui µ bi bu pu qTi
u,i (5.114)

λ b2i b2u kpu k2 kqi k2
in which λ is a regularization parameter and minimization is done

by all bias and latent factor variables simultaneously. This problem
can be solved by using a version of basic stochastic gradient descent
algorithm 5.1 with the following learning rules:
bu Ð b u α pe λ b u q
bi Ð bi α pe λ bi q
pud Ð pud α pe qid λ pud q
(5.115)
qid Ð pid α pe pud λ qid q
in which α is the learning rate. This model can be considered as

a practical version of the basic unconstrained factorization discussed
earlier and is often referred to as an SVD model, which is technically
not accurate.
implicit feedback The second model that we will consider lever-

ages the observation that the user selects items to rate not at ran-
dom but in accordance with personal interests and preferences. Conse-
quently, a useful signal is carried not only by the actual rating values
344 recommendations
but by the positions of the known ratings as well (see section 5.1.1 for
details). We can isolate this signal about user–item interactions in the
n m implicit feedback matrix, which contains ones in the positions of
known ratings and zeros in the positions of missing ratings. Normaliz-
ing each row to a unit length, we define the implicit feedback matrix F
as
$
fui |
& I
u | {
1 2
, if rui is known
(5.116)
%
0, otherwise
in which Iu is the set of items rated by user u. In a general case,

the implicit feedback matrix is not necessarily derived from the rating
matrix and can be created from a different data source. For example,
the implicit feedback matrix can be created based on purchase or web
browsing histories, so that each non-zero element indicates an interac-
tion between a user and item.
The idea of the factorized model with implicit feedback is to intro-
duce an additional set of item factors where each factor value yid
characterizes how much the act of rating item i increases or decreases the
proximity to concept d. Let us denote this set of factors as the m k
matrix Y. The product of the implicit feedback matrix and this new
item–factors matrix FY pzud q is an n k matrix in which rows cor-
respond to users, columns corresponds to concepts, and each element
zud can be interpreted as the incremental proximity of user u to con-
cept d attributed to the implicit feedback, that is, the act of rating. This
incremental proximity can be added directly to the main user–factor
matrix P, which characterizes the user–concept proximity derived from
the rating values; the result is the optimization problem
R pP FYq QT
2
min (5.117)

P, Q, Y
By adding user and item biases, we get the following rating predic-
tion formula:

¸
p
rui µ bi bu pu | Iu |1{2 yj qTi (5.118)
j P Iu
in which yj are the rows of matrix Y. The learning rules for
stochastic gradient descent can be straightforwardly derived from
expression 5.118 by including regularization terms and taking the

gradients:
bu Ð b u α pe λ 1 b u q
b i Ð b i α pe λ 1 b i q
pud Ð pud α pe qid λ2 pud q
(5.119)
qid Ð pid α e pud | Iu | { j P I yjd λ2 qid
° 1 2
u

yjd Ð yjd α e | Iu | { qid λ2 yjd
1 2
in which λ1 and λ2 are the regularization parameters. This model,

known as the SVD++ model, can offer better accuracy than basic SVD
because of more accurate handling of the implicit feedback signal [Ko-
ren, 2008]. The SVD++ model is often considered to be one of the most
advanced and efficient latent factor models.
fusing latent factors with neighborhoods Finally, we

consider a model that combines factorization with the nearest neigh-
bor approach. As we discussed in section 5.7.4, neighborhood-based
collaborative filtering can generally be considered as a regression
problem. This problem can be solved either analytically or by using
optimization methods such as stochastic gradient descent. The latter
approach enables us to roll the neighborhood model into the factor-
ization algorithm and optimize the latent factors together with the
neighborhood model weights. An integrated model can be obtained
by combining the latent factor expression 5.118 with the neighborhood
expression 5.69 that we discussed previously, which results in the
following formula for the rating prediction:

¸
p
rui µ bu bi pu | Iu | { 1 2
yj qTi
j P Iu (5.120)
¸
|1{2

| Qsui ruj buj wij cij
j P Qsui
in which Qsui is the neighborhood of item i in the set of items rated
by user u (in other words, the top s most similar items in Iu ) and
1{2
buj is the baseline prediction. Factors | Iu | {2 and Qsui
1
can be
interpreted as the reliabilities of the corresponding terms, that is, the
number of ratings that the estimate is based on, so contribution of the
terms is scaled up or down accordingly. The rating prediction error
is then minimized with respect to all underlined variables simultane-
ously by using a set of learning rules similar to set 5.119, but with
additional rules for weights wij and cij [Koren, 2008].
346 recommendations
5.9 hybrid methods
Making recommendations is a broad and challenging problem, so an

ideal recommender system should leverage multiple data sources and
account for a wide range of effects and signals, such as user–item in-
teractions, item content similarities, and many others. Most recommen-
dation methods, however, can use only one type of data and capture
only a certain type of effect. Basic collaborative filtering, for example,
is focused on the rating matrix analysis and neglects item content;
meanwhile, content-based filtering does the opposite. Consequently,
each method has its own strengths and weaknesses, and methods can
potentially complement each other. The hybrid approach attempts to
create superior recommendation systems by combining several basic
algorithms together.
We have already seen a few examples of how two or more rec-
ommendation methods can be combined together. For instance,
user-based and item-based neighborhood methods were blended
together by means of regression analysis in section 5.7.4.3, and
the basic SVD model was been augmented with implicit feedback
data in section 5.8.3.3. These hybrid solutions can generally achieve
substantially better performance than could be achieved from any
of the constituent algorithms alone. Our next goal is to develop a
more systematic and comprehensive framework for hybridization
that, ideally, can create the optimal blend from any set of recom-
mendation algorithms. This framework can help us not only to build
more powerful recommendation services but also to better and more
systematically understand some of the previously described methods.
The problem of hybrid recommendation models is closely related
to ensemble learning, which focuses on methods that generate and
combine multiple classification or regression models to obtain better
predictive performance than could be obtained by individual learning
algorithms. We use ensemble theory in the following sections to
build hybrid recommenders, by starting from very basic methods and
gradually increasing the complexity.
5.9.1 Switching
One of the most basic ways to combine several recommendation algo-

rithms together is simply to switch between them depending on some
condition. For example, we can assume that a collaborative filtering
method works well enough unless an item has too few known ratings,
in which case content-based filtering is likely to do a better job [Burke,
5.9 hybrid methods 347
2002]. Consequently, we can switch between these two algorithms de-

pending on the number of users who rated an item:
$
'
'
&p
(collaborative)
rui , if | Ui | ¡ 20
p
rui ' (5.121)
'
%p (content)
rui , otherwise
in which Ui is the set of users who rated item i. This solution can
help us to work around the cold-start problem, which is an issue for
collaborative filtering, and, at the same time, to improve trivial rec-
ommendations produced by content-based filtering whenever possible.
A generic schema of such a switching recommender is shown in Fig-
ure 5.16. This approach, however, is somewhat rudimentary because
it is based on heuristic rules rather than a formal optimization frame-
work. We can definitely achieve better results by leveraging machine
learning and optimization algorithms to properly mix the outputs of
the individual models.
Figure 5.16: A switching recommender.
5.9.2 Blending
Let us assume that several recommendation models have been trained

for the same set of users and items, so each of these models can es-
timate rating rui for a given pair of user and item. Our goal is to
combine these estimates together to produce the final rating estimate,
which is, ideally, more accurate than the predictions produced by any
single one of the models. This can be done by using heuristic rules,
as we did in the switching approach described in the previous section,
but, at the same time, it is naturally a regression problem that can be
efficiently solved by using the machine learning toolkit.
The problem of blending several rating estimates together can be for-
mally defined in the following way. Let us assume that we have s train-
ing samples, that is, known rating values in the training set. This set is
used to train q recommendation models, each of which can predict the
348 recommendations
rating for a given pair of user and item. For each training sample j, let
us denote the vector of q model outputs (predicted rating values) as xj
and the true rating value as yj . The problem of blending the available
estimates together can then be defined as finding the blending function
b pxq that minimizes the prediction error:
ş 2
min b xj yj (5.122)
b

j 1
This view of the problem is illustrated in Figure 5.17. This prob-

lem, that is, the combination of the predictions of several learning al-
gorithms by using another learning algorithm, is known as stacking, so
we use the terms blending and stacking interchangeably. Stacking is es-
sentially a standard supervised learning problem that can be solved by
using a variety of classification or regression algorithms.
Figure 5.17: A blending recommender.
One of the most basic solutions of problem 5.122 is, of course, lin-
ear regression. In this case, the combiner function is a linear function
defined as
b pxq xT w (5.123)
in which w is the vector of model weights. In other words, the final

rating prediction is a linear combination of predictions produced by
individual recommendation algorithms:
¸ pkq
q
p
rui wk prui (5.124)

k 1
The optimal weights for the blending function can be straightfor-

wardly calculated by using ridge regression:
1
w XT X λI XT y (5.125)
in which y is the vector of s known ratings, λ is a regularization

parameter, and X is an s q matrix of rating predictions. Each element
xjk is the rating predicted by algorithm k for the j-th training sample.
In practice, the best results are often obtained by using non-linear

blending models, such as neural networks and gradient boosted deci-
sion trees [Jahrer et al., 2010; Koren, 2009; Töscher et al., 2009]. The
blend can include dozens of recommendation models that can be of
different types (neighborhood-based, factorized, mixed, etc.), and each
type can be represented by multiple model variants that are trained
with different numerical parameters, such as the number of latent fac-
tors. The models can be trained on the entire training set, or this set
can be divided into subsets (bins) randomly or according to some cri-
teria and then the individual models can be trained for each bin sep-
arately. Blending is a powerful method that can substantially improve
the quality of predictions, so many blending techniques have either
been adopted from ensemble theory or developed specifically for rec-
ommender systems. We review some of these extensions and refine-
ments in the next sections.
5.9.2.1 Blending with Incremental Model Training

The basic blending approach assumes that all models in the blend are
trained in advance, and the blending function is then learned sepa-
rately to minimize the overall prediction error. This approach, however,
is not necessarily optimal because a set of models, each of which has
a low prediction error, does not necessarily produce a blend with the
minimal prediction error. To a certain extent, this can be explained by
correlation between the models – a good blend not only requires mod-
els to achieve low prediction errors individually but also to achieve
a certain level of error decorrelation [Töscher et al., 2009]. In fact, an
ideal solution would be to merge all models into one big model and si-
multaneously optimize all parameters with respect to the overall blend
prediction error. The hybrid model from section 5.8.3.3 that combines
the neighborhood approach with latent factors is actually an example
of such a solution. Unfortunately, this approach becomes intractable as
the number of models in the blend and, consequently, the number of
parameters grows. From this perspective, blending can be viewed as a
divide-and-conquer approximation of the globally optimal solution.
Blending can be improved by including the global error function of
the blend into the model training process. If it is assumed that individ-
ual models in the blend are trained by using gradient descent, which is
a fair assumption for many practical applications, one possible solution
is to redefine the convergence condition of the gradient descent loop
based on the overall prediction error of the blend. This solution is im-
plemented in algorithm 5.3. We initialize the matrix of model outputs X
with a single column of ones that can be interpreted as a constant term
in the blend. The recommendation models are then trained and added
350 recommendations
to the blend one after another. Each model is trained by using gradient
descent or stochastic gradient descent in the inner loop of algorithm 5.3.
For each iteration, we update the model by using its learning rules, pre-
dict ratings for all training samples, assemble a temporary matrix X by
appending a column of newly predicted ratings, re-optimize the blend
(the algorithm uses a linear blending function for illustration, but any
other blending model can be used), and estimate the blend prediction
error. The method does not change the error functions of individual
models and learning rules, but it changes the convergence condition
so that training stops when the overall prediction error of the blend is
minimized. In practice, the blend prediction error can continue to de-
crease after the model prediction error reaches its minimum and starts
to increase.
Xp0q s 1 column matrix of ones (constant component)

for recommendation model k 1, 2, . . . , q do
repeat
update model k (one step of model training)
predict ratings xk using model k

X Xpk1q xk (try adding xk to the blend)
1
w XT X XT y (optimize the blending function)
r Xw (calculate prediction of the blend)
SSE kr yk 2
(update the overall prediction error)
until SSE convergence
Xpkq X (permanently add xk to the blend)
end
Algorithm 5.3: Incremental model training by using a linear blending function

[Töscher et al., 2009].
5.9.2.2 Blending with Residual Training

From the error decorrelation standpoint, it can be beneficial to train
some models in the blend by using residual errors of other models as
inputs. More specifically, we chain several models so that the outputs
of models at the beginning of the chain are used as inputs for the
downstream models, as illustrated in Figure 5.18.
The models in the chain are trained sequentially. The first model is
trained based on the raw samples, the ratings predicted by this model
Figure 5.18: Model training on residual errors.
are then subtracted from the raw samples, the second model is trained
by using these residual errors as inputs, and so on. The final blend
can include predictions produced by models trained on the raw data
and predictions created based on the residual errors. Among the previ-
ously discussed models, the removal of the global rating average and
baseline predictors are basic examples of residual training techniques.
5.9.2.3 Feature-weighted Blending

The accuracy of rating predictions and, eventually, the quality of rec-
ommendations can be improved by blending together outputs of sev-
eral recommendation models. In previous sections, we have discussed
how the blending function can be constructed to mix the outputs of
models in an optimal way. We can, however, expect that the accuracy
can be improved even further if the blending function uses not only
the outputs of the recommendation models but also additional signals
about model reliability or some external signals about users or items.
For example, some models can produce very accurate results if many
ratings for a user or item are already known but can become very
inaccurate and unstable if the ratings are scarce. The weight of such
models in the blend can be increased or decreased depending on the
rating statistics. In fact, we can find some traces of this approach in
the previously described models. For instance, the discounted similar-
ity technique described in section 5.7.1 mixes the reliability data with
the user similarity measures. In the same vein, the latent factor model
with implicit feedback from section 5.8.3.3 amplifies certain factors by
using an external implicit feedback signal.
The blending framework can take advantage of external signals,
sometimes referred to as meta-features, in different ways. One possible
approach is to admix these signals into the blending function as
additional inputs, that is, to append the signals to the vector of the
recommender model outputs. Although this approach is generally
feasible, it does not work well with linear blending functions and
often requires complex nonlinear blending models to be learned, for
example, when gradient boosted decision trees are used [Sill et al.,
2009]. An alternative approach is to combine several linear models
into a pipeline with a predefined structure to mix the signals from
352 recommendations
recommendation models with the signals from meta-features. This

approach allows us to take advantage of the simplicity and stability of
linear regression but to achieve much better results than a plain linear
model that uses meta-features as additional inputs. We will spend
the rest of the section describing the details of this method [Sill et al.,
2009].
The idea of feature-weighted blending is to mix the outputs of the
recommendation models by using a linear blending function but to
calculate the blending weights as functions of the meta-features. Let us
assume a set of q recommendation models produces rating predictions
x1 , . . . , xq . In addition, let us assume p meta-features g1 , . . . , gp are
associated with each rating value. We choose to mix the predictions by
using the linear blending function
¸
q
b pxq wk xk (5.126)

k 1
but weights wk are dynamically calculated based on the meta-

features:
¸
p
wk fk pgq vki gi (5.127)

i 1
in which fk are called feature functions and vki are static weights. In
other words, the feature functions amplify or suppress the signals from
the recommendation models, as shown in Figure 5.19. Note that this
design is quite similar to the signal mixing pipelines that we discussed
in Chapter 4, in the context of search services.
Figure 5.19: Feature-weighted blending.

This model translates into the following optimization problem:

ş ¸ 2
min vki gji xjk yj (5.128)
v

j 1 k,i
in which the outer sum iterates over all s training samples, xjk is the
rating predicted by algorithm k for the j-th training sample, gji stands
for the j-th sample meta-features, and yj is the true rating value for
sample j. To solve this problem, let us first introduce the s pqpq matrix
A that contains the cross-products of predictions and meta-features for
each training sample:
aj, ppk1q i xjk gji , 1¤j¤s

1¤k¤q (5.129)
1¤i¤p
In each row of matrix A, the first p elements correspond to the first
recommendation model, followed by p elements that correspond to
the second model, and so on. In other words, the first segment of p
elements is the output of the first model xj1 , modulated by each of
the feature functions, and so on. Let us also introduce vector v with
the same structure, created by sequencing weights vki as a row with
qp elements, so that the first p elements correspond to the first model,
followed by p elements for the second model, and so forth. The opti-
mal weights can then be found by solving a regression problem that
corresponds to the following system of linear equations:

AT A λI v AT y (5.130)
in which λ is a regularization parameter and I is a diagonal identity

matrix.
Feature-weighted blending provides a relatively simple extension of
the basic weighting framework that allows one to modulate model pre-
dictions with additional signals or meta-features. Common examples
of such meta-features include basic statistics (e. g., the number of times
the item has been rated, the standard deviation of the user ratings),
time-dependent statistics (e. g., the number of distinct dates on which
a user has rated items), and correlation statistics (e. g., the maximum
correlation of the item with any other item). These statistics are impor-
tant for a hybrid recommender because the reliability of the estimates
produced by the constituent models depends on the number of avail-
able ratings and other similar factors reflected in the statistical signals.
Consequently, these features enable the hybrid model to learn how to
switch between the models depending on the expected reliability of
the estimates.
354 recommendations
5.9.3 Feature Augmentation
The next class of hybrid methods that we will consider is recom-

menders with feature augmentation. The feature augmentation
technique refers to a design where several recommendation models
are chained together in such a way that predictions produced by one
recommendation model are consumed by another model as inputs.
We already used this approach in blending with residual training, but
this idea can be exploited in several other ways.
One possible approach for chaining two recommenders together is
to use the first model in the chain to generate completely new features
that are not present in the raw data, so that the downstream models
can use them as inputs. For example, the content-based Naive Bayes
recommender described in section 5.5.2 can leverage item attributes
such as related authors and related titles to recommend books. These
attributes can be created by using an item similarity measure computed
over the rating matrix, that is, by using collaborative filtering [Mooney
and Roy, 1999]. This effectively creates a hybrid recommender with
feature augmentation, where collaborative filtering is the first model in
the chain that generates new features and the content-based classifier
is the second model that consumes them.
The second option for chaining recommenders is to use the first rec-
ommender in the chain to enhance the features that are present in the
raw data. For example, a content-based recommender can be used to
fill missing elements in the rating matrix, and this enhanced matrix can
then be used by some collaborative filtering method. This approach can
be contrasted with the previous example, which uses collaborative fil-
tering to augment the inputs of the content-based Naive Bayes recom-
mender. Let us flesh out this solution, called content-boosted collaborative
filtering, in detail, with the assumption that a user-based nearest neigh-
bor recommender is used as the collaborative filtering component of
the hybrid [Melville et al., 2002]. The first step is to use a content-based
recommender to fill in the missing elements of the rating matrix and
create a new pseudo rating matrix, defined as follows:
$
&r
zui % ui(c), if user u rated item i
(5.131)
p
rui , otherwise
(c)
in which prui is the rating predicted by the content-based recom-
mender. The second step is to apply the collaborative filtering part
of the hybrid to the pseudo rating matrix to predict the rating for a
given pair of user and item. In principle, one can use any out-of-the-
box collaborative filtering algorithm to make the final prediction. The
challenge, however, is that the ratings filled in during the previous step
skew the statistics of the number of known ratings used by many col-
laborative filtering algorithms. This requires the collaborative filtering
part to be modified and several additional factors and parameters to
be introduced to fix the statistics:
• The reliability of content-based rating predictions depends on the
number of known ratings for a given user. Consequently, predic-
tions that do not have enough support should be devalued when
used in collaborative filtering. If we use a user-based neighbor-
hood model for the collaborative filtering step, we can account
for the reliability of the incoming ratings by modifying the user
similarity measure. Let us first define a normalized support vari-
able that grows proportionally to the number of ratings provided
by a user but is limited to one if the number of ratings exceeds
the threshold parameter T :
$
qu
&1,
% | Iu | ¥ T (5.132)
| Iu | {T , otherwise
The similarity function can then be redefined by adding a factor

that equals the harmonic mean of the support variables for two
users:
sim 1 pu, vq simpu, vq

2qu qv
(5.133)
qu qv
The harmonic mean is chosen because it is biased toward the

minimum of two numbers, so the similarity measure will be sig-
nificantly penalized if either user has provided too few ratings.
• The hybrid system includes content-based and collaborative fil-

tering components that are both capable of predicting the rating
for a given user and item, so these two predictions need to mixed
together. To manage the balance between the predictions, let us
introduce amplification factor wu for the content-based predic-
tion. This factor is defined as the baseline amplification weight
wmax multiplied by the support variable in order to penalize un-
reliable predictions:
wu wmax qu (5.134)
The final rating prediction formula for the collaborative filtering part
can then be defined as follows:

wu prui µu
(c) °
v sim
1 pu, vq pzvi µv q
p
rui µu ° 1 pu, vq (5.135)
wu v sim
356 recommendations
in which µu is the average user rating computed over the pseudo

rating matrix. This is essentially the basic user-based nearest neighbor
(c)
model with the content-based estimate prui admixed and the similarity
function adjusted to account for the reliability of the content-based es-
timates. Note that these reliability-related adjustments are fundamen-
tally similar to the feature-weighted blending that we discussed in the
previous section: in both cases, the hybrid recommender uses the ba-
sic rating statistics to devalue models with low reliability and amplify
signals from models with high reliability.
The content-boosted collaborative filtering model defined by expres-
sion 5.135 can generally provide better accuracy than either of the two
constituent recommendation methods alone. If the raw rating matrix
is reasonably dense, the model outperforms both the content and col-
laborative components by taking advantage of two signals. In the case
of a sparse rating matrix, the accuracy of the collaborative filtering
component drops and the overall performance of the hybrid model
converges to the accuracy of the content-based recommender [Melville
et al., 2002].
5.9.4 Presentation Options for Hybrid Recommendations
We conclude the discussion of hybrid methods with a brief remark

about how a hybrid recommender can leverage the presentation ca-
pabilities of the recommendation service. First, it is worth noting that
recommendations produced by different models should not necessarily
be blended together – a recommender system can simply display sev-
eral lists of recommended items. Ecommerce web sites, for example,
very often display several recommendation bars with different seman-
tic meanings, such as Customers who viewed this item also viewed, Inspired
by your browsing history, Top rated items, Similar items, and so on. These
pieces of content can naturally be created by using different recommen-
dation algorithms, including both personalized and non-personalized
methods. Recommender systems that use this approach are typically
referenced as mixed hybrids.
In certain cases, the recommended items need to meet additional
requirements or conditions, depending on how the recommendations
are presented and how the user interacts with the recommendation
service. For example, a user can explicitly request the recommendation
of restaurants in a certain location or more books similar to the selected
one. In such cases, the recommender system can be used as a sorting
component that postprocesses the results created by a search service
or another recommendation model. For instance, a search service can
be used to fetch the list of items according to the criteria explicitly
5.10 contextual recommendations 357
specified by the user, and then this list can be sorted by a collaborative
filtering unit. This technique, sometimes referred to as cascading, can
be viewed as an extreme case of blending, where the signal from the
first recommender or search service is steeply pitched to sort items into
relevant and irrelevent buckets and the second recommender does the
secondary sorting within the buckets.
5.10 contextual recommendations
The majority of recommendation algorithms, including all of the meth-

ods we have discussed earlier in this chapter, are based on the assump-
tion that the relevance of a given item for a given user can be pre-
dicted by using only the profiles of the item and user. This approach
completely ignores the circumstances under which recommendations
are made, including time, user location, marketing channel, and other
pieces of information about the situation and environment. This con-
textual information, however, is very important because consumer de-
cisions are almost always contingent on the context of the decision
making. Consequently, the relevance of recommendations is unique for
each recommendation transaction rather than statistically determined
by the item and user profiles. Let us consider several specific cases to
better understand the notion of context:
location Recommendations made by a shoe store for users in
Alaska may not be relevant for users in Hawaii. Customers who
use a mobile application to find relevant restaurants nearby
may receive recommendations that will not be relevant for them
when they move to another location.
time Movie recommendations that are relevant for a fifteen-year-old

user today may not be relevant five years later when that user will
be twenty. TV program recommendations relevant in the morn-
ing may not be relevant in the evening. Recommendations made
by an apparel recommender system in one season may not be
relevant in another season.
intent The relevancy of restaurant recommendations made for

a user can change depending on whom the user dines out
with: alone, with their spouse, with co-workers, or with family.
Recommendations made for users shopping for themselves can
be different from recommendations made if they are shopping
for a gift. Hotel booking recommendations for business travelers
can be different from recommendations for leisure travelers.
358 recommendations
channel Recommendations made in emails can have a different

structure and presentation to recommendations made on a
website or in a store.
conditions Recommendations made by a department store may or

may not include umbrellas, depending on the current weather
conditions or forecast.
A recommender system should take the contextual information
about location, time, intent, and channel into account to deliver
relevant real-time user experiences. We will spend the rest of this
section discussing how recommendation algorithms can be extended
or modified to incorporate these contextual signals.
5.10.1 Multidimensional Framework
Traditional recommendation models predict the relevance of a given

item for a given user on the basis of the corresponding item and user
profiles. These models can be viewed as functions that take a user and
item as arguments and produce rating predictions:
p
rui R pu, iq (5.136)
Context-aware recommender systems extend this framework with

additional arguments, each of which represents a certain dimension
of the context, such as location, time, or channel [Adomavicius and
Tuzhilin, 2008]:
p
rui R pu, i, location, time, . . .q (5.137)
In other words, the basic rating function defined over two-

dimensional space
R: User Item Ñ Rating (5.138)
is replaced by a function defined over a multidimensional space that

includes user, item, and context dimensions:
R: User Item Location Time . . . Ñ Rating (5.139)
This idea is illustrated in Figure 5.20, which shows an example of

a three-dimensional recommendation space. In this example, each rat-
ing value is a function of a user, item, and time. All known ratings
are attributed with a time label and placed in the corresponding cells
of a three-dimensional array, instead of a two-dimensional rating ma-
trix. The goal of the recommendation model is, therefore, to predict the
rating values in the empty cells of the array. Note that a multidimen-
sional array can be collapsed onto a standard two-dimensional rating
matrix by discarding the contextual information. This may require the
merging of several rating values that are projected onto one element
of the matrix. For example, if a user rated the same item several times
on different dates, only the latest value or average value can be kept
in the rating matrix. Alternatively, the rating matrix can be obtained
by selecting a certain point on the context dimension and cutting out
a two-dimensional slice from the multidimensional cube at this point.
For example, the array depicted in Figure 5.20 can be viewed as a pile
of rating matrices Rptq, one for each time interval. Finally, a rating ma-
trix can be created not for a certain point on the contextual dimension
but for a certain range. In the case of the example in Figure 5.20, this
would be for a certain time interval.
R(t)
R(u,i,t)
users
time
items
Figure 5.20: Example of a three-dimensional recommendation space.
We generally assume that context dimensions can have a hierarchical

structure. For example, each known rating can be attributed with the
date it was given, so the time dimension is discrete and contains as
many intervals as there are distinct date labels in the rating data. Dates,
however, can be aggregated into weekly, monthly, quarterly, or yearly
intervals, and the rating matrix Rptq can thus be cut for a certain week,
month, quarter, or year. We can have more than one hierarchy for one
dimension. For instance, dates can also be categorized into weekdays
and weekends, so that rating matrices for weekdays and weekends can
be obtained. In a similar vein, fine-grained location attributes, such as
latitude and longitude, can be aggregated into zip codes, cities, states,
and countries. Finally, user and item dimensions can also be associated
360 recommendations
with certain hierarchies. For example, users can be categorized into

buckets based on their age and items can be categorized into genres.
Contextual information can be obtained from different sources.
Some attributes, such as a rating time stamp or user device location,
can be automatically collected by the recommender system or market-
ing channels that the system is integrated with. Some other attributes,
especially intent-related ones, may not be directly available but can be
obtained by using special features of the user interface (for example,
a This is a gift order checkbox on an online order placement form) or
inferred by using predictive models.
5.10.2 Context-Aware Recommendation Techniques
A non-contextual recommendation service can be viewed as a process

that consumes the training data in the form User Item Rating, cre-
ates a model that maps the pair of user and item to rating, and evalu-
ates this model for a given user to produce a sorted list of recommen-
dations. This pipeline is shown in Figure 5.21.
Data Model Recommendations
U I R U I R i, j, ...
u
Figure 5.21: The main steps of the non-contextual recommendation process
[Adomavicius and Tuzhilin, 2008]. U, I, and R are the dimensions
of users, items, and ratings, respectively. Recommended items are
denoted as i, j, . . .
The multidimensional framework described in the previous section

suggests several ideas for how this pipeline can be modified to incor-
porate contextual information [Adomavicius and Tuzhilin, 2008]:
contextual prefiltering The first possible solution is to create
a two-dimensional rating matrix from the original multidimen-
sional data and then apply a standard non-contextual recom-
mendation model or algorithm with this matrix as an input, as
shown in Figure 5.22. The rating matrix is a slice of the multidi-
mensional cube selected for a given value of the context. For ex-
ample, a movie recommendation service that stores ratings with
time stamps can make recommendations on the weekdays and
weekends by using two different matrices. The matrices are cre-
ated by selecting all weekday or all weekend ratings, respectively,
from the original data cube.
Contextualized Model Recommendations

Data Data
U I C R U I R U I R i, j, ...
c u
Figure 5.22: Context-aware recommender system with contextual prefiltering
[Adomavicius and Tuzhilin, 2008].
One of the key considerations for the prefiltering approach is the

trade-off between data sparsity and contextualization accuracy.
On the one hand, contextualization of the input data can improve
the accuracy of recommendations because the recommendation
algorithm uses only the ratings that are relevant for the context.
On the other hand, contextualization decreases the amount of
data available for the recommendation algorithm, which can neg-
atively impact the quality of recommendations. For example, a
movie recommendation service that makes weekend recommen-
dations by using only the ratings submitted on the weekends al-
most certainly loses some relevant signals carried by the weekday
ratings. Contextualization with narrow selection criteria can also
result in very sparse rating matrices, which can result in unreli-
able and skewed rating predictions. The trade-off between data
sparsity and context accuracy can be controlled by using the hi-
erarchical aggregation that we discussed in the previous section.
For example, the time context can be applied at a days-of-the-
week level of granularity (seven buckets in total) or a weekday–
weekend level (two buckets in total). Note that, in principle, we
can use the controlled precision reduction technique that we dis-
cussed in the context of search services in section 4.6.2 to try
different granularity levels and select the optimal one for a given
value of the context.
contextual postfiltering An alternative approach to contextu-

alization is postfiltering of recommendations, as depicted in Fig-
ure 5.23. A recommender system with postfiltering initially dis-
cards the contextual information and collapses the rating data
cube onto a plain rating matrix, so that a standard recommenda-
tion algorithm can be applied to produce non-contextual recom-
mendations. The list of recommended items is then contextual-
ized by using context-aware postprocessing rules that are applied
after the main recommendation method. These rules are typically
based on user or item attributes and can be either heuristic or
362 recommendations
driven by a predictive model. For example, an apparel recom-

mender system can create an initial list of non-contextual recom-
mendations by using any content-based or collaborative filtering
algorithm and can then filter or re-rank the items according to
the current season or weather. For instance, it can push warm
clothes to the top in winter. In this case, warm clothes can be de-
termined by using a heuristic rule or content classification model
such as a Naive Bayes text classifier.
Contextual
Data Model Recommendations Recommendations
U I C R U I R i, j, ... i’, j’, ...
u c
Figure 5.23: Context-aware recommender system with contextual postfiltering
contextual modeling The most generic solution for the contex-

tual recommendation problem is to create a model that can pre-
dict ratings directly as a function of multiple arguments, includ-
ing item, user, and context. This approach is illustrated in Fig-
ure 5.24. The key advantage of the contextual modeling approach
is that the context-related parameters of the model can be learned
and optimized together with other parts of the model. This can
lead to better results than heuristic pre- and postfiltering solu-
tions, which, as we have discussed, can filter out relevant data
and signals that do not formally meet the context criteria.
Contextual
Data Model Recommendations
U I C R U I C R i, j, ...
u, c
Figure 5.24: Context-aware recommender system with contextual modeling

Contextual models can often be obtained by extending standard

non-contextual content-based and collaborative filtering models.
Let us illustrate this idea with a conceptual example that uses
the neighborhood model as a basis. Recall that a non-contextual

nearest neighbor model can be expressed as follows:
¸
p
rui sim ppu, iq , pv, jqq rvj (5.140)
v,j
in which indexes v and j iterate over the neighborhood of users

and items, respectively. The similarity measure between user–
item pairs pu, iq and pv, iq can be specified in many different ways,
as we discussed earlier in the section dedicated to neighborhood-
based collaborative filtering. In the case of the user-based neigh-
borhood model, for example, the measure will be reduced to the
similarity between two users, which, in turn, can be computed as
the Pearson correlation coefficient or some other metric:
$
sim ppu, iq , pv, jqq

p q
&sim u, v , ij
(5.141)
%
0, otherwise
A time-aware contextual model can extend the notion of similar-

ity and include the time dimension:
¸
p
ruit sim ppu, i, tq , pv, j, sqq rvjs (5.142)
v,j,s
in which index s iterates over the neighborhood of ratings with

time stamps close to the target context time stamp t. The multidi-
mensional similarity measure can then be defined as the distance
between two cells in the three-dimensional data cube. For exam-
ple, one can use the Euclidean distance metric:
sim ppu, i, tq , pv, j, sqq

b (5.143)
sim2 pu, vq sim2 pi, jq sim2 pt, sq
This is just a conceptual example that illustrates the approach

for incorporating contextual information into a recommendation
model, but we will develop more practical solutions in the next
section.
Finally, let us note that multiple contextual and non-contextual rec-
ommendation algorithms can be combined together into hybrid mod-
els. The hybrid approach can help to overcome the limitations of in-
dividual methods (for example, sparsity issues caused by overly strict
contextualization of the input data) by blending multiple predictions
in the optimal way.
364 recommendations
5.10.3 Time-Aware Recommendation Models
The temporal dimension is one of the most important types of context

because of substantial variability in user–item interaction patterns
over time. A good example of such variability is item popularity
changes over time that can be caused by external factors, such as
fashion changes that make certain garments more or less popular, the
appearance of an actor in a new movie that boosts the popularity of
related movies, or advances in technology that make old electronic
devices outdated. Another example is the drift of user preferences
over time that can be caused by changes in tastes, social role, or
location. For instance, a user who used to rate an average item as
4 stars can become more critical or advanced over time and then
give only 3 stars to mediocre products. At the same time, temporal
context information is also one of the easiest types to collect because
time stamps can be set internally by a recommender system without
external dependencies on marketing channels or user interfaces. These
factors make time-aware recommendations a low-hanging fruit that
can substantially improve the accuracy of rating predictions with
relatively low cost.
The multidimensional framework and pre/postprocessing contextu-
alization techniques provide a simple toolkit that can be applied to
cases with a periodic temporal context, such as seasonality. This toolkit
is, however, very basic, and we will explore how to take advantage of
predictive and optimization methods to create more advanced and ac-
curate time-aware recommendation models. We will next discuss how
the three main collaborative filtering models – baseline estimates, near-
est neighbors, and latent factors – can be extended to account for a
temporal dimension. It is worth noting that all three solutions were
developed as components for the same hybrid model [Koren, 2009].
However, each model uses its own technique to account for temporal
changes.
5.10.3.1 Baseline Estimates with Temporal Dynamics

The goal of baseline estimates is to capture the average user and item
rating biases, as well as the global rating average. Recall that the stan-
dard baseline estimate of a rating is defined as follows (section 5.6.1):
bui µ bu bi (5.144)
in which bu and bi are the average user and item biases, respectively.
With the added assumption that user and item biases can change over
time, the time-aware version can be defined as
bui µ b u pt q b i pt q (5.145)
in which bu ptq and bi ptq are the functions of time that need to be
learned from the data. Parameter t can be defined as the number of
days measured off from some date zero in the past. In practice, user
and item components can have very different temporal dynamics and
properties, so one might need two different solutions for these two
functions [Koren, 2009]. In many practical applications, item popular-
ity changes slowly over time and each item has relatively many ratings.
Therefore, the time range can be split into multiple time intervals (for
example, several weeks each), and the item bias can be estimated for
each interval independently. This leads to the following simple time-
aware model for the item component:
b i pt q b i bi,∆t (5.146)
in which bi is the global stationary bias, ∆t is the time interval into

which t falls, and bi,∆t is the item bias estimate for this time bucket.
Note that the time-dependent part of the bias can be estimated only
for historical dates and has to be set to zero if the entire ∆t period is
in the future. It may seem strange that this model does not attempt
to extrapolate the time trend into the future, but it is important to
keep in mind that baseline estimates are used as a component in more
advanced models and their goal is to remove the observed trends and
refine the signal; extrapolation can be done by the remaining parts of
the model.
Although this approach can work well for items, it may not be ef-
ficient for users for at least two reasons. First, one user generally has
far fewer ratings than one item, so the bias cannot be reliably esti-
mated even for relatively large time intervals. Second, user biases can
change much faster than item popularity, and this requires the use of
even smaller time buckets. This problem can be solved by modeling
the user bias drift as a simple function, instead of point estimates. For
instance, one can use the following functional form to model the drift:
du ptq sgnpt tu q | t tu |β (5.147)
in which sgnpxq is a sign function that equals one when x is positive

and equals 1 when x is negative, tu is the mean rating date, and β
is the constant parameter selected by using cross-validation. Roughly
speaking, this is a linear function that can be bent down by setting the
366 recommendations
parameter β 1, as shown in Figure 5.25. The time-aware user bias

can then be defined as
b u pt q b u wu du ptq (5.148)
in which the stationary part bu and the scaling factor wu should

be learned from the data for each user u. This model can be improved
further by including additional terms that capture the day-specific vari-
ability of the user bias, such as the number of ratings provided by a
user on a given day.
β=1
du(t)
β=0.5
β=0.25
tu t
Figure 5.25: The user bias drift function for different values of β. The function
is linear when β 1.
5.10.3.2 Neighborhood Model with Time Decay

The nearest neighbor model predicts a rating for a given pair of user
and item by averaging ratings from similar users or items. The ratings
can be averaged by using a heuristic similarity measure or weights
learned from the data. For example, one of the basic item-based neigh-
borhood models is defined as follows (we discussed more complete
and practical versions of this model in section 5.7.4):
¸
p
rui bui wij ruj buj
(5.149)
j P Qkui
in which Qk ui is the neighborhood of item i in the set of items rated
by the user, buj is the baseline estimate, and weight wij is a measure of
the similarity between items i and j learned by using gradient descent
or, alternatively, defined as some heuristic similarity measure, such as
the Pearson correlation coefficient. Regardless of the weight estimation
approach, we can generally assert that if a user gives the same ratings
to two items i and j, it positively contributes to the corresponding

weight wij , which, in turn, boosts the corresponding rating ruj . It can
be argued that ratings can lose their relevancy over time, so the fact
that a user equally liked two items in the past should not necessarily
contribute to the similarity of these items in the future because of the
transient nature of user taste and identity [Ding and Li, 2005; Koren,
2009]. This consideration can be taken into account by adding a time
decay factor into the model to discount old ratings:
¸
ecu |ttuj | wij ruj buj

p
rui t p q bui (5.150)
j P Qkui
Decay rate cu is generally user specific and should be learned as an
additional variable as a part of the gradient descent process.
5.10.3.3 Latent Factor Model with Temporal Dynamics

The latent factor models predict ratings by computing a correlation
between user and item representations in the latent factor space. These
models can be extended to account for temporal effects by using the
techniques we developed earlier for baseline estimates. Recall that the
most basic latent factor model is defined as
p
rui pu qTi (5.151)
in which pu and qi are the k-dimensional user and item latent factor
vectors, respectively. One can account for variability of user tastes by
adding a time drift term to each element of the latent factor vector,
similar to the corresponding expression 5.148 for baseline estimates:
pus ptq pus wus du ptq, 1¤s¤k (5.152)
in which pus is the stationary term, du ptq is the drift defined by

function 5.147, and s is the latent dimension index. Both the stationary
and time-dependent parts of this model are estimated from the data by
using gradient descent.
In practice, one will most likely use a more advanced latent factor
model than basic solution 5.151. Fortunately, the time-dependent user
factor defined by expression 5.152 can be utilized in the majority of
latent factor models. One particularly important case is the SVD++
model defined in expression 5.118. Inserting the time-dependent user
factor into the SVD++ model, we obtain a model called TimeSVD++:

¸
µ pu ptq | Iu | 2
1
p
rui bi bu yj qTi (5.153)
j P Iu
368 recommendations
The TimeSVD++ model is one of the most accurate non-hybrid col-

laborative filtering models, and it is often considered to be a pinnacle
of recommender system engineering [Koren, 2009].
5.11 non-personalized recommendations
The quality of recommendations is generally determined by the ability

of a recommender system to recognize the intent of the user and find
offerings that can fulfill this intent. As a result of the wide variability
of user intents, recommender systems benefit greatly from accessing
and utilizing personal and behavioral data; systems that do not access
and use this data are likely to deliver poor quality recommendations.
Roughly speaking, we can expect that recommendations created for
everybody, that is, for an abstract average user, will actually be good
for almost nobody because only a few real users will precisely match
the average profile. However, recommender systems have to deal with
the fact that the personal and behavioral information may be available,
incomplete, or unreliable. For example, online systems often deal with
anonymous users who have no interaction history or have only a very
limited one collected during the current session. Such cases are very
common, so it is important to extend the range of recommendation
algorithms with non-personalized methods that can be used in a stan-
dalone mode or can be combined with personalized recommenders
into a hybrid system. Non-personalized recommendations are not nec-
essarily as accurate as personalized ones, but they can still be an effec-
tive solution for certain applications such as product cross-selling.
5.11.1 Types of Non-Personalized Recommendations
Non-personalized recommendation methods can generally be viewed

as an extreme case of contextual recommendations where the signals
about the target user are completely missing and the recommended
items are selected based on the contextual and background informa-
tion alone. The context, however, can include behavioral data about
other users, so the common user–item interaction patterns and item
popularity statistics can be known to the recommender. Let us con-
sider several examples of non-personalized recommendations that are
frequently used in practice:
popular items Popular categories and brands are often highlighted
in user interfaces to simplify navigation, and popular products
are often promoted in sections like Best Sellers or Top 10. This type
of recommendation does not necessarily use any request-level
context, but it relies on dynamically updated selling or browsing
5.11 non-personalized recommendations 369
statistics that can be considered as background context. From the

predictive modeling perspective, these methods simply exploit
the fact that the most frequent purchasing intents are the best
predictions for the purchasing intent of the target user, given that
no additional information is known.
trending items A recommender system can recommend products

that are trending upwards, instead of top sellers, based on the
assumption that such recommendations can be more serendipi-
tous. This approach, in particular, can better promote long-tail
items because even a slow-moving product can have popularity
bursts because of advertising or social media activity. A recom-
mender of trending items typically scores items based on some
smoothed version of the sales volume change history. For exam-
ple, the scoring function for item i can be defined as
spiq 1.00 ∆v1 piq 0.50 ∆v2 piq (5.154)
in which ∆v1 piq is the relative sales volume change (in percent)
for the previous day and ∆v2 piq is the volume change two days
ago.
new releases Some recommender systems highlight and promote

new items or items recently added to the assortment.
similar items In online marketing channels, non-personalized rec-

ommendations can often be created based on the browsing con-
text. One of the most typical examples is the More like this or You
might also like type of recommendations shown on the product
details page, that is, in the context of a certain product. Such
recommendations can often be created by using standard search
methods. For instance, one can define the distance between two
products as the weighted average of the TFIDF distances be-
tween the corresponding product document fields, and the most
similar items can be recommended. This strategy relies on the
item content and, consequently, tends to produce trivial non-
serendipitous recommendations.
frequent patterns Best seller recommendations use only a small

fraction of the information available in the purchase and brows-
ing histories, namely, the overall sales statistics. Contextual infor-
mation such as the currently browsed product or category is also
unused, which negatively impacts the cross-selling relevance of
the recommendations. More targeted recommendations can be
created by analyzing purchasing and browsing patterns and de-
tecting items that are frequently bought together with a given
370 recommendations
item or set of items. This type of recommendation is often pre-

sented on the product detail pages in sections like Customers who
bought this item also bought or What other items do customers buy
after viewing this item. We will discuss this approach in detail in
the next section.
Although we categorize the above methods as non-personalized, it
is important to understand that most of these methods can achieve at
least some level of personalization through segmentation and granu-
lar contextualization. For example, a news recommender system can
ask users to select the topics that they are most interested in (such
as politics, science, sports, etc.) and then recommend most popular or
trending items within the specified categories, instead of making rec-
ommendations based on the global statistics. A recommender system
can also filter or re-rank personalized recommendations based on pop-
ularity statistics or release dates.
5.11.2 Recommendations by Using Association Rules
A non-personalized recommender system can analyze historical trans-

actions from different marketing channels, including stores and on-
line, to discover the typical dependencies between products that can
be used to make recommendations. For example, if two items are fre-
quently bought together, it can indicate that the second item would be
a reasonable cross-selling recommendation for users who are currently
browsing the first item and vice versa. The recommendations are then
created in the context of an individual item (e. g., when a user browses
a certain product details page) or multiple items (e. g., when a user has
already added several products to the shopping basket). It is important
to note that the boundary between personalized and non-personalized
recommendations is quite blurry in this case. For example, recommen-
dations that are created in the context of a certain product and shown
on a product details page can be viewed as an integral static part of this
page. It is the same for all users and, consequently, non-personalized.
However, if the same context with a single item is persistently attached
to a user and their experience is changed based on this context else-
where, it can be viewed as personalization. A context that includes
multiple items can definitely be interpreted as an interaction history or
implicit feedback. In this case, recommendations made on the basis of
patterns discovered in transactional data can definitely be categorized
as collaborative filtering.
If the goal is to make a recommendation based on the currently
browsed item or multiple items, we will be most interested in discov-
5.11 non-personalized recommendations 371
ering regularities in historical transactions in the form of the following

rules:
if a user purchases items X ti1 , i2 , . . .u,

then that user will also purchase items Y tj1 , j2 , . . .u
The sets of items X and Y are called itemsets, and the association
rule described above is denoted as X Ñ Y. For example, the rule
tpasta, wineu Ñ tgarlicu indicates that if users buy pasta and wine
together, they are likely to also buy garlic. The number of association
rules that have at least some support in data, that is, there exists at
least one transaction where itemsets X and Y are bought together, can
be very high. The goal of a recommender system, however, is to find
rules that correspond to persistent patterns that can be used as predic-
tors of user behavior. We need to introduce more formal rule quality
metrics to select such rules.
Let us assume that we have a set of transactions T where each trans-
action is a collection of items purchased together. Support for itemset X
can then be defined as the fraction of the transactions that contains all
items from the itemset, that is, the empirical probability of X:
support pXq
|t : Xt|
tPT
|T | , (5.155)
Support for an association rule is the fraction of transactions that

contains both itemsets of the rule:
support pX Ñ Y q support pX Y Y q (5.156)
A high support level ensures that the rule corresponds to a persistent

pattern where the itemsets are frequently bought together. It can, how-
ever, be the case that these itemsets are frequently bought separately
as well, so high support merely confirms the fact that both itemsets
are popular, even if no dependency between them really exists. This
aspect is measured by the confidence of the rule, defined as the fraction
of transactions containing X that also contains Y:
support pX Y Y q
confidence pX Ñ Y q
support pXq
(5.157)
Confidence can be interpreted as the conditional probability of find-

ing itemset Y in the transaction given that this transaction contains X,
that is, Pr pY | Xq. Note that support and confidence are defined based
on the purchase probability and, consequently, can be linked to mon-
etary measures such as revenue [Ju et al., 2015; Geng and Hamilton,
2006]. For example, one can roughly estimate the expected revenue of
372 recommendations
the rule (and the recommendation created based on this rule) as fol-
lows:
¸
revenue pX Ñ Y q support pX Ñ Y q price piq (5.158)
i PY
given that a user is about to buy item X. More accurate estimates
of monetary metrics can be obtained by using the uplift modeling
techniques that we discussed earlier in the context of promotion op-
timization. A recommender system typically needs association rules
with high support and confidence levels to ensure that these rules are
reliable and discriminative. The creation of such rules from a given
transaction history is a standard data mining problem, known as fre-
quent itemset mining, affinity analysis, or market basket analysis. This
problem can be solved by using a wide range of specialized algorithms,
such as Apriory or FP-growth.
example 5.7
Let us consider a detailed example that illustrates how association

rules can be used to create recommendations. In contrast with tradi-
tional collaborative filtering, association rule learning requires more
granular transaction-level data, but the transactions do not need to be
linked to individual users (it does not matter which user performed
which transaction). We will analyze the sample transaction history for
a grocery shop presented in table 5.9.
Transaction ID Items
1 milk, bread, eggs
2 bread, sugar
3 milk, cereal
4 bread, cereal
5 milk, bread, sugar
6 cereal, milk, bread
7 bread, cereal
8 milk, cereal
9 milk, bread, cereal, eggs
Table 5.9: Sample transaction history.
In order to create recommendations, for example, for milk, we run

an association rule learning algorithm on the transaction history with
the constraint X tmilku and sort the rules by the corresponding
confidence levels. The result is shown in table 5.10. The confidence level
for the milk Ñ cereal rule, for example, is equal to 4/6 because there
are four transactions that contain both these items and six transactions
5.12 multiple objective optimization 373
that contain milk. The recommended items are then obtained from the
right-hand sides of the rules, so that the list of recommendations for
milk includes cereal, bread, eggs, and sugar, in order of relevance.
Rank Rule Support Confidence

1 milk Ñ cereal 4/9 4/6
2 milk Ñ bread 4/9 4/6
3 milk Ñ eggs 2/9 2/6
4 milk Ñ sugar 1/9 1/6
Table 5.10: Association rules for milk.
Note that although market basket analysis is a method of unsuper-

vised learning, we are essentially solving classification and feature se-
lection problems in this particular setting because the context (milk, in
this example) can be viewed as a training label, other items in the trans-
action can be viewed as features, and the goal is to identify the most
predictive features. In the association rule approach, these features cor-
respond to the right-hand side of the rules with high confidence levels.
N
As we discussed at the beginning of this section, association rules

can be used to create personalized recommendations as well, depend-
ing on how many items are included in the context (left-hand side of
the rule) and the semantic meaning of the context. For usage cases
such as anonymous web session personalization, the association rule
approach can be an effective alternative to other collaborative filtering
methods, such as neighborhood models, both in term of accuracy and
computational complexity [Mobasher et al., 2001].
5.12 multiple objective optimization
All of the recommendation methods discussed above are essentially

driven by a single objective – to provide the best semantic match or
predicted preference score. However, recommendation accuracy might
not be the only concern of the recommender system design: a marketer
might also be interested in incorporating multiple competing objectives
into the recommendations offered to the customers. For instance, gro-
cers might be interested in boosting perishables with a shorter shelf
life, fashion stores might want to promote sponsored brands or sea-
sonal collections, and a wide range of retailers can benefit from recom-
mending products with a higher margin or from taking into account
product stock levels to avoid stockouts [Jambor and Wang, 2010].
374 recommendations
One possible approach for implementing a multi-objective recom-

mender system is to mix the semantic relevance signals with signals
that correspond to the secondary objectives. From this perspective,
multi-objective methods can be compared to hybrid models where mul-
tiple signals are also mixed together to achieve optimal results. The key
difference, however, is that hybrid methods typically use a standard
loss function, such as the average rating prediction error, as the opti-
mization goal, whereas multi-objective recommenders use more cus-
tomized optimization targets. In this section, we will consider a recom-
mender system with multiple objectives that was developed and tested
in practice on a large scale by LinkedIn, an employment-oriented so-
cial networking service [Rodriguez et al., 2012]. In the case of LinkedIn,
the primary objective was to recommend candidates who semantically
match a job description and, also, as a secondary objective, display a
job-seeking behavior.
We start with the idea that the recommendations produced by the
core recommendation algorithm can be re-ranked by using a function
that optimizes some secondary objective, with the condition that devi-
ations from the original relevance-based ranking should be penalized.
Let us first consider the case with a single secondary objective and
define a relatively abstract framework that can be adapted for a wide
range of objectives. First, we can use the core recommendation algo-
rithm to rank all m items for each of n users. Let us denote these
original recommendations as an n m matrix Y, in which each row
corresponds to a user, each column corresponds to an item, and each
element is a rank of an item in the recommendation lists. We assume
that each user is actually presented with only k ! m recommenda-
tions, but all items are scored originally to give enough choice for the
re-ranking function. In practice, we do not necessarily need to rank all
m items and can limit the number of recommendations by some num-
ber that is substantially greater that k. Each recommended item can
then be scored according to the secondary objective, and we denote an
n m matrix of these scores as X. For instance, this matrix can con-
tain product gross margins. Note that the score can be a function of
both the item and its position in the recommendation matrix Y, that is,
the user and rank. The optimization problem can then be defined as
follows:
max g pϕ pY, X, wqq
w
(5.159)
subject to d topk pYq , topk pϕ pY, X, wqq ¤c
in which
• g represents the utility function that evaluates the quality of the
recommendations from the secondary objective perspective.
5.12 multiple objective optimization 375
• ϕ is a composite ranking function that combines pairs of rows

from matrices X and Y into a new list of recommendations that
balances the two objectives.
• w is the parameter (weight) that defines the mixing balance be-

tween the two objectives. This parameter is the subject of opti-
mization.
• topk pq denotes the first k elements with a maximal ranking score.
This operation truncates the original matrices X and Y to the size
of n k.
• d is the distance function that measures the discrepancy between

the two recommendation matrices. One possible way to measure
the discrepancy between the two score vectors x and y is to cal-
culate the sum of squared errors between their histograms:
b̧
d px, yq pHpxqi Hpyqi q2 (5.160)

i 1
in which the histogram Hpxq is a vector with b elements (buckets)

and each element corresponds to the number of scores in x that
fall into the corresponding range. The overall distance between
the matrices is then defined as the sum of distances over all users,
that is, rows of the matrices.
• c is a threshold that limits the discrepancy between the original

and re-ranked recommendation lists.
The main idea of the optimization problem above is to increase the
utility of the re-ranked recommendations that mix the relevance scores
with the secondary objective but to penalize the difference between the
original relevance-based recommendations and the re-ranked recom-
mendations to make sure that relevance will not be completely sacri-
ficed in pursuit of the secondary objective. The design of function ϕ
should include tunable parameters that control the trade-off between
the two objectives and will be the subject of optimization. This ap-
proach can be straightforwardly extended to incorporate more than
two objectives and multiple divergence constraints. Denoting the num-
ber of objectives as q, we can define the following multi-objective opti-
mization problem:
max g pϕ pY, X, wqq
w
(5.161)
subject to dj topk pYq , topk pϕ pY, X, wqq ¤ cj
in which X is now the n m q matrix of scores, w is a vector of q
weight parameters, and j iterates over all divergence criteria.
376 recommendations
We can illustrate how the optimization model above can be adapted

to practical problems by using a couple of examples. First, consider the
case of a retailer who wants to incorporate a revenue objective into the
recommendation scores. The overall utility function can be defined as
the expected gross margin, with the assumption that Mpiq P r0, 1s is a
normalized gross margin of item i and the probability of purchase is
modeled as a reciprocal to the ranking position (i. e. , the lower the item
in the list of recommendations, the lower the probability of conversion):
Mpiq
m̧
g pzq I pz i ¤ kq
1
(5.162)
k

i 1
zi
in which z is the vector of ranks produced by the composite rank-

ing function ϕ and I is an indicator function that equals one if the
argument is true and zero otherwise. As the secondary objective is the
expected gross margin, matrix X is straightforwardly defined as
xui Mpiq (5.163)
The composite ranking function can then be specified as the mix

of the original relevance score y, produced by the core recommender
algorithm, and the margin score x:
z ϕ py, xq : zi yi xwi (5.164)
in which w is the parameter that controls the trade-off between the

relevance and pitching of high-margin products. This parameter is the
subject of the optimization in problem 5.159.
Our second example of re-ranking according to a secondary objec-
tive is the boosting of featured items, such as on-sale products or per-
ishables. The utility function can be specified as the average number of
featured products in the short list of k recommendations:
m̧
gpzq Fpiq I pzi ¤ kq
1
(5.165)
k

i 1
in which Fpiq is a feature label that equals one if the item is featured
and zero otherwise. The matrix X is defined as
xui Fpiq (5.166)
The composite ranking function combines the relevance score and

feature labels with a trade-off parameter w, which is the subject of
optimization:
z ϕ py, xq : zi yi wx i
(5.167)
5.13 architecture of recommender systems 377
The ranking function above can be straightforwardly extended to

incorporate multiple separate features, each of which contributes to
the final ranking score in accordance with its own trade-off parameter
(recall that X can be an n m q matrix of scores, so that x can be a
q m matrix):
z ϕ py, xq : zi yi w1x wx2 . . . wqx
1,i 2,i q,i
(5.168)
Optimization problem 5.159 depends on the ranking function, so the
standard optimization methods for smooth functions, such as gradient
descent, are not directly applicable. In a general case, this problem
can be approached by using learning to rank algorithms [Rodriguez
et al., 2012]. In many practical applications, however, we can expect to
have only one or two parameters w. In this case, the problem can be
straightforwardly solved with an exhaustive search over all possible
values.
5.13 architecture of recommender systems
Thus far, we have discussed a wide range of recommendation models

and algorithms, as well as methods that can be used to combine mul-
tiple models into a hybrid or adjust recommendations based on con-
textual information or secondary objectives. A recommender system,
however, is more that just the implementation of some algorithm. It is a
complex software system that includes multiple components and mod-
ules that connect the recommendation model with the outside world
and make it function. In this section, we discuss a possible reference ar-
chitecture of a recommender system, presented in Figure 5.26, as well
as some possible variations and trade-offs [Jack et al., 2016].
user interface A recommender system can be connected to mul-
tiple user interfaces, such as websites, email service providers,
mobile notifications, or news feeds. These channels interact with
the core recommender system through an interface that receives
the contextual information as an argument and returns ranked
recommendations. In the most basic cases, the interface of a rec-
ommender system can be as simple as a side bar containing rec-
ommendations on a web page. Services that heavily depend on
recommendations, such as online video services, often provide
much more comprehensive interfaces that include multiple sec-
tions, such as personalized recommendations, popular items, and
latest trends.
data collection The total number of data sources used by a rec-
ommender system can be quite high. One reason is that indus-
trial recommender systems often use many different algorithms,
378 recommendations
Figure 5.26: High-level architecture of a recommender system.
either for hybridization or experimentation purposes, so the sys-

tem should consume enough data to support content-based, col-
laborative filtering, and popularity-based methods. The system
may also have access to some external or third party data or
signals that can help to refine the recommendations. This of-
ten requires the creation of a comprehensive data collection in-
frastructure that consolidates user profiles (e. g., preferences and
personal information), item information that generally requires
some sort of content analyzer to be used to parse and transform
the raw content data, and the user–item interaction data, which
is typically captured by the user interface as ratings and other
types of feedback.
modeling The collected, cleaned, and consolidated data is used to

create recommendation models. The models are deployed into
a repository and periodically retrained to catch up with data
changes. The model can be evaluated in the modeling (offline)
5.14 summary 379
tier, service (online) tier, or both. If the model is completely eval-

uated offline, the output of this process is a set of recommen-
dations for all users. The recommendations can updated on a
schedule, for example, on a daily basis, and loaded into the rec-
ommendation service. If the model is partially evaluated online,
the offline part precalculates only certain data elements, such as
item similarity matrices or latent factor vectors. This approach al-
lows one to do the heavy computations offline in batch mode but
to keep the flexibility of online recommendations. Industrial rec-
ommender systems typically have a repository with multiple rec-
ommendation algorithms that are constantly updated and tested.
recommendation service The primary purpose of the rec-

ommendation service is to complete the evaluation of the
recommendation model and serve the recommendations to the
user interfaces. The recommendation service can implement a
number of functions related to contextualization and operational
control. First, recommendations produced by the core algorithm
can be postprocessed to make additional improvements and
adjustments. For example, a service can track the recommen-
dations already seen by the user in real time and rotate or
randomly shuffle the list of recommendations to make the
user experience more dynamic, productive, and serendipitous.
Second, the recommendation service can check the quality of
recommendations by using validation rules and log issues or can
make automatic adjustments in the case of validation failures.
Examples of automatic quality checks include validation of the
total number of recommendations in the list and monitoring of
the algorithm execution time.
5.14 summary
• Digital channels enable marketers to carry extremely wide and deep

assortments with a large number of slow-moving niche products.
This is one of the key differentiators in a comparison with tradi-
tional distribution channels, where the assortment is limited by dis-
tribution costs.
• Extremely wide and deep assortments with a long tail of niche prod-
ucts create a need for efficient discovery services, including search
and recommendations.
• The main goal of recommendation systems is to provide customers

with relevant offerings when the purchasing intent is not explicitly
380 recommendations
expressed. This can be viewed as a counterpart to search services,

where the intent is explicitly specified as a search query.
• Recommendation systems can generally leverage user–item interac-

tion data, which include explicitly provided ratings and implicitly
collected browsing histories, catalog data, and contextual informa-
tion. The main output of the recommendation service is a ranked
list of recommended items.
• One of the main inputs of a recommender system is customer rat-

ings. The ratings are typically represented as a rating matrix, where
rows are customers, columns are items, and each element is a numer-
ical rating value. Rating values can also be attributed with contextual
information, such as time stamps or the marketing channel that the
rating came from. In practice, rating matrices are very sparse.
• The main business objectives of recommender systems include rel-

evance, novelty, serendipity, and diversity of the recommendations.
The relevance of recommendations can be measured through rating
prediction accuracy and ranking accuracy. Quantitative metrics can
also be defined for novelty, serendipity, and diversity.
• The most important families of recommendation algorithms are con-

tent filtering and collaborative filtering. These basic algorithms can
be extended by using hybridization, contextualization, and admix-
ing of additional objectives and signals.
• Content-based filtering recommends items that are similar to the

items that the user liked in the past. Content-based filtering can be
viewed as an item classification problem. The key advantages of con-
tent filtering are the ability to recommend items based only on a
user’s own ratings and to recommend new unrated items, as well
as the interpretability of the results. The main shortcomings are the
complex feature engineering that is required for content classifica-
tion and the bias toward trivial recommendations.
• Content filtering can use document similarity measures and other

search methods, such as latent semantic analysis or latent Dirichlet
allocation, to find the most similar items. An alternative approach is
to use text classification models, such as the Naive Bayes classifier.
• Collaborative filtering uses the rating matrix to find items or users

with similar rating patterns. It generally creates more diverse and
serendipitous recommendations than content filtering. Collaborative
filtering can be viewed as a matrix completion problem.
5.14 summary 381
• The most important collaborative filtering methods include nearest

neighbor recommenders and latent factor models.
• Multiple recommendation models can be combined together into a

hybrid model. Hybrid recommenders use the same techniques as
signal mixing pipelines in search services. A hybrid recommender
can switch between relevance signals from different models, blend
them together, or use the outputs from one model as inputs for other
models.
• Contextual recommendations use additional attributes such as

time or location to make more targeted recommendations. These
attributes can be considered additional dimensions that expand
the rating matrix into a multidimensional cube. A recommender
system can use contextual information to prefilter the input data,
postfilter the recommendations, or extend the inputs of the rating
prediction model. Contextual recommendations use the same ideas
as controlled precision reduction in search services.
• Recommendations for unknown users and users with limited inter-

action histories and profiles are important usage cases. Recommen-
dation systems can leverage basic sales statistics (best sellers), con-
tent (similar items), and purchasing patterns (frequent itemsets) to
make non-personalized or partially personalized recommendations.
• The main components of a recommender system include the user

interface, data collection tier, modeling tier, and a recommendation
service.
6
P R I C I N G A N D A S S O RT M E N T
The problem of price management has a very long history. The fun-
damental aspects of pricing have been studied for centuries to explain
the interplay of supply and demand on a market. This has resulted in
the development of a comprehensive theory that describes the strate-
gic aspects of pricing, such as price structures, relationships between
price and demand, and others. These methods provide relatively coarse
price optimization methods that can, however, inform strategic pricing
decisions. The opportunity to automatically improve tactical price de-
cisions was first recognized and seized by the airline industry at the
beginning of the 1980s and can be partially attributed to the advance-
ment of digital reservation systems that enabled dynamic and agile
resource and price management. This required the development of a
totally new set of optimization methods that were later adopted in
other service industries, such as hotels and car rentals. This new, truly
algorithmic, approach, commonly referred to as revenue management or
yield management, has clearly demonstrated the power of automated
price and resource management by multiple cases of late adopters of
new techniques being bankrupted or defeated by the pioneers of auto-
mated price management.
Price management is closely related to other programmatic services,
especially promotions and advertisements. Price management meth-
ods can be used both to optimize discount values of promotions and
to price advertising and media resources sold to service clients. We
will start this chapter with a review of the basic principles of strategic
pricing and price optimization. We will then continue with the devel-
opment of more tactical and practical demand prediction and price op-
timization methods for market segmentation, markdowns, and clear-
ance sales. We will also briefly review the major resource allocation
methods used in service industries to set booking limits. Finally, we
will consider the assortment optimization problem, for which we can
reuse some of the building blocks developed for price management.
383
384 pricing and assortment
6.1 environment
As we will demonstrate later in this chapter, price management is a crit-

ically important determinant of enterprise profitability and, eventually,
survival. Consequently, real-life price management processes often in-
clude multiple layers of decision making, which range from strategic
executive decisions down to micro-decisions at the level of individual
transactions. We are distilling this complexity into the relatively simple
model presented in Figure 6.1. This model does not explicitly account
for some strategic aspects of pricing, but it captures the major features
of the pricing environment that are important for micro-decisioning:
• We generally assume that a company sells some products to its
customers and earns the following profit G for each product i:
Gi Qi pPi Vi q Ci (6.1)
in which Qi is quantity sold, Pi is price, Vi is variable costs (e. g.,

wholesale product price), Ci is fixed costs associated with the
product, and i iterates over the assortment of products offered
to the customers. Most methods considered in this chapter are
focused on maximization of the profit G as a function of price,
although it is important to keep in mind that this optimization
can be a subject of external strategic constraints. For example, a
company can choose a competitor-oriented strategy that pursues
market share, rather than profits, and restrict the optimization
process with a price-match policy.
• The profit is a function of the quantity sold which, in turn, de-

pends on the demand. The key assumption made by the model
is the heterogeneity of the demand, that is, the variability of the
demand according to one or many dimensions, such as customer
segments, store locations, seasons, classes of service, and so forth.
This provides an opportunity to differentiate prices according to
these dimensions as well and, thus, tune profits at the level of
individual customer segments or time intervals. We will spend
most of this chapter discussing price structures and price opti-
mization techniques that deal with different aspects of the het-
erogeneity of the demand.
• The demand is a function of price and other variables that range

from competitor prices to weather. In the most basic case, the
revenue management system can optimize prices by creating re-
gression models for individual product demands and finding op-
6.1 environment 385
Figure 6.1: Revenue management environment.
timal prices that maximize the profits through maximization of

the demand:
Pi
opt
argmax Qi pPi q pPi Vi q Ci (6.2)
Pi
in which Qi pPi q is the demand prediction model for product i.

In practice, the problem is typically more complicated than this
because of various constraints and interdependencies.
– One notable example is stock level constraint – if a seller
has a limited stock of a product, the quantity sold Q is the
minimum of the demand and the stock level.
– Another important factor is dependent demands – because
products in one category are often substitutable, a change in
the price of one product can make customers switch to an-
other product. This makes the optimization problem more
complex because product prices have to be optimized jointly,
not individually.
– Finally, a seller can pursue additional goals that translate

into additional constraints. For example, a fashion retailer
can aim at selling out the inventory by the end of the season.
Consequently, the revenue management system has a number
of inputs, including historical demand data, fixed and variable
costs, stock levels, and other business constraints.
• Although optimization of the price level is a natural task for a

revenue management system, many environments provide other
important controls. One set of such controls is related to how
the price is communicated to the customers. As we will discuss
later, it is often more efficient to communicate price changes as
discounts and special offers, rather than to change the baseline
prices. This establishes a link between price optimization and
discount optimization tasks. The second set of controls is related
to product availability and classes of service. A classic example
is an airline that requires low-cost tickets to be booked in ad-
vance and makes this option unavailable a few days before the
flight. Finally, the revenue management system can control the
assortment of products, their presentation, and the placement
options. Examples include shelf-space optimization by removing
low-performing products and store-layout optimization by plac-
ing related products together to leverage cross-selling opportuni-
ties.
The model above outlines the basic areas where programmatic price
management methods can be applied. We will spend the next few
pages introducing the basic concepts and principles of pricing and will
then start to design price optimization methods for the environment
described in this section.
6.2 the impact of pricing
Pricing plays a critically important role in the enterprise economy be-

cause prices are the key determinants of revenues and profitability. The
right pricing decisions can deliver great profitability improvements,
whereas the wrong decisions can have grave consequences. One of the
reasons is that the price determines how the product or service is po-
sitioned on the market and perceived by customers. Prices that are too
low undermine a firm’s sustainability by leaving unharvested profits
and setting the wrong expectations about product value and quality
for the customers; prices that are too high harm sales and the firm’s
reputation, which slows down business growth. Another reason is that
the quantitative dependency between prices and profits is very strong
6.2 the impact of pricing 387
in most industries and enterprises, so price often dominates in profit

equations. Let us examine an example that illustrates the importance
of pricing decisions.
example 6.1
Consider an imaginary apparel retailer that sells 100,000 garments

monthly at $40 per item, with a wholesale price of $25 per item and

fixed costs of $500,000 per month. We can express the profits of the
retailer as a function of price, costs, and sales volume by using the
following basic profit equation:
G Q pP V q C (6.3)
in which Q is quantity sold, P is price, V denotes variable costs, and C
denotes fixed costs. Consequently, the baseline profit will be
G 100,000 p$40 $25q $500,000 $1,000,000 (6.4)
The retailer can choose from several strategies to improve the base-
line profit. One alternative is to increase the quantity sold by investing
in marketing campaigns or the development of new sales channels.
Other approaches are to raise the sale price, change supplier to reduce
the variable cost, or cut down the fixed costs. We evaluate all of these
strategies in table 6.1 and calculate how a one percent change in sales
volume, price, variable cost, and fixed costs will impact profit. It turns
out that profits are most sensitive to changes in price, which indicates
the high importance of pricing decisions.
Baseline +1% in Q +1% in P -1% in V -1% in C

Q 100,000 101,000 100,000 100,000 100,000
P $40.00 $40.00 $40.40 $40.00 $40.00
V $25.00 $25.00 $25.00 $24.75 $25.00
C $500,000 $500,000 $500,000 $500,000 $495,000
G $1,000,000 $1,015,000 $1,040,000 $1,025,000 $1,005,000
∆G% +1.5% +4.0% +2.5% +0.5%
Table 6.1: Quantitative example that illustrates how changes in prices, costs, and
quantity sold influence profits.
Although we have used somewhat arbitrary numbers in this exam-

ple, this pattern prevails in a huge variety of enterprises across many
industries. For example, a study by McKinsey and Associates analyzes

the profits of 2,463 companies and comes to the conclusion that a 1%
change in price results in 11.1% profit improvement, whereas a 1%
improvement in sales volume, variable cost, or fixed costs yields im-
provements of 3.3%, 7.8%, and 2.3%, respectively [Marn and Roseillo,
1992].
6.3 price and value
The development of automated price management systems requires us

to break down pricing into formal optimization problems that treat
profits merely as mathematical functions. On the other hand, price is a
complex matter that depends on the nature of the product, the compe-
tition, and customer psychology. In this section, we will start to bridge
the gap between the fundamental pricing problem and optimization
tasks by discussing what the price is and how it can be determined.
This discussion is very strategic and gives us very few clues about
how automated price management can be implemented, but it pro-
vides guidelines that will help us to develop more elaborate methods.
6.3.1 Price Boundaries
Economic theory argues that price is determined by the interplay of

supply and demand on the market. Although every product or service
has its production cost, which can sometimes be considered as a “fair”
baseline for the price, pricing requires us to delve into the valuation
logic of both the seller and the buyer.
On the one hand, we can assume that the seller of a product or
service has a minimum profitable price point. Selling above this point
generates profits; selling below this price incurs losses. In many cases,
this price point can be assumed to be equal to the marginal cost of the
product.
On the other hand, a buyer extracts a certain utility from the pur-
chased product. The utility depends on the functional properties of
the product, the customer’s ability to achieve useful goals by using
these properties, the availability of the product in the right place and
at the right time, and other factors. In some cases, a relatively precise
estimate of the utility can be obtained. For example, the utility of an
industrial electrical generator can be assessed through the price of the
electricity it produces. In other cases, only a gross approximation of
the utility can be provided. An example of this is innovative medical
equipment that may be valued as highly as a human life.
6.3 price and value 389
A seller and buyer can commit to a mutually beneficial transaction

if the marginal cost is below the utility. If this does not hold, they
both will be better off passing on the deal. Consequently, the price is
essentially a range, not a single point. At a very high level, the goal of
a price management system is to dynamically select an optimal point
in this range for individual transactions.
The marginal cost and utility often provide very broad price brackets
that can be useless in practice. For instance, the purchase of a bottle of
soda on a hot day can literally be a question of life and death, which
skyrockets the utility, but high competition keeps the price close to the
lower boundary. At the other extreme, the negligible marginal cost of
software distributions does not prevent the price from staying close to
the upper boundary determined by utility. More narrow brackets can
often be obtained by comparing a product or service with available
alternatives and carefully evaluating the product features.
If a comparable alternative is available, its price can be taken as a
baseline. A given product can be superior or inferior relative to the
alternative, and the difference in value can be estimated by building a
value exchange model [Smith, 2012]. A value exchange model estimates
the potential difference in price between two products by analyzing
and evaluating individual product features that can be advantageous
(contribute toward higher price) or disadvantageous (contribute to-
ward lower price), as illustrated in Figure 6.2. The final price, also
known as the exchange value, can be obtained by adding this differ-
ential value to the baseline price of the alternative.
Figure 6.2: Price boundaries and the exchange value.
The exact design of the value exchange model depends on the nature
of the products and their differences. In many cases, product features
can be evaluated with methodologies, such as conjoint analysis, that rely
on consumer surveys [Green and Srinivasan, 1978]. In certain cases, the
model can be created by analyzing the potential outcomes of choosing

one or another product. Let us consider two examples:
• The differential value of a more reliable product relative to a less
reliable alternative can be estimated by considering the probabil-
ity of failure and the potential replacement cost. If the price of the
alternative pA and the replacement cost pR are known, the price
of a new reliable product can be roughly estimated as follows:
p pR p1 Prpfailureqq ppA pR q Prpfailureq (6.5)
• Accessories and complementary products can incur switching

costs that can be incorporated into the value exchange model.
For example, manufacturers of razors and blades typically de-
sign their products in such a way that blades are not interchange-
able between brands, and, consequently, the exchange value of
a blade is boosted by the relatively high cost of switching to a
razor from a different brand.
Value exchange considerations can be factored directly into price
structures, which we will discuss in detail in later sections.
6.3.2 Perceived Value
The concept of utility may suggest that buyers make decisions by com-
paring their willingness to pay against the price: the product is pur-
chased if, and only if, the price is below the utility. This “rational be-
havior”, however, is not an adequate model of real consumers. Valu-
ation is a subjective process that depends on how exactly the value
and price are communicated to the prospect and how the prospect per-
ceives it. Failure to properly communicate the value or price can set
the wrong expectations and displace the price boundaries in the un-
desirable direction. Efficient communication, by contrast, can improve
the perceived value of a product or diminish the value of comparable
alternatives.
Value communication, at the first glance, might not look very rel-
evant for a discussion of algorithmic methods because it deals with
psychological aspects of value and price perception that can probably
not be codified in the software. It turns out, however, that analysis of
these psychological patterns can produce applicable rules that can be
incorporated into price structures and, consequently, can be accounted
for in price optimization problems.
One of the most solid frameworks that captures many important
aspects of value and price perception is prospect theory [Kahneman
and Tversky, 1979]. Prospect theory considers the evaluation process
6.3 price and value 391
from the risk assessment standpoint and can be characterized by the

following propositions:
point of reference The potential gains and losses associated with
a deal are evaluated by comparison with some point of reference.
The point of reference is based on past experience (e.g., the last
observed price for a given or similar product) or judgement and
tends to be persistent once it is set.
diminishing sensitivity Changes in gains or losses are sharply

perceived in the zone around the reference point but become less
noticeable as the magnitudes of gain or loss increase. The differ-
ence between $9 and $19 discounts appears to be substantial, but
the same ten dollar difference is not perceived equally valuable
for discounts of $719 and $729.
loss aversion Losses are generally perceived more sharply than

gains of the same magnitude. A loss of $100 is typically perceived
more important than a gain of $100.
risk aversion for gains Guaranteed gains are preferred to oppor-

tunistic gains of the same magnitude. A prospect who has the
choice of getting a guaranteed $450 or of winning $1000 with
a 50% chance (and a 50% chance of winning nothing) generally
prefers the first option.
risk seeking for losses In contrast to gains, potential losses are

preferred to guaranteed losses. A prospect who has the choice
of definitely losing $450 or of losing $1000 with a 50% chance
(and a 50% chance of losing nothing) generally prefers the second
alternative.
The propositions above imply a certain shape to the dependency be-
tween the real and perceived gains and losses, as shown in Figure 6.3.
The slope of the curve in the negative zone is steeper than in the pos-
itive zone, in accordance with the loss aversion hypothesis, but the
steepness decreases at both ends to follow the principle of diminishing
sensitivity.
Prospect theory suggests several important guidelines that can be
used to optimize price structures:
• Price rises (surcharges) are viewed much more negatively than
price drops (discounts). This is the reason why prices are typi-
cally communicated as a list price and discount. This allows the
list price to be kept constant and the discount value to be moved
in either direction, including personalized pricing, without ex-
plicit surcharges.
Figure 6.3: A prospect theory value function. The actual value increment of ∆
can be perceived as a relatively small gain, whereas the actual decre-

ment of the same magnitude ∆ can be perceived as a huge loss.
• The reference point should be kept at a high level. This com-

plements the previous point about list prices and discounts be-
cause reductions in baseline prices can lead to undesired persis-
tent shifts in the reference point.
• It is generally better to subdivide gain into multiple smaller gains.

A single gain will be discounted because of diminishing sensitiv-
ity, so a price structure with multiple smaller benefits can have a
higher perceived value.
• The risk-seeking behavior for losses suggests that some explicit

product benefits can be replaced by equivalent potential bene-
fits without substantial loss in perceived value. For example, flat
pack furniture requires time and effort to be spent in assembling
it, but reduced price and delivery costs can still be perceived as
unconditional advantages.
We will use some elements of prospect theory in price optimization
to account for behavioral factors that are inconsistent with the basic
principle of a rational consumer.
6.4 price and demand 393
6.4 price and demand
Utility, value exchange models, and other valuation methods help us

to predict the expected willingness to pay for a given product or ser-
vice. We can also recall that consumer choice theory (see Section 2.6.1
for details) provides us with the tools to predict a consumer choice in
the case of multiple available alternatives. Although one can attempt
to sell a product to each customer at a price that is derived directly
from an individual willingness to pay, we choose to start with the tra-
ditional problem of setting a common price structure for all customers.
This means that we need to deal with thousands or even millions of in-
dividual buying decisions that can be described in probabilistic terms.
Let us define the customer’s willingness to pay for a given product or
service as the maximum price acceptable for a customer. The customer
will buy a product if, and only if, the price is less than the willingness
to pay. We can describe the population of customers by using the dis-
tribution of the willingness to pay wppq: for every pair of prices p1
and p2 , the fraction of customers fpp1 , p2 q whose willingness to pay is
between p1 and p2 is
» p2
fp p 1 , p 2 q wppqdp (6.6)
p1
The demand function qppq, also referred to as the price–response func-

tion, can be expressed through wppq as follows:
»8
qppq Qmax wpxqdx (6.7)
p
in which Qmax qp0q is the maximum achievable demand for a

given seller. The demand function can be viewed not only as an aggre-
gate market metric determined by variance of willingness to pay but
also as a model of a single customer behavior, in the sense that a given
consumer might be willing to buy different quantities of product at dif-
ferent prices. In the latter case, willingness to pay should be considered
as willingness to pay per unit.
Mathematical analysis of the demand function can provide us with
additional insights and help to define useful metrics and properties.
First, let us take a closer look at the derivative of the demand function:
B
Bp qppq Qmax wppq (6.8)
As wppq is non-negative, the derivative is non-positive for any p,

which means that the demand function is downward sloping. The
slope of the demand function given by its derivative is a measure of

the price sensitivity. A steep slope means that customers are sensitive
to price changes and demand rapidly drops as the price rises, whereas
a shallow slope means that customers are relatively insensitive to price
changes. It is more common, however, to measure price sensitivity not
as a slope of the demand function but as the elasticity of demand, defined
as the ratio of the percent change in demand to the percent change in
price:
∆q{q

∆p{p
qpppq BBp qppq (6.9)
Although the elasticity in equation 6.9 is a function of price, meaning

that it can be different at different points of the demand curve, this term
is often used more loosely by assuming that the elasticity is roughly
constant in the range of interest, so the demand can be characterized by
a single value of . The elasticity of demand does not depend on mag-
nitudes of price or volume, so it provides a convenient way to measure
and compare price sensitivities. Elastic markets, with ¡ 1, respond
to a small change in price with a large change in demand. For example,
restaurant meals are reported to have an elasticity of about 2.3, which
means that a 10% increase in price can drive a 23% decrease in demand.
Inelastic markets, with 1, respond to price changes with a small
change in demand. For example, the price elasticity of motor gasoline
in the US was estimated to be around 0.04, which means that it takes a
50% increase in the price of gasoline to reduce automobile travel by 1%.
We should, however, distinguish elasticity for a category of goods and
for an individual brand in the category. Substitution of one category by
another is generally difficult, so category-level demand is relatively in-
elastic in many industries. A switch between brands is simpler, which
makes the demand curves more elastic from a single seller’s stand-
point.
We will now explore several commonly used demand models by
expressing them in terms of wppq and qppq.
6.4.1 Linear Demand Curve
A simple demand model can be derived under the assumption that the
willingness to pay is uniformly distributed in the range from 0 to the
maximum acceptable price P:
$
wppq unifp0, Pq
{
&1 P, 0¤p¤P
(6.10)
%
0, otherwise
The demand function can be obtained by integrating wppq in accor-

dance with equation 6.7:
»P
qppq Qmax wpxqdx
p

Qmax 1 Pp (6.11)
Qmax
P
p Qmax
We can see that the uniformly distributed willingness to pay results

into a linear demand function, as illustrated in Figure 6.4. We will
assume linear demand curves for the optimization of basic price struc-
tures because of their analytical convenience, although this is typically
a very gross approximation of real demand functions. One of the short-
comings of the linear demand model is that it assumes that every dollar
of price change yields the same increment in demand. This is generally
not true because price sensitivity is typically high near the point of ref-
erence (e. g., competitive prices) and diminishes as the price moves
away from it.
Figure 6.4: Uniform willingness to pay and the corresponding linear demand
curve. Note that we follow the traditional economic notation in the
right-hand plot by placing price on the vertical axis, although de-
mand is considered as a function of price.
6.4.2 Constant-Elasticity Demand Curve
The constant-elasticity demand function can be obtained from the def-

inition of elasticity under the assumption that the elasticity is globally
constant. This means that we need to solve the following equation for
q pp q:
p
qppq
BBp qppq (6.12)
This is a differential equation, and its solution is a family of functions

given by
q p p q C p (6.13)
in which C ¡ 0 is an arbitrary coefficient. This coefficient is essen-

tially a parameter of the demand function that can be chosen to fit it to
known data points, namely, pairs of observed price and corresponding
demand. We can calculate the willingness to pay that corresponds to
the constant-elasticity demand by substituting 6.13 into 6.8:
wppq
B 1
Bp qppq qp0q p
1
(6.14)
The demand curve qppq and willingness to pay wppq with constant-
elasticity demand are depicted in Figure 6.5. Similarly to the linear
demand function, constant-elasticity demand can be a reasonable ap-
proximation for relatively small price changes. The constant-elasticity
demand correctly captures a smooth decrease in the willingness to pay
as the price grows, but it also implies that the willingness to pay – re-
call that this is the maximum acceptable price – is concentrated near
zero, which is not necessarily a realistic assumption.
Figure 6.5: The constant-elasticity demand function q pp q p and corre-

sponding willingness to pay w pp q p 1 for different values
of .
6.4.3 Logit Demand Curve
The logit demand function attempts to overcome the limitations of the

linear and constant-elasticity models by accounting for the fact that the
price elasticity reaches its maximum near the point of reference. More
specifically, we can expect that demand for a product is persistently
low if its price is set much higher than competitive prices and minor
changes in the price have very little stimulus, that is, the local demand
elasticity is low. In the same vein, prices that are set much below the
market are likely to drive persistently high demand that will be rela-
tively insensitive to small price changes – customers will perceive the
deal as a bargain anyway. The highest price sensitivity is likely to be
in a zone around the competitive prices, where small changes in price
can significantly impact the demand. These considerations, as well as
empirical support, suggest a sigmoid shape for the demand curve, as
depicted in Figure 6.6. The sigmoid demand curve, also referred to as
the logit demand function, can be specified as follows:
qppq Qmax
1
(6.15)
1 ea bp
in which Qmax is the maximum achievable demand, and b is a param-

eter that controls the steepness of the demand curve. For any a and
b, maximum price sensitivity will be achieved at a price numerically
equal to pa{bq, so parameter a can be used to shift the point of ref-
erence if b is given. The parameters Qmax , a, and b can be estimated
from data to fit the logistic curve to observed data points.
The willingness to pay function for logit demand can be straightfor-
wardly obtained by differentiating the demand:
B ea bp
wppq
B p q pp q q p0 q b p1 ea q
1
2 (6.16)
1 ea bp
As shown in Figure 6.6, the logistic willingness to pay is a bell-

shaped curve that is relatively similar to a normal distribution.
Logit demand is closely related to the multinomial logit (MNL)
model that we considered in Section 2.6.1.1. Recall that if a given
customer n chooses a product or service among several alternatives
p1, . . . , Jq and the utility from choosing option i is measured as Vni ,
then the MNL model states that the probability of choosing option i is
as follows:
Vni
Pni °Je (6.17)

j 1e
Vnj
Figure 6.6: Logit demand function for several values of the parameter b and the
corresponding willingness to pay. The line thickness is proportional
to the magnitude of b.
Utility Vni is typically measured by using a regression model that

accounts for the different properties of customers and products. How-
ever, if we model it as just a linear function of price Vni bj pj , in
which bj are regression coefficients, then the probability Pni does not
depend on individual consumers anymore and becomes equal to the
average probability of product selection, which is the market share of
the product µi :
ebi pi
µi ppq ° b p (6.18)
je
j j
Equation 6.18 is essentially a new demand model that is even more

flexible than our logit demand one because it allows individual com-
petitive prices to be explicitly accounted for. However, if we choose to
treat competitive prices as a single parameter that can be defined as
¸
c ebj pj (6.19)

j i
then we can express the market share of a given product i as follows

(note that we use the identity c1 e ln c ):
ebi pi bi pi 1
µ i pp i q ° c e eb p cc1

j ie
bj pj ebi pi i i
(6.20)

e ln c bi pi
1 e ln cbi pi
which is the same as the logit demand model from equation 6.15.
6.5 basic price structures 399
6.5 basic price structures
A demand curve describes the relationship between the price and de-
manded quantity. It allows us to express a firm’s revenues and profits
as a function of price and to solve the optimization problem to deter-
mine the profit-optimal price level. Although this approach might look
like a precise way to calculate optimal prices, it can rarely produce
acceptable results because it is difficult, perhaps impossible, to esti-
mate a globally accurate demand curve that takes into account all of
the consequences of a price change, including competitors’ responses
and other strategic moves. The formal optimization problem, however,
can provide useful insights and support decision making, which is an
important step towards a programmatic solution. The analysis of de-
mand curves also helps to justify different price structures and their
key properties, which is necessary for the more advanced and auto-
mated optimizations that we will consider in later sections.
6.5.1 Unit Price
The first price structure we consider is the pricing of individual items

or units, such as one book, one shirt, or one pound of oranges. First,
let us write down the standard profit equation of the firm by using the
demand function qppq:
G q p p q pp V q (6.21)
in which G is profit, p is price, and V is variable costs. We omit the

fixed costs here for the sake of brevity. We then recall that the linear
demand curve is defined by
p
qppq Qmax 1 (6.22)
P
in which P is the maximum willingness to pay and, consequently, the
maximum acceptable price. The price can be optimized by taking a
derivative of the profit with respect to the price and setting it to zero:
B G B q p p q pp V q q pp q 0
Bp Bp (6.23)
Solving this equation for p, we obtain the optimal price, which is the
average of P and V:
popt P 2V (6.24)
We can substitute the optimal price into equation 6.22 to determine the
number of units that the firm is expected to sell at this price:
qopt Q2P
max
pP V q (6.25)
Finally, the profit at this price will be
Gopt Q4P
max
pP V q2 (6.26)
A geometrical interpretation of the equations above is shown in Fig-

ure 6.7. Note that the profit is numerically equal to the area of the
rectangle bounded by popt and V.
Figure 6.7: Unit price optimization for a linear demand curve.
A similar result can be obtained for the constant-elasticity demand.

According to the definition of elasticity we get
Bq ppq qppq (6.27)

p p
Substituting this into equation 6.23, we find the optimal price:
popt V 1 (6.28)
Equation 6.28 is very convenient for demonstrating some weaknesses

of strategic price optimization with the basic demand curves. Let us
consider the example of a firm that manufactures an item at a cost of
$10 and wants to determine the optimal sale price by using the esti-
mated elasticity. If the elasticity estimated from the transactional data
is 1.5, the optimal price is $30. However, it is challenging to estimate
the elasticity with high accuracy, and it is likely that the estimate is ac-
tually more like 1.5 0.4. This leads us to a range of “optimal” prices
from $21 to $110, as shown in table 6.2. It is clear that this result has
limited practical applicability.
Elasticity 1.10 1.20 1.30 1.40 1.50 1.60 1.70 1.80 1.90
Optimal $110 $60 $43 $35 $30 $27 $24 $23 $21
Price
Table 6.2: Optimal prices calculated for different values of elasticity and variable
costs of $10 with the cost-elasticity demand model.
6.5.2 Market Segmentation
Virtually all markets demonstrate heterogeneity of demand caused by

the fact that different customers, and even the same customers at dif-
ferent moments of time, value products differently. This diversity of
valuations has numerous reasons. Consumer markets, for example, are
fundamentally heterogeneous because the number of human needs,
such as meals or clothing, is relatively limited but incomes vary tremen-
dously, which results in very different amounts of money spent on the
same needs by different individuals. Customers can use the same or
similar products in different ways and derive different values from
product features, have more or less information about competitive of-
ferings, and so on. We have already seen how this heterogeneity creates
a fertile field for targeted promotions and advertisements, and we can
now explore how it influences price decisions. Fortunately, the analysis
of unit price optimization provides a very convenient basis for this.
We can see in Figure 6.7 that the maximum achievable revenue is
numerically equal to the total area under the demand curve, so the
maximum achievable profit can be estimated as
Gmax 12 P Qmax (6.29)
At the same time, any single price popt , no matter how optimal it is
or isn’t, represents a trade-off because some customers will not buy a
product if they consider it to be too expensive, although they would be
willing to buy it at a lower price, in between popt and V, and would
thereby positively contribute to the profit. Moreover, some customers
will tolerate prices higher than popt , although the sales volume that
they will generate is relatively small. In both cases, a firm fails to cap-
ture additional profits that lie in the triangle in between the demand
curve and the variable costs line. Price segmentation is a natural way
to overcome the limitations of a single regular price by segmenting
customers according to their willingness to pay and offering different
prices to different segments. A particular case of this strategy where
the regular price has been complemented by a higher premium price
and a lower discounted price is shown in Figure 6.8. Note how the
profit area increases relative to that in the single-price strategy.
Figure 6.8: Profit optimization by using price segmentation.
This consideration leads to the challenging question of how we can

sell the same or similar products to different customers at different
prices. Broadly speaking, this requires the setting of fences between
customers with different willingness to pay in such a way that cus-
tomers with higher willingness will not be able to a pay the lower price
intended for the lower segments. The fencing mechanisms are multi-
tudinous and vary across industries, although most variations can be
reduced to a few basic principles. Let us consider a few examples of
price-fencing techniques from the retail industry that demonstrate the
remarkable ingenuity in this regard:
store zone Stores in a retail chain are typically located in different
neighborhoods with different demographic and competitive fac-
tors, such as average household income, average family size, dis-
tance to the nearest competitive store, etc. This naturally sepa-
rates customers by levels of price sensitivity and ability or will-
ingness to look for an alternative supplier. It enables a retailer to
set store-level prices that differ in different zones.
package size Fast-moving consumer goods (FMCG), such as soft
drinks or toiletries, have high turnover rates and consumers natu-
rally have a choice between buying small amounts of product fre-

quently and stockpiling larger amounts. This trade-off is also im-
pacted by demographic factors, such as household size. This cre-
ates fences by willingness to buy large or small packages and en-
ables the setting of different per-unit margins for different pack-
age sizes. Quantity-based and purchase-frequency-based offers
and discounts also relate to this category.
sale events Customers can be differentiated by their willingness to

wait for a lower price versus their willingness to buy immediately
at the regular price. This type of segmentation is widely used in
the apparel domain, where seasonal sales are one of the main
marketing mechanisms.
coupons Many customers might not be willing to buy a given prod-

uct at the regular price but might consider buying it at a dis-
counted price. Hence, a retailer can benefit from a discount be-
cause it generates additional customers, although the margins
are lower relative to the purchasing habits of regular customers.
On the other hand, it might be harmful to offer a discount to an
excessively wide audience because it will be used even by those
customers who would be willing to pay the regular price (in the
absence of the discount). The response modeling techniques dis-
cussed in Chapter 3 help to solve this problem. However, there is
a traditional solution that has been in use since the 19th century
– couponing. A coupon represents a price discount that requires
some effort to earn or redeem (e. g., the customer has to find it
in a newspaper, cut it out, and present at a store), which fences
customers by willingness to spend time and effort in getting a
discount.
sale channels Sale channels naturally represent fences because

customers select channels by criteria that strongly correlate
with their willingness to pay. For instance, the price sensitivity
of liquor store shoppers is consistently lower than that of
customers who buy the same wine in grocery stores [Cuellar
and Brunamonti, 2013].
departments Retailers and manufacturers often differentiate

markups in accordance with differences in price sensitivity
among genders and ages. For example, women’s clothing is
generally more expensive than men’s.
club cards Membership helps to differentiate occasional shoppers

from high-spending loyal customers for whom private deals out-
weigh membership fees.
branding Retailers and manufacturers create separate brands that

are targeted at higher or lower consumer segments relative to
the main brand. The brands can be positioned as inferior to
sell products more cheaply without cannibalizing sales from the
main brand or as superior to capture additional revenue from
consumers with lower price sensitivity.
This list of price fences can be extended with more techniques from
other industries, such as airline fare classes or credit card deals. The
optimization strategy for price segments can be demonstrated with the
setting of prices for n segments in a way that maximizes profits. We
can start with the following equation that straightforwardly follows
from Figure 6.8:
ņ
G pqi qi1 q ppi V q (6.30)

i 1
in which pi and qi are the prices and quantity sold for a segment i,
respectively, and q0 0. The quantity sold at price pi is

qi Qmax 1 pPi (6.31)
We can find prices that maximize the profit by taking partial deriva-
tives of G and equating them to zero. By inserting equation 6.31 into
equation 6.30, setting p0 S and pn 1 V, and doing algebraic sim-
plifications, we find
BG Qmax
Bpi P ppi1 2pi pi 1 q, 1¤i¤n (6.32)
Equating these partial derivatives to zero, we find a recurrence rela-

tionship for segment prices:
pi pi1 2 pi 1
(6.33)
We can easily check that this relationship, as well as initial conditions

p0 S and pn 1 V, are satisfied by the following segment prices:

opt
pi n 1 1 pn 1 iq P iV (6.34)
Consequently, the optimal prices should be equally distributed in

between the variable costs V and maximum acceptable price P. This is
illustrated by an example in Figure 6.9.
With these prices, the profit earned can be obtained by inserting
equation 6.34 into equation 6.30:
Qmax n2ppnP V1qqP

2
Gopt (6.35)
Figure 6.9: Example of optimal prices for four segments with variable costs of
$10 and a maximum acceptable price of $25.
This equation is a generalization of the profit equation for unit

prices 6.26. We can see that the profit increases proportionally to
n{pn 1q as the number of segments grows and approaches the
maximum achievable profit.
Price segmentation is arguably the most powerful and widely used

pricing technique. Its power, however, completely depends on the abil-
ity to fence segments. Strict segmentation can be achieved in relatively
few cases. For example, a theme park can price tickets differently for
different age groups by verifying a proof of age. Most segmentations,
however, are imperfect and create an opportunity for customers from
a higher segment to buy a product marketed for a lower segment at a
lower price. For example, online customers are known to have higher
price sensitivity than store customers as a result of the ease of price
comparisons and other factors, so retailers quite often lower online
prices. This creates so-called showroom behavior where customers look
at objects in a store and then buy them later online. This market can-
nibalization can be quite damaging, as illustrated in the example in ta-
ble 6.3. The addition of a low-price segment increases the total profits
in the case of perfect segmentation, but the leakage of 500 customers
to this segment from the higher one completely erases the gains.
Total
Price Demand Revenue Profit
Profit
Single segment $19 5000 $95,000 $45,000 $45,000
Perfect A $19 5000 $95,000 $45,000

$48,000
segments B $13 1000 $13,000 $3,000
Imperfect A $19 4500 $85,500 $40,500

$45,000
segments B $13 1500 $19,500 $4,500
Table 6.3: Example of segmentation with and without demand cannibalization.

The variable costs V are assumed to be $10 per item and Profit

Revenue V Demand.
6.5.3 Multipart Pricing
We have seen that segmentation is a powerful method for capturing ad-

ditional revenues from heterogeneous demand. One popular method
of creating fences between customer segments is to leverage the differ-
ences in product usage patterns and intensity. For example, manufac-
turers of photographic cameras typically offer a wide range of prod-
ucts, including entry-level models, advanced enthusiast cameras, and
professional cameras. Although these categories are substantially dif-
ferent in image quality and other functional properties, manufacturers
also try to capture profits from the more intensive usage of profes-
sional equipment by offering better durability and performance. In this
regard, the most accurate segmentation can be achieved by charging a
camera holder based on the number of shots they will take to capture
revenues in direct proportion to usage. Although this approach might
not be feasible for cameras because of implementation and competition
reasons, variations of this are successfully applied in other industries.
The two structurally similar price strategies that exploit this idea are
two-part tariffs and tying arrangements:
two-part tariffs A two-part tariff is a price structure with two

components – an entrance fee and a metered price. The entrance fee
is charged for access to a product or service. The metered price
is a per-unit charge that depends on the consumed quantity. A
classic example of two-part tariffs is telecommunication services
that charge a connection fee in addition to metered fees for every
minute or gigabyte used. Other notable examples include utili-
ties, such as electricity, natural gas, or water, enterprise software,
which is often priced in proportion to the number of users, or-
ders, or concurrent connections on top of the baseline price, and

amusement parks, which can charge an entrance fee and a price
per ride.
tying arrangements Some products are very closely related to

each other, in the sense that a customer cannot extract much
value from one product without another. This enables a man-
ufacturer to create tying arrangements that prevent a customer
from switching between brands and to rebalance the prices of
the related products. Examples of tying arrangements can often
be found when a durable product (the tying product) is com-
plemented with a consumable part (the tied product), such as
a razor handle and razor blades or a printer and ink cartridges.
The revenue contributed by the consumable parts can dominate
the total consumer lifetime value, so the durable product can be
priced lower or even below its marginal costs.
Let us now consider quantitative models that provide directions on

how the entrance price pe and metered price pm of a two-part tariff can
be optimized. We have previously mentioned that the demand curve
can be interpreted both as an aggregate demand of the market deter-
mined by the maximum willingness to pay and as the consumption
level of a single customer at a given price. Two-part tariffs explicitly
depend on the level of consumption and, consequently, require us to
account for both the level of consumption and the demand heterogene-
ity, so we have to use a more complicated model than the one we used
for unit pricing and price segmentation [Smith, 2012; Oi, 1971].
Let us first consider the case of a single consumer, for whom the
demand curve is shown in Figure 6.10.
Recall that the equation of the linear demand curve is
pm
q Qmax 1 (6.36)
P
The consumer evaluates the product or service at price P, so the
surplus gained by buying a unit at metered price pm is numerically
equal to the area of the triangle under the demand curve bounded by
the pm line. Consequently, we can assume that the optimal entrance fee
is equal to this surplus because a lower fee will leave available profits
on the table and a higher fee will drive the consumer out of the market.
This means that the entrance fee can be expressed as an area under the
demand curve:
opt
pe q2 pP pm q Q2P
max
pP pm q2 (6.37)
Figure 6.10: Price optimization for a two-part tariff.
The total profit earned for a customer is a sum of the entrance fee
and metered charges:
G pe q pp m V q
opt
(6.38)
The optimal metered price can be obtained by taking the derivative

of the profit and equating it to zero:
BG Bpopt Bq
B p m B p m B p m pp m V q q0
e (6.39)
Inserting equations 6.37 and 6.38 into equation 6.39 and solving it
for pm , we find that the optimal metered price should be set equal to
the marginal costs:
opt
pm V (6.40)
This means that two-part tariff pricing encourages the seller to set
the entrance fee as high as possible and lower the metered price to the
minimum, so the profit will be extracted exclusively from the entrance
fee. This strategy, for example, is widely adopted by amusement parks
that tend to charge high entrance fees rather than charge per ride.
The approach with the high entrance fee, however, faces a strong
headwind in the case of competition or heterogeneous demand where
customers are willing to purchase different quantities of a product at
a given price. This situation is depicted in Figure 6.11, where multiple
demand curves have the same slope but differ in quantity purchased.
Let us assume that each demand curve corresponds to a certain cus-
tomer segment. We now cannot set the entrance fee higher than the sur-
plus that corresponds to the lowest demand curve because we would
Figure 6.11: Two-part tariff in the case of heterogeneous consumption levels.
lose customers otherwise. Let us denote the ratio between the demand
of the i-th segment to the demand of the segment that corresponds to
the lowest demand curve as ki , so the equations for the demand curves
can be written as

qi Qmax ki 1
pm
ki P
, ki ¥1 (6.41)
The profit can be expressed by summing the entrance fee with the
metered charges from each of the segments:
¸
G pe µ i q i pp m V q (6.42)
i
in which µi is the share of the segment i, that is, if the total number
of customers is N then the segment i contains N µi customers. The
optimal metered price can be found by taking the derivative of the
profit with respect to the metered price and equating it to zero:
B G B pe pp V q ¸ µ B q i ¸
0
B pm B pm m i
i
B pm i
µi q i (6.43)
Equation 6.37 for pe still holds, under the assumption that P is the
maximum acceptable price for the lowest demand curve, so we can
insert it into the equation above and, by using the fact that the sum
of all µi equals one, we derive a simple expression for the optimal
metered price:
¸
V µ i pk i 1 q V PpE rks 1q
opt
pm P (6.44)
i
in which E rks is the weighted average of the segment demand mul-

tipliers. This result demonstrates that heterogeneity in demand levels
drives the metered price up and, consequently, decreases the entrance
fee, which is already limited by the lowest demand curve. This can
reverse the price balance of the original two-part tariff – the metered
element can become dominant in the total profit relative to the entrance
element. This shift can also be driven by competition, which limits the
seller’s ability to extract consumer surpluses though the entrance fee.
6.5.4 Bundling
Bundling is often defined in economic texts as the sale of two or more

separate products in one package at a single price. This definition is not
very precise because virtually any product can be considered as a bun-
dle of its parts, just as a car is a bundle of an engine, wheels, and other
components that can be sold separately, at least on industrial markets.
From the price optimization standpoint, we are most interested in price
bundling that offers two or more products at a discounted price relative
to the sum of the individual product prices; this discount is the only
advantage to the alternative of buying unbundled items.
Price bundling is a popular price structure that can be found in many
industries. Examples of price bundles include sport and opera season
tickets, restaurant meals that bundle an appetizer, entree, and dessert,
luggage sets with several bags of different size, kitchenware, and soft-
ware suites. A seller typically has a choice between three options: to
sell unbundled products, to sell a group of products only as a bundle,
which is known as pure bundling, and to offer the individual products
as well as the discounted bundle, which is called mixed bundling. It
is intuitively clear that the discount provided by a bundle should be
counterbalanced by capturing more profits than the separate products
do. We have seen earlier that complex price structures often accom-
plish this by exploiting heterogeneity in the willingness to pay. Conse-
quently, we can make the assumption that bundling can leverage the
difference in the willingness to pay for different products.
example 6.2
As an example, let us consider an office software suite that includes a

spreadsheet application and slide presentation application. The most
basic scenario that we can analyze is a single-segment market in which
the value of each application is the same for all customers. Let’s say
that the spreadsheet application is valued at $100 and the presenta-
tion application is valued at $150 per user license. We can price the
individual products according to this valuation and the bundling of

the two products will not be beneficial because any bundle price other
than $250 will harm profits. The scenario with two segments shown in
table 6.4 is more interesting because the willingness to pay differs be-
tween the segments – sales departments highly value the presentation
application, whereas accounting departments prefer the spreadsheet.
The maximum price levels for the individual products that keep both
segments in the market are $100 for the spreadsheet application and
$100 for the presentation application. If the two segments are equal in
size, these prices are optimal and earn a total profit of $400. The price
of a bundle that includes the two products, however, can be set at $250
because both sales and accounting segments value the pair of products
at $250. The total profit earned from the two segments will be $500,
which is better than the $400 earned with the unbundled pricing.
Customer Segment Willingness To Pay
Spreadsheet Presentation
Sales $100 $150
Accounting $150 $100
Table 6.4: Example of price bundling with two products and two consumer seg-
ments.
The example above exploits the asymmetry of willingness to pay

between two segments. If all customers consistently value one product
higher than another, bundling is not able to capture more profits than
unbundled pricing. We can use this observation to build a quantitative
model for bundle price optimization. The advantage of this model is
that it makes relatively few assumptions about how customer segments
are distributed and can take advantage of numerical optimization or
simulations to optimize pricing for an arbitrary number of segments or
even individual users. Let us consider a scenario where we are selling
two products, X and Y, and there are several customer segments in
which the maximum willingness to pay Pix for product X is equal or
proportional to the willingness to pay Piy for product Y, as illustrated
in Figure 6.12. For any product prices px and py , a consumer segment
falls into one of four areas (buy nothing, buy only product X, buy
only product Y, or buy two products) depending on the relationship
between the willingness to pay and the corresponding price. In our
scenario with positively correlated valuations, customers buy either all

or nothing, so bundling cannot add value.
Figure 6.12: Customer segments in which the willingness to pay for two prod-
ucts is positively correlated. The segments are depicted as black
dots.
The case with asymmetrical willingness to pay is totally different.

For the sake of simplicity, let us assume that all segments have the same
aggregate willingness to pay, but segments can allocate it to products
differently, as shown in Figure 6.13. This was exactly the case in the
example with the office software suite that we just studied.
pure bundling The price of bundle pB can be set equal to the ag-
gregate willingness to pay beause it is constant for all segments:
pB Pix Piy , constant for all i (6.45)
The profit of selling the bundle will be

¸
GB ni pB Vx Vy
i
¸ (6.46)
ni Pix Piy Vx Vy
i
in which ni is the number of customers in segment i, and Vx and

Vy are variable costs for products X and Y, respectively. On the
other hand, we find that selling unbundled products earns the

following profit:
¸ ¸
GU ni ppx Vx q ni py Vy
(6.47)
¥
i: Pxi px ¥
i: Pyi py
The first and second terms of equation 6.47 are the total profits
for product X and Y, respectively. Comparing equations 6.46 and
6.47, we find that GB is greater than or equal to GU for any
product prices px and py , so pure bundling is a more effective
strategy in this scenario than selling unbundled products.
Figure 6.13: Customer segments with a negatively correlated willingness to pay.
mixed bundling We can try to enhance pure bundling by selling

products both separately and in a bundle. This requires us to
set prices on separately sold products in a way that does not
cannibalize the profits we gain from bundling. We can do this by
comparing the bundle profit with the profit of a separate product.
Let’s take product X as an example:
px Vx ¡ pB Vx Vy (6.48)
Consequently, we find the prices for individual products to be
px ¡ pB Vy
py ¡ pB Vx
(6.49)
This condition corresponds to the small triangles located at the

extreme ends of the bundle price line in Figure 6.13. Customer
segments that fall into these areas will choose to buy either prod-
uct X (rightmost triangle) or Y (leftmost triangle) instead of a
bundle, which will generate a higher revenue than that with pure
bundling.
The approach above provides relatively high flexibility for the opti-
mization of bundle prices. With the assumption that the willingness to
pay is estimated for sample segments or individual consumers, the cor-
responding points can be pinned on a plane and the optimal prices can
be searched by using numerical optimization methods and tessellating
the plane into areas with different profits depending on the segment
locations and sizes.
6.6 demand prediction
The basic demand models we considered earlier are convenient for

strategic analysis but often too coarse for the actual price optimization.
The challenge is that the demand on a given product depends on many
factors, including a product’s own properties such as price or brand,
the prices of competing products in the category, sales events, and even
the weather. We need to build a more advanced demand model that
incorporates these factors and allows one to perform what-if analysis
to forecast the response to price changes, assortment extensions and
reductions, and shelf-space re-allocation. This model is an important
building block that can be used in many applications that depend on
quantitative demand estimation, including the following:
static price optimization Baseline prices and markdowns can
be customized for different customer segments defined by chan-
nels, locations, and propensity models. This requires segment
properties to be incorporated into the demand model.
dynamic price optimization Price markdowns, sales event plan-

ning, and the pricing of products with constrained supplies re-
quire prices to be optimized as a function of time. This implies
that the demand model should also take temporal changes into
account.
category management Shelf-space optimization and assortment

optimization require an understanding of the dependencies be-
tween the demands on different products.
stock-level optimization Supply chain management and re-

plenishment benefit from demand modeling. Demand prediction
becomes especially important in the case of major sales events
and flash-sales business models.
6.6 demand prediction 415
Demand prediction can be considered as a relatively straightforward

data mining problem that boils down to the building of a regression
model and evaluation of it over historical data. However, the design
of the regression model is not so straightforward because the demand
is influenced by many factors with complex dependencies. We might
need to combine multiple basic demand functions, such as linear de-
mand or logit demand, to assemble a model that is flexible enough to
properly capture seasonal changes, consumer choices, price elasticity,
and other factors.
The design of demand models is definitely an art because differ-
ent optimization problems require different demand prediction mod-
els, and it is hardly possible to build a universal demand model that
incorporates the wide variety of factors that influence demand, such as
the following:
• Model usage. A model may or may not include controls for con-
sumer choice, demand changes over time, competitive prices, and
so on. The choice of controls depends on the application that the
model is created for.
• Available data. The availability, trustworthiness, and complete-

ness of data influence the design of a model and its capabilities.
• Business domain, model, and process. The demand model re-

flects the terminology, constraints, and structure of a particular
business. For example, the demand model can predict the de-
mand rate for individual products or groups that include multi-
ple product variants of different sizes, colors, or flavors.
• Experimentation. Demand models, just like most real-life predic-

tive models, explicitly and implicitly incorporate a lot of domain
knowledge and require a lot of experimentation and tuning.
We can, however, learn many useful techniques and design ideas by
studying industrial demand prediction models. The structure of mod-
els and choice of features provides us with reusable building blocks
and hints that can be leveraged for future demand prediction prob-
lems. We choose to study two real-life demand models from the retail
domain and also to examine the difference between these two exam-
ples and a few models reported by other companies. All of these mod-
els were created in the context of price and assortment optimization,
so they are well aligned with the optimization methods that we will
review later in this chapter.
6.6.1 Demand Model for Assortment Optimization
The first demand model we will consider was designed for assortment
optimization at Albert Heijn, a supermarket chain in the Netherlands
[Kök and Fisher, 2007]. It places a strong emphasis on consumer de-
cisions to enable a fine-grained analysis of the factors that influence
consumer choice.
A supermarket chain carries a large number of products that are
divided into merchandising categories, such as cheese, wine, cookies,
and milk. Each category is further divided into subcategories such that
products within a subcategory are similar and are often good substi-
tutes for one another but the difference between subcategories is sub-
stantial. For example, a fluid milk category can include subcategories
such as whole milk, fat-free milk, flavored milk, and so on. Supermar-
kets typically achieve very high service levels for the products that they
carry and stockouts are quite rare, so the model we will consider does
not take stockouts into account. At the same time, this demand model
was designed for assortment analysis and optimization, so it explic-
itly accounts for consumer choice, which makes it well suited to the
assortment-related problem that we will consider in later sections.
The demand for a single product can be broken down into three
separate decisions that apply for every consumer visiting a store:
• First, a consumer purchases or does not purchase from a subcat-
egory. Let us denote the probability that a consumer purchases
any product from a subcategory during the visit to the store as
Prppurchase | visitq.
• Second, a consumer chooses which product to buy within a

subcategory. The probability that a consumer chooses product
j among other alternatives when a purchase takes place is
Prpj | purchaseq.
• Finally, a consumer decides how many units to buy. We can cap-

ture this choice by using the mathematical expectation of the
quantity (number of units) purchased by the consumer given that
product j has been chosen and purchased. Let us denote it as
E rQ | j, purchases.
The demand for product j can then be expressed by using the choice
probabilities and expected purchase quantity as follows:
Dj N Prppurchase | visitq
Prpj | purchaseq (6.50)
E rQ | j, purchases
in which N is the number of consumers visiting the store within a

given time frame (e. g., during the day). All factors in equation 6.50
can be estimated from the transactional data from the stores. The de-
mand generally depends on the date (day of the week, holidays, etc.)
and store (size, neighborhood demographics, etc.), so we introduce sub-
scripts t and h to denote date and store, respectively, and estimate the
demand as a function of these parameters. Alternatively, store proper-
ties, such as size, location, and average consumer’s income, can be in-
corporated into the model as predictive variables. The number of store
visitors can be modeled by using a log-linear regression as follows:
7̧ ¸E
N
logpNht q α1 α2 Tt α3 Wt α3 1 Bti α10 i Eti (6.51)

i 1
i 1
in which Tt is the weather temperature, Wt is the weather comfort in-

dex (humidity, cloudiness, etc.), Bti and Eti are 0/1 dummy variables
for a day of the week and public holidays, respectively, NE is the total
number of public holidays, and α represent regression coefficients.
The purchase incidence is a binary choice variable (purchase/no pur-
chase), so we can use a standard modeling approach – express the pur-
chase probability as a sigmoid function that approximates the binary
decision and estimate its exponential parameter from the data. The
sigmoid function can be specified as
Prppurchase | visitq
1
e x
(6.52)
1
which is equivalent to

Prppurchase | visitq
x log
1 Prppurchase | visitq
(6.53)
Exponential parameter x is estimated for a given date t and store h

by using a regression model with the following structure:
xht β1 β 2 Tt β3 Wt β4 Aht
7̧ ¸E
N (6.54)
β4 i Bti β11 i Eti

i 1
i 1
in which Aht is the share of products that are currently on promo-

tion in the subcategory, that is, the ratio between the number of prod-
ucts that are promoted in a given store at a given date to the total num-
ber of products in the subcategory. As we will need to build further
product-level models, let us express the share of promoted products
with indicator variables Ajth for individual products that equal one if
the product is promoted and zero otherwise:
J̧
Aht 1J Ajht (6.55)

j 1
in which J is the total number of products in the subcategory.

Estimation of the probability of the purchase of a given product in
a subcategory is a little bit more tricky. As we have seen earlier, con-
sumer choice can be modeled by using the multinomial logit model
(MNL), so we express the probability of the purchase of a product
among alternatives as follows:
exppyj q
Prpj | purchaseq °
i exppyi q
(6.56)
in which i iterates over all products in a subcategory and yj is a pa-

rameter variable. Similarly to the probability of the purchase incidence,
we can build a regression model for parameter variable yj for a given
store and date:
yjht γj γN 1 pRjht Rht q γN 2 pAjht Aht q (6.57)
in which the regression coefficients γN 1 and γN 2 are shared for
all products, Rjht and Rht are the product price and the average price
in the subcategory, respectively, and Ajht and Aht are the promotion
dummy variables and average promotion rate, as described above for
the purchase incidence regression model.
Finally, the average number of units sold can be modeled as follows:
E rQ | j, purchases λj λN 1 Ajht λN 2 Wt
¸
N H (6.58)
λN 2 i Eti

i 1
in which λ are regression coefficients and the other variables have
been defined and explained above. By substituting the individual re-
gression models above into the root equation 6.50, we obtain a fully
specified demand prediction model. This model can be adjusted to the
retailer’s business usage cases by adding more explanatory variables,
such as marketing events.
Competing products and their attributes play an important role in
demand modeling even if the assortment is not the main concern. For
example, the online fashion retailer Rue La La reported that the relative
price of competing styles and the number of competing styles are in the
top three most important features in their demand prediction model
[Ferreira et al., 2016].
6.6.2 Demand Model for Seasonal Sales
The second demand model we will review was developed for Zara,
a Spanish fashion retailer and the main brand of Inditex, the world’s
largest fashion group [Caro and Gallien, 2012]. The model is geared
towards sales events optimization and places strong emphasis on the
temporal dimension of the demand.
Seasonal clearance sales are an integral part of the business strategy
for many apparel retailers. A regular selling season, which is typically
biannual (fall–winter and spring–summer), is followed by a relatively
short clearance sale period that aims to sell off the remaining inventory
and free up space for the new collection for the next season. Some
retailers exercise even smaller sales cycles to overrun competitors and
get more revenues from customers by offering more diverse and fluid
assortment. Price optimization in such an environment requires the
creation of a demand model that properly accounts for seasonal effects
and stockouts caused by the exhaustion of inventory and deliberate
assortment changes.
We describe the demand model in two steps, in accordaance with
the original report [Caro and Gallien, 2012]. The first step is to pre-
pare the available demand data for regression analysis by removing
seasonal variations and accounting for demand censoring due to stock-
outs. Next, the regression model itself is specified.
6.6.2.1 Demand Data Preparation

Most clothing items come in multiple colors and sizes, so every stock
keeping unit (SKU) can be referenced by using a product number r and
size–color variant v. With the assumption that the historical sales and
inventory data are available on a store level for individual days, let us
denote sales of SKU pr, vq at store h on day d as Spr, v, d, hq and the cor-
responding inventory level at the beginning of the day as Lpr, v, d, hq.
We also define a function Fpr, v, d, hq equal to one if a given SKU was
on display at store h on day d and zero otherwise. The on-display in-
formation can be available in the data explicitly or one can estimate it
from the sales and inventory data by checking if the inventory level or
quantity sold is zero for a given item.
First, let us introduce a seasonality factor that incorporates intra-
week and inter-week demand variations. We define the following ag-
gregates of the sales data:
• SW pdq is the total sales volume for a week in which day d falls.
This volume is aggregated for all products, size–color variants,
and stores.
• SW is the average total weekly sales volume calculated from the

historical data.
• SW prq is the average weekly sales volume for product r calcu-

lated from the historical data.
• SW pr, v, hq is the average weekly sales volume for SKU pr, vq at

store h calculated from the historical data.
• SD pweekdaypdqq is the average sales volume for a given day of

the week defined for seven days from Monday to Sunday.
The seasonality factor can then be defined as follows:
δpdq
SW pdq
SD°pweekday pdqq (6.59)
i 1 D q
p
SW 7
S i
The first and second terms account for inter-week and intra-week
demand variations, respectively. Next, we introduce the following fac-
tor that accounts for both seasonality and on-display information to
normalize the demand for product r and week w:
kpr, wq
¸ S p
W r, v, h q ¸
δpdq Fpr, v, d, hq
SW prq
(6.60)
h,v d in w
The fraction in the equation above corresponds to the share of sales

at store h relative to all stores, so the contribution of the on-display
status variable for the store is properly weighted by the store’s sales
share. Finally, we define the normalized demand for product r and
week w as follows:
¸
qpr, wq Spr, v, d, hq
1
kpr, wq (6.61)
v,h,d in w
6.6.2.2 Model Specification

Our next step is to build a regression model that predicts the normal-
ized demand rate qpr, wq. This was achieved by Zara by using a rela-
tively small log-linear model with the following specification:
logpqpr, wqq α0,r α1 logpQr q α2 Ar,w
α3 logpqpr, w 1qq
" *
Lpr, wq
1 (6.62)
α4,w log min 1,
T

pr,w
α5,w log
pr,0
in which α are regression coefficients and the features are defined as
follows:
• α1 : Qr is the quantity of product r purchased by a retailer. Al-

though this value is not directly related to the demand, it is im-
plicitly related to fashion and style because a retailer often pur-
chases large quantities of basic and popular items, whereas niche
items are purchased in smaller quantities.
• α2 : Ar,w is the number of days since product r was introduced

at the stores. The demand for fashion products often depends on
their novelty and tends to decrease over time.
• α3 : qpr, w 1q is the demand rate for the previous time inter-

val. This variable helps to capture demand correlation between
adjacent time intervals.
• α4,w : The broken assortment effect refers to the fact that the de-
mand for a given product can decrease as the inventory level
gets low. In the fashion retail context, this can often be attributed
to unpopular sizes and colors that remain after the most popular
ones have sold out. This effect can be accounted for by introduc-
ing a threshold T for Lpr, wq, which is the aggregated inventory
level for product r across all variants and stores.
• α5,w : The discount depth is defined as the ratio between the cur-
rent price pr,w and the regular price pr,0 . This term is effectively
a price sensitivity factor.
We can see that the model is heavily focused on the demand variabil-
ity over time because it was created for the optimization of seasonal
sales events. Demand models reported by other fashion retailers can
include more features such as brand, color and size popularity, rela-
tive prices of competing styles, and different statistics about past sales
events that can shift price sensitivity [Ferreira et al., 2016].
6.6.3 Demand Prediction with Stockouts
The demand models described in the previous sections are merely re-
gression models that are trained to forecast the sales numbers. In prac-
tice, the observed sales numbers do not necessarily match the actual de-
mand because of stockout events. If this is the case, the observed sales
volume will be lower than the actual demand, that is, the sales volume
that could potentially be achieved given an unlimited supply without
stockouts. The problem of stockouts can be especially important for
business models with seasonal sales or flash sales, where stockouts
are very frequent and a demand prediction model created based on
the observed sales volume is likely to be biased and inappropriate for
optimization of the inventory levels or prices. Therefore, a demand pre-

diction method that explicitly accounts for out-of-stock events and lost
sales needs to be developed. This problem has been studied intensively,
and a number of methods exist that address demand prediction with
stockout events [Anupindi et al., 1998; Musalem et al., 2010; Vulcano
et al., 2012]. In this section, we discuss a heuristic technique developed
by Rue La La, an online fashion retailer, to account for stockouts that oc-
cur during flash sales events, that is, extremely time-limited discounts
[Ferreira et al., 2016]. The advantage of this method is the simplicity
and ability to work with low inventory levels (if a retailer stocks only a
few instances of each SKU) that can be insufficient to fit more advanced
models.
On the assumption of a setting where a retailer sells multiple prod-
ucts, let us introduce the following notation:
• di is the actual demand for product i
• ci is the inventory level for product i
• qi is the actual quantity sold, which can be expressed as follows
qi min tci , di u (6.63)
If a product can be represented by multiple size–color variants, we

treat each variant as a separate product and measure the above values
for individual variants. Next, we assume that the retailer runs limited-
time sale events for individual products. At the beginning of the event,
the stock level is equal to ci . If the product sells out before the end of
the event, we can observe that qi ci but we do not observe the true
demand di . If the product does not sell out, we can assume that we
observe the true demand, qi di . The main problem we are trying to
solve is how to estimate the expected sales volume qi given inventory
level ci as a parameter. We can distinguish the following cases:
• If product i has already been on sale and did not sell out, the
p ,
observed quantity sold qi can be used as the demand estimate d i
and the expected quantity that will be sold given the inventory
level cnew
i in a new sales event can be predicted based on this
estimate. This can be summarized as
! )
p
d i qi Ñ pi
q min cnew p
i , di (6.64)
• If the product i has already been on sale and did sell out, we
have not observed the true demand. This requires us to perform
demand unconstraining, that is, to estimate the true demand based
on the quantity sold and the historical data for other products.
This unconstrained demand estimate d p can then be used to pre-

i
dict the expected quantity that will be sold in a new event. This
strategy can be summarized as follows:
! )
−−−−−−−−−Ñ Ñ min
unconstraining p p
qi d i pi
q cnew
i , di (6.65)
• If the product is new and has never been on sale, the demand
needs to be predicted with a regression model that uses product
and event properties as features and the unconstrained demand
value as the response variable. This problem can be solved by
using the methods discussed in the previous section. This case
can be summarized as
! )
p
d i fpproduct, eventq Ñ pi
q min cnew p
i , di (6.66)
in which f is the predictive demand model. In the case of multiple

product variants, the model can be created to predict the demand
at product level and then the demands on the product variants
can be derived from this based on the observed distribution of
demand for different sizes and colors.
In the approach described above, the key task is demand uncon-
straining. One possible approach is to use historical data for products
that did not sell out to estimate the demand for products that did sell
out. First, we can create demand curves for products that did not sell
out, with each curve describing the percentage of total sales for a given
product as a function of time into the event measured in hours or days,
as shown in Figure 6.14.
In practice, the number of products and corresponding demand
curves can be very high, so we can split the curves into several
classes by using standard clustering methods, determine the rules that
differentiate the classes (clusters) from one another, and determine
the typical demand curve for each class. For example, it can be the
case that the key factor that determines the demand curve shape is
the event start time, such as morning, early afternoon, late afternoon,
and so on. Most events that start, for example, in the morning may
have similar demand curves and thus belong to the same cluster. The
clustering rule can be more complex and involve multiple properties.
Once we have identified the rules that assign products and events to
classes and have also determined a typical curve for each class, we
can unconstrain the demand for products that sold out based on the
curve. First, we use the clustering rules to assign the product and
event to a class. Second, we use the demand curve for this class to
map the stockout time to the proportion of total sales k, as illustrated
Figure 6.14: Example of demand curves and the determination of the demand-
unconstraining proportion [Ferreira et al., 2016]. In this example,
there are three classes of events, each of which is represented by a
typical demand curve, and the event that needs to be unconstrained
falls into the first class.
in Figure 6.14. Once this proportion is determined, the true demand

for product i can be estimated by dividing the observed quantity sold
by the proportion value:
p
d i qki cki (6.67)
This estimate can then be used for sales volume prediction, demand
modeling, and stock level optimization.
6.7 price optimization
A demand model enables us to search for the optimal price by analyz-

ing how the price changes increase or decrease the profits. We have
already seen that this optimization is not particularly difficult for ba-
sic price structures and an oversimplified environment that does not
account for the special properties of supply, demand, and operations
that can be found in real applications. In practice, however, we face
numerous constraints and interdependencies that require much more
elaborate and specialized optimization models to be developed.
Most of the constraints fall into one of the following three categories:
supply constraints, demand constraints, and structural constraints.
Supply constraints can be imposed by limited resource capacity,
such as the fixed number of seats in an opera house, constrained or
6.7 price optimization 425
expensive replenishment, and product perishability, which can take

different forms from the limited shelf life of grocery products to the
seasonal collections at a clothing store to airline tickets that must be
sold before the departure of a plane. Demand constraints are often
related to imperfect consumer segmentation, interdependency of the
demands on substitutable products, demand changes over time, and
demand uncertainty, in the sense that the demand cannot be predicted
accurately. Structural constraints are related to operational and legal
conditions that might require suboptimal but practicable solutions to
be selected.
We will continue this section with price optimization models for dif-
ferent market segmentation strategies and constraints. This group of
methods can be considered as static price optimization, although one
can repeat the optimization procedures regularly to adjust prices over
time or define time-based customer segments, such as weekend and
workday cinema tickets. We will then demonstrate that the dynamic
pricing that explicitly optimizes temporal price changes is closely re-
lated to market segmentation, and we will describe dynamic price man-
agement methods.
6.7.1 Price Differentiation
The goal of price differentiation, more commonly referred to in eco-

nomic texts as price discrimination, is to find the optimal prices for sep-
arate customer segments or individual customers. Price optimization
at the segment or customer level requires the creation of a demand
model that takes customer or segment properties as parameters or the
creation of a separate demand model for each segment. This can be
achieved with the demand prediction methods that we have previously
discussed. The basic price optimization problem can then be specified
as follows:
¸
max
p
pps vs q qs pps q (6.68)
s
in which s is a segment, ps is the price for segment s, p is a vec-

tor of prices for all segments, qs is the demand function for segment
s, and vs is variable costs, which can be constant or vary depending
on the segment. This optimization problem is separable by segments,
so the basic unit price optimization can be applied for each segment
separately.
It is often the case that the number and structure of segments is
limited by operational constraints, so a programmatic system might
need to evaluate the impact of combining several segments together
into a group and assigning that group a single price. This can be done
by rewriting equation 6.68 for N segment groups Si as
Ņ ¸
max pi pp i v s q q s pp i q (6.69)
p

i 1 P
s Si
and solving a separate optimization problem for each group to find

N optimal prices.
example 6.3
Let us illustrate the optimization models above with an example from

the retail domain. Consider a retailer that operates multiple stores
and sells a product that comes in several sizes, for example, analgesic
tablets that are packaged in bottles of 25 or 50 tablets [Khan and Jain,
2005]. The retailer can offer quantity discounts based on package sizes
and set prices separately for each store. Regression analysis of the
transactional data has shown that the demand for analgesics is well
described by the following model:
qpp, s, hq 2000 1400p 8s 10s h (6.70)
in which p is the price per tablet, s is the package size (the num-
ber of tablets in the bottle), and h is the average household size fac-
tor, which is positive when the average household size in the store
area is relatively large and negative when the average household is
small. The demand is negatively correlated with price, as we expect.
It is also negatively correlated with the package size, which indicates
that consumers prefer smaller packages to larger ones. The last term
is positively correlated with the package size for large households and
negatively correlated for small ones, so the demand for large packages
is higher in areas with large households, which is also intuitive.
We will optimize prices for a setting with two stores with different
values of the household size factor h and two package sizes with dif-
ferent wholesale prices v, as shown in Figure 6.15.
The first scenario we consider is a fine-grained price differentiation
that jointly optimizes quantity discounts based on package sizes and
store-level prices. The goal is to find four different prices pij , in which
i corresponds to one of two package sizes and j corresponds to one of
two stores. The optimization problem can then be stated as follows:
¸ ¸
max si pij vi qppij , si , hj q (6.71)
p

i 1,2 j 1,2
This optimization problem is separable and quadratic with respect to
prices because the demand function is linear. Solving the problem for
Figure 6.15: Parameters of the price optimization problem for two stores and
two package sizes.
the values from Figure 6.15, we obtain the results presented in table 6.5.
We can see that the solution justifies quantity discounts and suggests
that per-tablet prices are lower for a large package in both stores. It
also exploits the higher demand for large packages in the area with
large households by raising the price in the corresponding store.
Package Size Household Price per Demand

(tablets) Size Tablet (bottles)
25 Small $0.67 607

25 Large $0.81 794
50 Small $0.49 410
50 Large $0.76 785
Total profit $45,863
Table 6.5: The optimal prices for a scenario with four segments.
Our second scenario assumes that it is not possible to vary prices

at a store level because of operational constraints, so we can only set
two prices, for the small and large packages. We change the optimiza-
tion problem accordingly by following the generic approach from equa-
tion 6.69:
¸ ¸
max ps i p i v i q q p p i , si , h j q (6.72)
p

i 1,2
j 1,2
This problem is also separable and quadratic with respect to prices.

Solving it, we find that the total profit decreases in comparison with
that of the first scenario, as one can see in table 6.6 where the optimiza-
tion results are shown.
Package Size
Price per tablet Demand (bottles)
(tablets)
25 $0.74 1401
50 $0.63 1195
Table 6.6: The optimal prices for a scenario with two segments.
This analysis allows us to evaluate different segmentation strate-

gies and find the optimal solution under structural constraints.
N
6.7.1.1 Differentiation with Demand Shifting
One of the biggest challenges of price differentiation is imperfect
fences between price segments that allow customers to move from
one segment to another depending on the price differences. Although
it sounds as though this demand shifting is necessarily harmful for
a seller, it can actually impact profits both positively and negatively.
On the one hand, customers with high willingness to pay can find
a way to buy a product at a relatively low price, thereby decreasing
the profit of the high-margin segments. On the other hand, the
corresponding increase in the segment that the customer migrated to
can counterbalance the loss.
The shifting effect is especially important in cases of constrained sup-
ply because it can help to achieve a more even demand distribution and
reduce stockouts. For example, the supply of seats in an opera house
is fixed but the demand can vary significantly, typically reaching its
maximum on the weekends and its minimum on weekdays. The opera
house can lose potential revenue if the peak demand on the weekends
exceeds the capacity. We can expect that setting variable ticket prices
for different days of the week can improve profits because the higher
demand on the weekends enables the opera house to charge higher
prices and make better margins, as we have seen in the previous sec-
tion. High weekends prices, however, can make some customers buy
cheaper weekday tickets, which shifts the demand to days with avail-
able seat capacity and improves the revenues even further.
It can be quite challenging to create a demand shifting model. The
challenge is partly related to the design and training of a demand
model that simultaneously accounts for the prices of all related seg-
ments. It is often impossible to measure cross-price elasticities (how the
demand in one segment is impacted by the price in another segment)
for all possible pairs of segments, so one has to use grosser approxima-
tions, such as the ratio between the price in a given segment and the
average prices in other segments. Another challenge of demand shift-
ing models is that cross-segment dependencies make the optimization
problem inseparable and sharply increase its computational complex-
ity. Assuming that we draw prices from a discrete set of size m and the
number of segments is n, we might need to evaluate as many as mn
price combinations if the demand model does not follow any particular
functional form that exhibits properties such as linearity or convexity.
One possible approach to price optimization with demand shifting
is to make the assumption that the demand shift is proportional to the
price difference between the segments. More specifically, if the price
for segment i is higher than the price for segment j, then the demand
for segment i decreases by Kppi pj q and the demand for segment
j increases by the same amount. Parameter K determines the amount
of demand transferred between the two segments for every dollar of
price difference. The basic optimization problem can then be rewritten
as follows to adjust the demand in each segment by the total of the
demands shifted from other segments:

¸ ¸
max
p
ppi vi q qi ppi q K pj pi (6.73)
i j
in which i and j iterate over all segments. Note that this demand
shifting model does not change the total demand, in the sense that the
sum of all shifts is zero. However, this does not mean that the total
quantity sold remains constant for any K because the demand shift
causes the optimal prices to change, which, in turn, changes the values
of the demand functions.
example 6.4
We can illustrate the impact of demand shifting by continuing our ex-

ample with analgesic tablets. Inserting the demand shifting terms into

equation 6.71, which describes the scenario with four price segments,
we get
¸ ¸
max si pij vi qppij , si , hj q ∆ppij q (6.74)
p

i 1,2 j 1,2
in which the demand shift is the sum of the pairwise price differ-
ences with other segments, or
¸ ¸
∆ppij q K pxy pij (6.75)

x 1,2 y 1,2
We solve this optimization problem and obtain the prices for the four
segments presented in table 6.7. From a comparison with tables 6.5 and
6.7, we find that the demand shifting has increased price sensitivity, so
the demand for small packages has decreased but the demand for a
large package has increased because of the relatively low price per
tablet. This change in demand can be considered positive for a retailer
because it increases the total profit.
Package Size Household Price per Demand

(tablets) Size tablet (bottles)
25 Small $0.75 420

25 Large $0.81 608
50 Small $0.56 535
50 Large $0.69 910
Table 6.7: The optimal prices for a scenario with four segments and demand

shift. The shifting parameter K 400.
N
6.7.1.2 Differentiation with Constrained Supply
The price optimization models that we have considered so far were fo-
cused on setting prices that achieve the highest possible profits allowed
by a given demand curve. This view on price optimization assumes
perfect replenishment, such that a seller is always able to deliver the
quantity of a product demanded at the profit-optimal price. This as-
sumption is reasonably fair for some industries, such as supermarket
retail, where it is possible to build a supply chain that almost perfectly
replenishes the inventory and stockouts are rare. However, as we have
already discussed, it does not hold in many other industries that face
different supply constraints. In this section, we will discuss a relatively
simple case, in which each market segment has a fixed capacity of a
product and we need to find the optimal global price or segment-level
prices.
Let us first consider how a product can be priced for a single mar-
keting segment if the available quantity is fixed. This problem is a stan-
dard unit-price optimization with the addition of a quantity constraint:
max x pp V q
p,x
subject to x ¤ q pp q
(6.76)
x¤C
p¥0
in which C is the available quantity (capacity), so that a stockout

occurs if the demand exceeds C. As usual, we denote the price, de-
mand, and variable costs as p, qppq, and V, respectively. Variable x
corresponds to the actual quantity sold.
This problem is quite trivial because the demand is a monotonically
decreasing function of price. First, we can find the optimal price, re-
gardless of the constraint and calculate the demand that corresponds to
this unconstrained optimal price. Next, we can compare this demand to
the available quantity and set the price based on the maximum of these
two values. If the demand that corresponds to the unconstrained opti-
mal price is below the available quantity, then this price is the solution
because the stock level is less restrictive than the demand. Otherwise,
we take the price that corresponds to the maximum available quantity.
This price, referred to as the stockout price, will be higher than the un-
constrained optimal price to slow down sales and avoid stockouts. The
latter case is illustrated in Figure 6.16.
Figure 6.16: Unit price optimization with constrained supply.

The principle above can be combined with the price differentiation

and demand shifting considered in the previous sections. However, it
is only applicable if the supply constraints are set to each segment
separately. Global capacity constraints when all price segments share
the same inventory require more advanced optimization methods that
we will describe later in this chapter.
example 6.5
As an example, let us take an opera house that optimizes its ticket

prices. We assume that the opera gives a performance daily, the capac-
ity C of the auditorium is 1200 seats, and the demand varies over the
week in accordance with the following formulas:

1800 50p, Monday

1350 50p, Tuesday
1200 50p,

Wednesday

qpp, tq
1350 50p, Thursday (6.77)

1800 50p, Friday
2250 50p,

Saturday
3600 50p, Sunday
in which p is the ticket price and t is the day of the week. We also
assume that the variable costs per seat are negligible because the build-
ing maintenance costs, cost of the performance, and other expenses are
pretty much constant.
The opera house might decide to set a flat price for all days, which
would lead to the following constrained optimization problem:
¸
max p xt
p,x
t
subject to xt ¤ qpp, tq (6.78)
xt ¤ C
p¥0
in which t iterates through the seven days of the week. This prob-
lem instance is not particularly difficult because we can assume that
the price belongs to a relatively small discrete set, so we can evaluate
each candidate solution. We find that the optimal price is $19.80, which
corresponds to a revenue of $98,010.
The result above can be contrasted with variable pricing, for which
each day is considered as a separate segment and optimized accord-
ingly. We can rewrite the optimization problem for seven different

prices:
¸
max pt xt
p,x
t
subject to xt ¤ q pp t , t q (6.79)
xt ¤ C
pt ¥ 0
This problem is separable by segments, so we can optimize each

day in accordance with equation 6.76. The optimal prices, as well as
the number of sold seats xt , are shown in Figure 6.17, which clearly
demonstrates the increased seat utilization. Price differentiation turns
out to be extremely efficient in this example and increases the total
revenue up to $147,825.
Figure 6.17: Example of ticket price optimization for an opera house. The verti-
cal bars represent the number of seats sold, and the points are the
ticket prices for the corresponding days.
6.7.2 Dynamic Pricing
Dynamic pricing is a group of pricing strategies and methods that op-

timize profits by changing prices over time. As we will see shortly,
these methods are structurally similar to other price differentiation

techniques, but they offer additional advantages for the algorithmic
approach. In particular, dynamic pricing is more focused on incremen-
tal continually adjusted pricing decisions that make the problem more
tractable than global price optimization. In addition, dynamic pricing
typically deals with strict capacity and time constraints that introduce
additional complexity, which, in turn, makes automated optimization
appealing.
Although it is intuitive that dynamic pricing can be beneficial by
providing an additional degree of freedom in comparison with static
prices, we need to explore this concept more deeply to understand why
one would benefit from changing prices over time.
First, we note that dynamic pricing can be used as a market seg-
mentation technique. Consider a seller who initially offers a product
at some baseline price. The product will be purchased by customers
whose willingness to pay is above the baseline price and will not be
purchased by other customers. Assuming that all customers have made
their purchasing decision with regard to the baseline price, the seller
can then lower the price. This will generate additional revenue from
purchasers whose willingness to pay is in between the baseline and dis-
counted prices. Thus, dynamic pricing is able to create price segments
by exploiting only variance in willingness to pay and time without ad-
ditional segment fences. We will use this concept in the next section
to build a quantitative model that provides useful properties for the
optimal price trajectories.
The heterogeneity of willingness to pay can be a sufficient condition
for dynamic pricing, but in a significant number of business cases, the
demand exhibits even higher variability. Dynamic pricing then plays
the role of an equalizer that adjusts prices to changing demand condi-
tions:
variable demand Demand for a product or service can change over
time, often following seasonal patterns. Industries that face vari-
able demand include apparel retailers, entertainment businesses,
and hotels to name a few.
variable inventory value Demand changes can often be con-

nected to objective or subjective changes in the inventory value.
Fashion products, electronic devices, and automobiles drop their
value as newer models enter the market. Perishable grocery
products lose their value as the expiration date approaches,
whereas airline tickets tend to be more valuable for last-minute
buyers.
demand uncertainty Dynamic price changes can help to find the

optimal price by trial and error when a seller is uncertain about
the demand [Pashigian, 1987]. For example, an apparel retailer
who buys the inventory upfront for the next season might not be
able to predict how popular a new fashion garment will be. It can,
however, try different markdowns to find a price that maximizes
profits and matches the inventory constraints.
From the optimization standpoint, dynamic pricing does not neces-
sarily require specialized optimization methods. Assuming that time is
discrete, so that every time interval can be considered as a segment and
there are no constraints that make the intervals interdependent, we can
optimize intervals separately by similar methods to the price differen-
tiation cases (recall the opera house example). Dependencies between
the time intervals might require specialized optimization models to be
created, and this is often the case in dynamic pricing. These depen-
dencies are typically related to the supply constraints because all time
intervals are typically served from the same inventory resource pool.
We already stated the paradigm that dynamic pricing can be consid-
ered as an equalizer of the demand. We can extend this view and say
that dynamic pricing equalizes both demand and supply. The main two
attributes of a constrained supply are fixed capacity and perishability:
fixed capacity A global capacity constraint means that a seller sells
off a fixed stock of a product that cannot be replenished. A good
example is apparel retail, when a retailer buys fixed quantities of
inventory upfront. It can also be the case that replenishment is
possible but limited in some way.
perishability Perishability means that the stock of a product must

be sold within a limited time frame. The unsold inventory loses
its value or can be sold at a relatively small salvage price. Exam-
ples of perishable products include service resources, such as ho-
tel rooms and flight tickets, seasonal collections of apparel, and
consumer packaged goods.
The presence of these two constraints in a business is typically a good
indicator for dynamic pricing feasibility. These two properties are
equally important because we are trying to match the demand rate
with the supply rate, which is essentially a ratio of the capacity to the
selling time period defined from the perishability.
6.7.2.1 Markdowns and Clearance Sales

Dynamic price optimization with a fixed capacity of perishable inven-
tory can be expressed as the following mathematical problem:
Ţ
max pt xt
p,x

t 1
Ţ
subject to xt ¤C (6.80)

t 1
xt ¤ q pp t , t q, for t 1, . . . , T
pt ¥ 0, for t 1, . . . , T
In the formulation above, we assume that capacity C should be sold

out within the time frame that consists of T discrete time intervals. Our
objective is to maximize the revenue by setting optimal prices for each
of T time intervals. We also assume that the unsold inventory has zero
value after point T and variable costs are negligible, although both the
salvage inventory price and variable costs can be included into the
equation in a relatively straightforward way.
Problem 6.80 models a number of important business cases. In the
retail space, markdowns and seasonal clearance sales match this model
because the sale period is typically fixed as is the stock to be sold. A
wide range of service providers, including airlines, railways, hotels,
theaters, stadiums, and freight companies, face conceptually similar
optimization problems in selling a fixed capacity of seats, rooms, or
freight spaces within a fixed time frame determined by the vehicle
schedule, check-in date, or event time. The service industries, however,
often face many additional constraints and use different methods of
revenue optimization that are based on resource allocation, so we will
focus on markdowns and clearance sales in this section. Resource allo-
cation methods will be considered later in this chapter.
The first conclusion we can make from equation 6.80 is that prices
can vary over time only if the demand rate varies. If the demand rate is
constant, all time intervals are identical, so we can apply the standard
unit price optimization and then select the maximum of this uncon-
strained optimal price and the stockout price determined by the capac-
ity constraint C, in accordance with the logic we described in section
6.7.1.2.
Variability of the demand rate over time can be attributed to differ-
ent factors, such as seasonality or changes of the product value. In a
general case, the demand changes over time can take different shapes,
including increasing and decreasing trends, as does the price. We can,
however, show that markdown prices tend to follow a certain pattern
by considering a finite population of customers [Talluri and Van Ryzin,

2004]. Let us assume that one sells a durable product to a finite number
of customers, such that a customer does not buy a product more than
once during the sale period. If a seller sets a certain price pt for time
interval t, it means that all customers with willingness to pay higher
than or equal to pt will buy the product and become inactive until the
end of the sale period. The seller, however, can decrease the price to
attract customers with a lower willingness to pay. This indicates that
markdown prices should be set close to the customer valuation at the
beginning of sales and then monotonically decrease.
To build a quantitative model, let us also assume that customers’
willingness to pay is uniformly distributed between 0 and some maxi-
mum price P:
$
wppq unifp0, Pq
{
&1 P, 0¤p¤P
(6.81)
%
0, otherwise
Recall that uniform willingness to pay implies a linear demand
curve, which, in this context, can be interpreted as the number of
customers who buy a product at a given price and become inactive
until the end of the sale. Consequently, we can visualize the markdown
process as sliding down the demand curve, as shown in Figure 6.18.
Note that the optimization of markdown prices in this interpretation is
almost identical to the market segmentation problem that we studied
earlier, so the equations below are structurally similar to the equations
for market segmentation but have different meanings.
We find that the quantity sold during period t is
pt
Qmax 1 1 t1
p
Qt
P P
(6.82)
Qmax
P
p p t 1 p t q
Consequently, the total sales revenue is
Ţ
G pt
Qmax
p p t 1 p t q (6.83)

t 1
P
We take partial derivatives of the revenue to find the revenue-

maximizing prices:
BG Qmax
Bpt P ppt1 2pt pt 1 q (6.84)
Equating these derivatives to zero, we find that
pt pt1 2 pt 1
(6.85)
Figure 6.18: Markdown optimization model for uniformly distributed willing-

ness to pay.
The initial conditions for this recurrence equation should be set to

meet the capacity constraints, so that the markdown prices are dis-
tributed in the range between the maximum price P and the stockout
price PS :
p0 P
1 PS
(6.86)
pT
The markdown prices that meet relationship 6.85 and conditions 6.86
are given by

opt
pt PS pP PS q 1 T t
1
(6.87)
This result suggests that markdown prices should be evenly dis-

tributed between the list price and stockout price. Although this is an
interesting conceptual insight, the finite population model is too gross
for applicable pricing decisions, so we need to return to basic optimiza-
tion problem 6.80 and solve it directly for arbitrary demand functions.
This approach provides more flexibility because demand prediction
models are able to incorporate various patterns observed in the past,
including the effects related to the finite population.
6.7.2.2 Markdown Price Optimization

Optimization problem 6.80 plays a very important role in algorithmic
pricing, so we will spend this section discussing how it can be solved
efficiently and working through a numerical example.
Let us assume that the set of allowed prices is discrete, which is the
case in most practical applications where prices are integer values of
cents or dollars. If the set of allowed prices has K price levels and the
number of markdown rounds is T , then we can find the optimal prices
for each round by evaluating KT possible price combinations. This ap-
proach is computationally feasible for business cases with small K and
T values. For example, it is common for many discount retailers to
set prices that end in 4.90 or 9.90, such as $34.90 or $59.90, so the
total number of price points is relatively low. However, the problem
becomes intractable even if we only have twenty price levels and opti-
mize daily prices for a horizon of two weeks, which would require us
to evaluate 2014 price combinations.
One possible solution is to approximate the original non-linear
optimization problem with a linear programming relaxation [Talluri
and Van Ryzin, 2004]. Let us denote the set of allowed price levels as
tP1 , . . . , PK u and introduce weight variables z that control the price
levels selected for each time interval, such that
Ķ
pt zit Pi (6.88)

i 1
The optimization problem can then be restated as finding K T

revenue-maximizing variables z, such that, for any time interval t, one
zit is equal to one and the other z variables are equal to zero. Let us
relax the zero-or-one constraint for variables z and just require them to
sum to one for each time interval:
Ķ
zit 1, for t 1, . . . , T
i 1 (6.89)
zit ¥0
This effectively means that we allow fractional prices, that is, two
or more prices from the discrete set can have non-zero weights in the
same time interval. We assume this to be acceptable, which means that
a markdown round can be broken down into smaller intervals. For ex-
ample, if we have two price levels with non-zero weights 0.2 and 0.8,
then the first price level should be set for one fifth of the round du-
ration and the second price level should be set for the remaining four
fifths of the round. This relaxation allows us to rewrite the dynamic

price optimization problem as follows:
Ţ Ķ
max zit Pi q pPi , tq
z

t 1i 1
Ţ Ķ
subject to zit q pPi , tq ¤ C

t 1i 1 (6.90)
Ķ
zit 1, for t 1, . . . , T

i 1
zit ¥0
Problem 6.90 is a linear program: if we arrange all variables z in a
flat vector with K T elements and calculate the vectors of revenue
and demand values for the corresponding pairs of time t and price
level i, then the objective function and capacity constraint can be ex-
pressed as vector products. This formulation allows us to use standard
optimization software for linear programming.
example 6.6
Let us now consider an example of a retailer who plans a four-week

sale campaign for a single product. The set of allowed prices includes
five levels: $89, $79, $69, $59, and $49. Demand functions are estimated
for each week of the sale as follows:

1800 10p, week 1

qpp, tq 15p,
1300 week 2
(6.91)

15p,
1200 week 3
1100 18p, week 4
We substitute these parameters into problem 6.90 and solve it for dif-
ferent values of product capacity C. These solutions are presented in
the small tables below. Each solution is a price schedule with 20 values
of z: each column corresponds to one of the four weeks, and each row
corresponds to one of the five price levels, with the topmost row cor-
responding to the highest price and the lowermost row corresponding
to the lowest price. For example, the optimal price is $89 for the first,
third, and fourth weeks when the capacity is 700 items. For the same
capacity, the price is fractional for the second week, which means a mix
with 22% at $89 and 78% at $79.
C = 700, G = $61,400 C = 1000, G = $81,507

1.00 0.22 1.00 1.00 1.00 0.00 0.00 1.00
0.00 0.78 0.00 0.00 0.00 0.27 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.73 1.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
C = 1300, G = $96,778 C = 1600, G = $109,218

1.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00 1.00 0.60 0.00 1.00 0.72 0.00
0.00 0.00 0.00 0.40 0.00 0.00 0.28 1.00
C = 1600, G = $116,198 C = 2200, G = $117,242

0.58 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.42 0.00 0.00 0.00 1.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 1.00 1.00 1.00 0.00 1.00 1.00 1.00
We can see that the prices generally decrease over time, which re-
flects the decreasing trend in the demand rates. Another tendency
is that tight capacity constraints generally diminish and delay mark-
downs, which is also expected.
N
6.7.2.3 Price Optimization for Competing Products
One of the main challenges of price optimization is the dependen-
cies between the products. In many business domains, especially in
retail, customers constantly make a choice between competing or sub-
stitutable products by using the price of one product as a reference
point for another product, so the demand for a given product is usually
a function of both the product price and the prices of the competing
products. If this is the case, the prices cannot be optimized for each
product in isolation, and the prices for all competing products need to
be optimized jointly instead. The number of competing products can
be as high as several hundreds in some applications, so the optimiza-
tion problem can become computationally intractable. In this section,
we dive into this problem and discuss the framework developed by
Rue La La, an online fashion retailer, which can significantly reduce
the computational effort [Ferreira et al., 2016].
Let us assume that a retailer is selling n distinct products. Let us

denote the set of products as N, so |N| n, and the set of possible
price levels as P, so |P| k. In practice, the set of prices is often rela-
tively small because it is a common practice to use prices with certain
endings, such as $9.95 and $14.95. We also assume that the retailer
has a demand prediction model that estimates the demand for a given
product as a function of the product price and the prices of competing
products. In the case of an unlimited inventory, the model can predict
the true (unconstrained) demand. If the retailer has a fixed quantity
of product in stock, the model can predict the expected quantity sold,
which is the minimum of the true demand and the available quantity,
as we discussed earlier in section 6.6.3.
The optimization goal is to assign a price to each product in a way
that the total revenue is maximized. This problem statement is appli-
cable both in the case of static and dynamic pricing. In the case of
dynamic pricing, for example, multiple concurrent sales events, this
optimization problem needs to be solved repeatedly. For example, a re-
tailer can start n sales events for all n products at the same time with a
certain quantity of each product in stock. The initial prices are assigned
by solving the optimization problem for the initial settings. The next
day, the prices are updated by solving the problem once more with the
latest stock levels, and so forth. A naive solution for the problem is to
try all kn possible price assignments and evaluate the demands and
revenue for each assignment. This problem can be formally defined as
¸
max pi qi ppq
p
i PN (6.92)
subject to pi P P, for i 1, . . . , n
in which p is an n-dimensional vector of product prices, pi is the price

of product i, and qi is the demand model for product i that uses all
prices, including the product price pi and the prices of all competing
products, as inputs. Both the number of products n and the number of
possible prices k can be relatively large, so this approach is inappropri-
ate in many practical applications. One possible way to work around
this problem is to predict the demand not as a function of all individual
product prices but as a function of some aggregate that has a smaller
number of possible states than vector p. For example, one can use the
sum of the prices for the competing products as an aggregate:
ņ
Q pi (6.93)

i 1
It is easy to see that the total number m of possible values of Q is
m npk 1q 1 (6.94)
For example, if the number of products is 10 and the set of pos-

sible prices includes 10 levels P t$1, . . . , $10u, the value of Q will
be in the range from $10 to $100 with steps of $1. The assumption
that the demand depends mainly on the sum of the prices and not on
the individual prices may or may not be true in practice, but there is
evidence that this works in at least some applications [Ferreira et al.,
2016]. This assumption essentially reduces the search space from kn
price combinations to Opnkq possible values of Q. The optimal price
assignment can then be found by solving m optimization problems for
each possible value of Q and selecting the best solution. To define the
optimization problem, let us denote the j-th element of the set of price
levels P as ppjq and introduce binary variables zij P t0, 1u, such that zij
equals one if product i is assigned price ppjq and zero otherwise. With
the assumption that the value of Q is given, the optimization problem
can then be defined as the following integer program:
¸ ¸
max ppjq qi pppjq , Qq zij
z
i PNjPP
¸
subject to zij 1
j PP (6.95)
¸ ¸
ppjq zij Q
i PNjPP
zij P t0, 1u
The first constraint ensures that each product has exactly one price,
and the second constraint guarantees that all product prices sum up to
Q. The demand function qi has to predict the demand for product i
as a function of the price assigned to the product and the sum of the
prices Q, so the sum of the prices is used as one of the features for
model training. The demand function, as we have already discussed,
can take into account the available inventory, so the predicted demand
level can be bounded by the product stock level. This can be especially
important if the model is used for sales events and markdown price
optimization when the goal is to sell off the stock.
Integer programming problem 6.95 has substantially lower computa-
tional complexity than the naive exhaustive search approach, but it can
still be challenging if the number of products and prices is high. If this
is the case, the linear programming relaxation of the problem (where
zij is not binary but is a continuous variable, such that 0 ¤ zij ¤ 1)
can be used to approximately evaluate all possible values of Q, and the
exact optimal solution can then be found by solving integer program-

ming problem 6.95 for only a subset of Q values [Ferreira et al., 2016].
This approach can greatly reduce the computational complexity and
make the problem tractable, even for a large number of products and
price levels.
6.7.3 Personalized Discounts
The price optimization methods that we discussed earlier leverage the

diversity of willingness to pay by setting different discounts for dif-
ferent customer segments or time intervals. Eventually, we would like
to combine these two approaches to manage both the monetary and
temporal properties of a discount at a segment level. Moreover, we can
attempt to improve the efficiency of price differentiation by replacing
segmentation with personalized discount levels. At this point, pricing
methods converge with the promotion methods discussed in Chapter 3:
pricing services can take advantage of targeting techniques to make
pricing decisions based on individual customer profiles, and promo-
tion services can optimize the monetary aspects of a promotion, such
as discount depth, by using price optimization methods. We spend
this section developing a method that optimizes the depth of the dis-
count and tries to find the optimal time and duration for the offer of
a discount to a given user [Johnson et al., 2013]. The idea of temporal
properties optimization comes from the assumption that a customer’s
probability to purchase is not uniform and varies over time, so there is
an optimal discount time window for each user.
In order to model the temporal properties of a discount, we will
decompose the probability of the purchase of brand k by customer u at
time t with discount value d into two multipliers: the brand purchase
probability and the probability to make a purchase at time t:
pktud ppbrand k | u; dq pptime t | u; dq (6.96)
Now we need to model the probability density functions ppbrand

k | u; dq and pptime t | u; dq separately. Nevertheless, we will use
a common approach for both of them. First, we define the form of
probability distribution and describe it with an utility function as a
parameter. Second, we build a regression model to estimate the utility
function from the data.
The probability density function of the purchase of a given brand
ppbrand k | u; dq is a typical case of a multiple choice model because
the consumer chooses a brand from several alternatives (let us denote
the total number of competing brands as K) that can substitute for one
another. Consequently, we can use the multinomial logit (MNL) model

to define the distribution:
exppxkut q
ppbrand k | u; dq °K
i1 exppxiut q
(6.97)
The utility function xkut can be learned from the data by building a
regression model such as the following:
W̧
xkut βuw Fkutw (6.98)

w 1
in which Fkutw are W explanatory variables that include discount

d and other features like loyalty and price, and βuw are W regression
coefficients.
The probability density function of a purchase at time t is modeled
in [Johnson et al., 2013] in the form of an Erlang distribution:
pptime t | u; dq y2u t exppyu tq (6.99)
in which the parameter variable yu can be estimated by means of a

regression model that, similarly to the model for the parameter vari-
able x in equation 6.97, includes the discount value as an explanatory
variable, so it can later be a subject of optimization.
The probability of purchase defined above enables us to model the
sales volume for a given customer Qu as a function of the discount
value in dollars d, discount start time t, and discount duration T :
t» T
Qu pd, t, T q pktud dt (6.100)

t
This leads us to the following optimization problem for the gross

margin:
¸
max m pQu p0, 0, tq Qu pd, t, t Tq Qu p0, t T , 8qq
d,t,T u (6.101)
d Qu pd, t, T q
in which m is the margin at the regular price. The first term in the
equation above corresponds to the revenue, which, in turn, consists of
three components – revenue received before the promotion, during the
promotion, and after the promotion – and the second term corresponds
to promotional costs. Figure 6.19 illustrates this breakdown.
The plot at the top shows the probability density of purchase by cus-
tomer u when the expected sales volume for a given product at the
regular price corresponds to the S0 area. A flat permanent discount
Figure 6.19: Optimization of the promotion time frame.
will lift this volume by adding area S1 , so the total revenue and pro-
motional costs (shown in the middle plot) will both be proportional
to S0 S1 . A time-optimized promotion will make the revenue pro-
portional to S0 S2 , and its costs will be proportional to S02 S2 (the
plot at the bottom). This difference between a flat promotion and an op-
timized promotion shows the potential to take advantage of temporal
optimization in the case of certain quantitative properties of probability
density functions.
6.8 resource allocation
In a nutshell, dynamic pricing provides a way to segment customers

by their willingness to pay and a way to optimize the prices for each
segment given that the total capacity is limited. As we discussed ear-
lier, one of the greatest advantages of this approach is that it does not
require predefined price segments and is able to create and tune them
dynamically. On the other hand, it requires the business model and op-
erational environment to be flexible enough in how prices are set and
updated. This flexibility varies across industries. For example, the re-
tail and eCommerce environments typically provide good capabilities
for dynamic pricing, whereas service industries, such as airlines, ho-
tels, and freight distribution, might be less flexible in that regard. This
difference can be partly attributed to historical reasons, namely the
practice of setting fixed fares for different classes of service. This has
resulted in the development of a large group of methods that are based
on an alternative interpretation of the problem. These methods first ap-
peared in the airline industry and historically preceded dynamic pric-
ing, as well as most programmatic methods in general.
Assuming that a seller has operational, legal, or business constraints
that limit the ability to vary prices arbitrarily, we can turn the dynamic
price optimization problem upside down and consider an alternative
approach. The idea is to define a set of fixed price segments, typically
referred to as fare classes, and allocate a fraction of the total capacity to
every class in a way that maximizes profits. Consequently, the subject
of optimization is the capacity limits allocated for each class. A classic
example of this problem is an airline that offers three fare classes (e. g.,
economy, business, and first class) and decides how many seats of each
class should be reserved, given that the total airplane capacity is fixed.
6.8.1 Environment
We discussed earlier that dynamic pricing is feasible for environments

that exhibit certain properties, such as variability of demand and fixed
resource capacity. These fundamental considerations are generally also
applicable to resource allocation because it is essentially a different so-
lution for the same problem. Many resource allocation methods, how-
ever, were developed primarily for service applications and address a
number of constraints that are specific for this domain. We confine our
discussion to the basic environment described below and only briefly
review additional challenges that exist in the theory and practice of
resource allocation.
• A seller offers a product or service to several market segments at
different fares. The segments can be defined based on the level
of service, such as economy or business classes in the airline in-
dustry, or can be based on more complex and fine-grained busi-
ness rules that aim to improve fencing between the classes. For
example, a hotel chain or airline might be willing to sell their
services to business customers at a higher price than to leisure
customers. To prevent business customers from buying the ser-
vice at the lower price, the service provider can set a condition
that low-price offerings must be booked a few weeks in advance
or not include overnight Saturday stays.
• All fare classes are served from the same fixed capacity of a re-
source, and the booking limits for classes can be changed dynam-
ically. For example, an airline can allocate different percentages
of standard economy and discounted economy seats for differ-
ent flights, although the total number of economy seats remains
fixed.
• An optimization system analyzes historical and ongoing demand

data to set or update booking limits for each fare class. The book-
ing limits are loaded into a reservation system, which is a trans-
actional application that receives booking requests. A booking
request is a request to reserve a unit of capacity for a specified
fare class, such as a request to reserve one discounted economy
seat. The reservation system either accepts the request if the cor-
responding booking limit is more than zero, subsequently decre-
menting the limit counter, or rejects the request if the capacity is
exhausted.
• The basic optimization models that we will consider make sev-

eral important assumptions about the demand. First, the demand
for each fare class is assumed to be a random variable with a
known distribution. Secondly, it is assumed that all demand vari-
ables are independent. In particular, the demand for a given class
does not depend on the availability or unavailability of other fare
classes. This is a very gross approximation because a customer
who has a request for a certain fare class rejected can consider
other fare classes and increase the corresponding demands, just
as in other cases with imperfect segmentation. Finally, it is as-
sumed that the requests arrive sequentially from the lowest class
(the cheapest one) to the highest class (the most expensive one).
This assumption is also a relatively gross approximation, but it is
widely accepted in practice and often matches the real demand
patterns – for example, fencing rules for low-fare leisure travel-
ers often include a condition that the booking should be done in
advance.
It should be noted that many applications of resource allocation re-
quire two major requirements that are not included in our basic en-
vironment model to be addressed. First, resources are often allocated
not as single units but as products that include multiple units. For ex-
ample, a hotel booking is a product that includes one or more room
nights, and an airline itinerary may be a chain of flight legs. This re-
quires a network of resources to be jointly managed and optimized.
Secondly, booking requests are eligible for cancellation in many indus-
tries, including airlines and hotels, and the share of canceled requests
can be significant. American Airlines has reported that about half of

reservations are canceled or become no-shows [Smith et al., 1992]. This
results in the practice of overbooking, when a service provider allows
resources to be booked beyond the capacity with the anticipation that
some of these reservations will be canceled in the future. Overbook-
ing also requires the development of specialized methods that adjust
booking levels according to the expected cancellation rates.
As we already mentioned, resource allocation considers booking lim-
its as the subject of optimization. A straightforward way to define a
booking limit is to assign a capacity for each fare class separately. The
major issue with this approach is that the capacity for a higher class
may be exhausted when lower classes remain available. Consequently,
a reservation system can reject a request for a higher class to preserve
the capacity it could use for future low-fare requests. This behavior
is obviously damaging from the profitability standpoint. A better ap-
proach, accepted as a standard in most theoretical models and practical
applications, is nested limits. The idea of nesting is to set limits not for
a given fare class individually but for all classes higher or equal to a
given one, as illustrated in Figure 6.20. We can see that these limits,
commonly called protection levels, are set in such a way that the first
limit is equal to the capacity reserved for the first class and the last
limit is equal to the total capacity.
Figure 6.20: Three fare classes and protection levels y1 , y2 , and y3 .
The reservation system accepts or rejects a request in accordance

with the following logic:
• A request for class i is accepted only if yi ¡ yi1 . This is intuitive
because the reservation limit for class i is the difference between
yi and yi1 .
• If the request is accepted, yn is decremented and all protection

levels greater than the new value of yn are set equal to yn .
This can visualized in Figure 6.20 as the right-hand boundary
of yn being moved to the left as more and more capacity is re-

served; the right-hand boundaries of the other protection levels
are bumped on the way.
The reservation process is illustrated in table 6.8. Initially, we have
2 units allocated for the first class, 2 units for the second class, and 4
units for the third class, which totals the capacity to 8. Any request is
accepted until round 5, when the total capacity decreases to 4 units.
The difference between y3 and y2 is zero, so class 3 closes. Note that
the process is irreversible – once the difference between yi and yi1
becomes zero, class i remains permanently closed. Requests for the
second class are accepted until round 7, and only requests for the first
class are accepted after that.
Protection Reservation
Units Sold Action
Levels Request
y1 y2 y3 C1 C2 C3
1 2 4 8 0 0 0 1 unit in C 2 Accept
2 2 4 7 0 1 0 1 unit in C 3 Accept
3 2 4 6 0 1 1 1 unit in C 3 Accept
4 2 4 5 0 1 2 1 unit in C 1 Accept
5 2 4 4 1 1 2 1 unit in C 3 Reject
6 2 4 4 1 1 2 1 unit in C 1 Accept
7 2 3 3 2 1 2 1 unit in C 2 Accept
8 2 2 2 2 2 2 1 unit in C 2 Reject
9 2 2 2 2 2 2 1 unit in C 3 Reject
10 2 2 2 2 2 2 1 unit in C 1 Accept
11 1 1 1 3 2 2 1 unit in C 1 Accept
12 0 0 0 4 2 2 — Reject
Table 6.8: Example of the reservation process with nested booking limits.
6.8.2 Allocation with Two Classes
The allocation problem is challenging, so we start with the most basic

scenario with two fare classes. Let us assume that we have a capacity
of C units and denote prices for the first and second classes as p1 and
p2 , respectively, so p1 ¡ p2 .
In accordance with our assumptions about the environment, the de-
mand for each of the two classes is a random variable Qi and its cumu-
lative distribution function Fi is known. We receive allocation requests
sequentially and, according to another assumption, the requests for
the second, less expensive class come first. Consequently, our goal is
to determine the optimal value of the protection level y, such that we
accept no more than C y requests for the second class and reserve
the remaining y units for the first class requests.
Any time we receive a request for the second class, we have the
choice of accepting it or rejecting it and switching the space to the
first class. This decision can be easily analyzed in terms of expected
outcomes, as illustrated in Figure 6.21. If we accept the request, we
earn revenue of p2 . If we reject the request, close the second class,
and switch to the first class, then two outcomes are possible. On the
one hand, if the demand for the first class will eventually exceed the
remaining capacity y, we will book this unit at price p1 . On the other
hand, if the demand for the first class is lower than the remaining
capacity, the unit would not be booked at all and we would earn zero
revenue. Consequently, the acceptance condition for the second class
can be written as follows:
p2 ¥ p1 PrpQ1 ¥ yq (6.102)
which is the same as
p2 ¥ p1 p1 F1 pyqq (6.103)
Inverting the cumulative distribution function, we find the optimal

protection level for the first class:

yopt F1 1 1
p2
p1
(6.104)
Equation 6.104 is known as Littlewood’s rule [Littlewood, 1972]. The

optimal protection level does not depend on the demand distribution
for the second class because we are looking for the larger quantity to be
reserved in the first class and assume that whatever remains is booked
in the second class, regardless of the distribution.
Figure 6.21: A decision tree for a two-class allocation problem.

example 6.7
Let us illustrate the optimization of two fare classes with the following
example. Consider a service provider that has 20 units of a resource
and sells them at a full price of $300 and a discounted price of $200.
The demand for the full price service is estimated to be normally dis-
tributed with a mean of 8 and a standard deviation of 2. Consequently,
the probability that the demand for the full price offering will exceed
y units can be expressed by using the cumulative distribution function
Φ of the standard normal distribution

y 8 0.5
PrpQ1 ¥ yq 1 Φ (6.105)
2
Note that we have added a shift of 0.5 because of the discrete nature
of the reservation units – the probability that the demand is exactly y
units can be approximated by integrating the cumulative distribution
function over the interval from y 0.5 to y 0.5. Given that we reserve
y units to be sold at full price, the marginal revenue for a full-price
unit can be defined as
r1 pyq $300 PrpQ1 ¥ yq (6.106)
In other words, the marginal revenue is the difference between the
expected revenue from the full price segment, given that y units are
allocated for this segment, and the corresponding revenue, given that
only y 1 units are allocated. The total expected revenue from the
full-price segment is the sum of the marginal revenues:
¸
y
R1 p y q r1 piq (6.107)

i 1
The revenue from the discounted segment is simply the number of

remaining units multiplied by the discounted price:
R2 pyq $200 pC yq (6.108)
The total revenue earned by a provider is the sum of the revenues
for the full price and discounted segments. Let us now calculate these
metrics for all possible values of the protection level y and put them
into table 6.9.
We can see that the highest revenue is achieved when the protection
level is 7, that is, 7 units are allocated for the full price segment and 13
units are allocated for the discounted segment. Littlewood’s rule leads
us to the same result – the marginal revenue r1 pyq that corresponds to
the right-hand side of equation 6.102 goes below the discounted price
of $200 from protection level 8 onwards.
N
y r1 y p q p q
R1 y p q
R2 y p q
R1 y p q
R2 y
1 $299 $299 $3800 $4099
2 $299 $599 $3600 $4199
3 $299 $898 $3400 $4298
4 $296 $1195 $3200 $4395
5 $287 $1483 $3000 $4483
6 $268 $1751 $2800 $4551
7 $232 $1983 $2600 $4583
8 $179 $2163 $2400 $4563
9 $120 $2283 $2200 $4483
10 $67 $2351 $2000 $4351
11 $31 $2383 $1800 $4183
12 $12 $2395 $1600 $3995
13 $3 $2398 $1400 $3798
14 $0 $2399 $1200 $3599
15 $0 $2399 $1000 $3399
16 $0 $2400 $800 $3200
17 $0 $2400 $600 $3000
18 $0 $2400 $400 $2800
19 $0 $2400 $200 $2600
20 $0 $2400 $0 $2400
Table 6.9: Example of the protection level optimization for two fare classes.
6.8.3 Allocation with Multiple Classes
Littlewood’s rule provides a compact expression for the capacity al-

location problem with two classes. In practice, however, the problem
typically has to be solved for more than two classes. The determination
of an optimal solution is quite challenging in this case, but there are
several ways to cope with it. One possible approach is to solve the prob-
lem recursively by using the assumption about sequentially arriving
demand classes, so a decision about the protection level for one class
leads to a problem of smaller dimensionality. This method allows us to
express and solve the allocation problem in terms of dynamic program-
ming. A different approach is to extend the probabilistic analysis that
we did for Littlewood’s rule and use simulations to find the optimal
protection levels [Brumelle and McGill, 1993; Talluri and Van Ryzin,
2004]. We choose to follow this latter approach here.
We have shown that, for the two-class problem, the optimal protec-
tion level for the first class is given by

p2 p1 Pr Q1 ¥ yopt
1 (6.109)
Let us take one step forward and consider a decision tree for a third
class, as shown in Figure 6.22.
Figure 6.22: A decision tree for a three-class allocation problem.
Similarly to the two-class problem, a request for the third class can
be either accepted, which earns us the revenue of p3 , or rejected. The
latter action means that the third class will be closed and all further
requests will be handled in the two-class mode. This can result in two
possible outcomes:
• If the total demand for the first and second classes is below the
protection level y2 , the unit will be lost – we closed the third class
too early.
• Otherwise, the rest of the requests will be handled as a standard

two-class problem. As we have shown in the previous section,
the average revenue per unit R2 is equal to p2 in this case, given
that the protection level y1 is set optimally, that is, in accordance
with Littlewood’s rule.
Consequently, the optimal value for y2 can be expressed as follows:

p3 p2 Pr Q1 Q2 ¥ yopt
2
Q1 ¥ yopt
1 (6.110)
We can compare equations 6.109 and 6.110 and apply the decision
tree approach recursively to find the following relationship for the op-
timal protection levels:

pj 1
pj
Pr Q1 ... Qj ¥ yopt
j

(6.111)
Q1 ¥ yopt
1 and . . . and Q1 ... Qj1 ¥ yopt
j1
Although equation 6.111 does not look as simple as Littlewood’s

rule, it provides a relatively simple way to estimate protection levels
by using simulations. Let us consider an example with three classes for
the sake of convenience, although the method can be straightforwardly
applied to any number of classes. We assume that the distributions of
demands Q1 and Q2 are known, so we can generate a relatively large
number of two-dimensional points with the coordinates defined as Q1
and Q1 Q2 .
The optimal value for protection level y1 can be estimated by using

equation 6.109 – we need to find a line that splits the points by the
first coordinate in such a way that the number of points on the left
and the number of points on the right are in the same proportion as
p1 and p2 . This is illustrated in Figure 6.23. The points on the right
satisfy the condition Q1 ¥ yopt
1 from equation 6.110, so we split them
by the second coordinate in such a way that the number of points in
the bottom section and the number of points in the top section are in
the same proportion as p2 and p1 to estimate protection level y2 .
Figure 6.23: Optimization of the protection levels for three classes by using sim-
ulations.
6.8.4 Heuristics for Multiple Classes
Although the optimal protection levels can be calculated by using

the simulation method considered in the previous section, as well as
alternative algorithms, many implementations use simpler heuristic
methods that give suboptimal solutions but are known to be very
close to optimality in most practical applications. The most important
method of this kind is the expected marginal seat revenue (EMSR)
algorithm, which has two versions: EMSRa and EMSRb [Belobaba,
1987, 1989]. Both versions attempt to adapt Littlewood’s rule to
multiple fare classes heuristically.
6.8.4.1 EMSRa
Recall that the protection level for class j is the total capacity reserved
for classes from j down to 1. If we have already calculated the protec-
tion levels for the cheaper classes from n down to j 1, the protection
level for j determines how to split the capacity between class j 1 and
the more expensive classes. The idea of EMSRa is to approximate this
protection level by the sum of the protection levels obtained by apply-
ing Littlewood’s rule for class j 1 and each of the classes from j down
to 1 separately. This means that we first calculate j pairwise protection
levels from the following equations:

pj 1 pj Pr Qj ¥ ypjjq1

pj1q
pj 1 pj1 Pr Qj1 ¥ yj 1
(6.112)
..
.

pj 1 p1 Pr Q1 ¥ ypj1q1
The final protection level is calculated as a sum of the pairwise levels,
as illustrated in Figure 6.24.
¸ pkq
j
yj yj 1
(6.113)
k 1
Comparing equations 6.112 and 6.113 with the optimal solu-
tion 6.111, we see that EMSRa uses the probabilities that separate
demands exceed the levels determined by the corresponding price
ratios to approximate the probability that the sum of demands exceeds
a certain level. As a result, EMSRa tends to be excessively conservative,
in the sense that it reserves too many units for the higher classes and,
thereby, rejects too many low-fare bookings.
6.8.4.2 EMSRb
The alternative approach to the problem is to merge the classes from j
to 1 into one virtual aggregate class that has its own demand and price
and to then apply Littlewood’s rule. The demand for an aggregate class
can be estimated as the sum of demands for the included classes:
¸
j
Qj Qk (6.114)

k 1
6.9 assortment optimization 457
Figure 6.24: Example of EMSRa for three fare classes.
The “price” of the aggregated class can be defined as the weighted

average price of the included classes:
°j
k 1 pk E
rQk s
pj °j (6.115)
k1 E rQk s
The protection level for class j can then be estimated by applying

Littlewood’s rule to the aggregated class and the previous class j 1,
that is

pj 1 pj Pr Qj ¥ yj (6.116)
It is generally accepted that both EMSRa and EMSRb provide ap-

proximations that are very close to the optimal solutions. EMSRb was
developed to improve EMSRa and is sometimes said to perform better
that EMSRa, but the experimental results with both real data and sim-
ulations show that neither method consistently outperforms the other
[Talluri and Van Ryzin, 2004].
6.9 assortment optimization
The customer’s willingness to pay for a given product or service is

almost always influenced by alternative options, that is, the ability
to choose a competing or substitutable offering. The impact of these
forces on demand functions and, eventually, profits can be more or less
significant depending on the particular industry and business case. We
have already discussed that pricing decisions for multiple products or
customer segments might need to be jointly optimized in the case of
demand shifting or dependent demands within a category. This prob-
lem grows more challenging in retail industries that deal with tens or
hundreds of thousands of products, so not only prices but also other

marketing and enterprise resources should be optimized to take de-
mand dependencies and substitution effects into account. We devote
this section to examining this type of optimization task in detail.
6.9.1 Store-Layout Optimization
Demand dependencies and affinities between products and categories

suggest an opportunity to cross-sell related or complementary offer-
ings to buyers of certain products. Examples of such complementary
products include coffee and sugar, cosmetics and handbags, and even
beer and diapers. One possible way to leverage these opportunities is
to identify products that customers tend to buy together and optimize
the store layout or placement of content items on a media site to sim-
plify navigation and encourage customers to buy product sets instead
of single products.
The first part of this strategy, the identification of items that are fre-
quently bought together, can be accomplished by using a basic market
basket analysis. Let us assume that we have a history of sales transac-
tions where each transaction t is represented as a set of items r pur-
chased in this transaction:
tn trn1 , rn2 , . . . , rnk u (6.117)
The support for an item or item set is the fraction of transactions in

the history that contain that item or item set. In other words, it is the
empirical probability that a randomly selected transaction contains a
given item or item set. The lift is defined for a pair of items as the ratio
of the support for the pair divided by the product of the supports for
each item separately:
support pra and rb q

λpra , rb q
support pra q support prb q
(6.118)
In other words, the lift is the ratio between the observed probability
of two items co-occurring and the co-occurrence probability calculated
under the assumption that the items are independent. Consequently,
a lift higher than one indicates affinity of the items (assuming statisti-
cal significance of the results, of course). We can measure the lift not
only for pairs of products but for pairs of categories as well, by map-
ping each item in the transaction history to its category and evaluating
expression 6.118 with the assumption that ra and rb are categories.
Let us now return to the store-layout optimization problem. We have
n product categories and n product locations, such as aisles or shelves.
The optimization problem can then be stated as the assignment of all
categories to different locations with the goal of placing categories with

high pairwise affinity close to each other. First, let us calculate the
matrix of pairwise lifts for the categories:
(
L λij , i, j 1, . . . , n (6.119)
Next, we need some sort of a distance matrix for locations:

(
D dij , i, j 1, . . . , n (6.120)
in which dij is the distance between locations i and j. The distance

can be defined in different ways, for example, it can be a binary vari-
able that is equal to one if the locations are adjacent and zero otherwise.
The optimization problem can then be defined as
ņ ņ
max λij dπpiq,πpjq (6.121)
π

i 1j 1
in which πpiq is a permutation function that maps categories to lo-

cations, that is, πpxq equals y when category number x is assigned to
location y. Problem 6.121 is an instance of the quadratic assignment
problem1 (QAP). It is a well-studied combinatorial optimization prob-
lem that, however, is fairly difficult computationally. Nevertheless, it
has been reported that this method has been used in practice to opti-
mize layouts for convenience stores [Winston, 2014].
example 6.8
The following example illustrates store-layout optimization. Consider

a grocery store that offers six categories of products: dairy, deli, bakery,

drinks, produce, and frozen food. The matrix of pairwise lifts for these
categories is estimated from the historical data:
Dairy Deli Bakery Drinks Produce Frozen

Dairy 1.00 0.80 1.30 0.90 1.00 0.90
Deli 0.80 1.00 1.20 1.10 1.30 0.80

Bakery
L 1.30 1.20 1.00 1.30 1.20 0.90
Drinks
0.90 1.10 1.30 1.00 1.20 1.50

Produce 1.00 1.30 1.20 1.20 1.00 0.80
Frozen 0.90 0.80 0.90 1.50 0.80 1.00
(6.122)
1 QAP was first introduced in the context of operations research to model the following real-
life problem. There is a set of facilities and a set of locations. The objective is to assign each
facility to a location such that the total cost is minimized, with the assignment cost being a
product of the distance between locations and the flow between the facilities.
The floor plan is the 2 3 grid shown in Figure 6.25, in which each
placement represents a display shelf. In total, we have six available
placements for six categories.
Figure 6.25: Store floor plan.
The distance matrix for this layout is defined as follows:
1 2 3 4 5 6

1 0 1 0 1 0 0
2 1 0 1 0 1 0

D
3 0 1 0 0 0 1 (6.123)
4 1 0 0 0 1 0

5 0 1 0 1 0 1
6 0 0 1 0 1 0
We assume that the distance is only equal to one between adjacent
placements and is zero otherwise. For example, placement number five
has a distance of one to cells two, four, and six. By solving optimization
problem 6.121 for the matrices specified above, we find the optimal lay-
out presented in Figure 6.26. This small example can be easily solved
by evaluating all 6! 720 possible permutations, but larger problems
require the use of optimization software that can handle QAP or one
of its relaxations.
N
6.9.2 Category Management
The problem of category management appears when a seller needs to

optimize the overall performance of a product category as opposed to
the performance of a single product. This problem is very typical in
the retail industry because a retailer can change product assortment
within a category relatively easily and the primary goal is to utilize the
available resources, such as shelf space, in the most efficient way.
A category represents a relatively cohesive set of products that have
a lot in common such as “dairy desserts” or “women’s jeans”, so it is
Figure 6.26: One of the optimal store layouts for the example with six categories.
Alternative optimal layouts can, of course, be obtained by mirroring
this grid horizontally or vertically.
generally possible that customers might be willing to substitute one

product with another if the product of their choice is not available for
some reason. The reasons for product unavailability can include both
deliberate assortment changes and temporary stockouts. One of the
main goals of category management is to find a subset of products
that meets physical constraints, such as available shelf space, and max-
imizes the profit by taking advantage of the substitution effect in the
optimal way. Alternatively, this problem can be stated as the identifica-
tion of the least significant products, those that can be excluded from
the assortment and substituted by other products without a negative
impact on the profits. The outputs of this analysis can then be applied
in the optimization of several different controls:
• Product stock levels can be optimized to account for substitution
effects and potential losses caused by stockouts.
• Shelf layouts can be optimized to adjust the relative product

shares on a shelf.
• Assortment can be optimized by the introduction or removal of

products from the assortment.
From an econometric perspective, the problem of category manage-
ment arises from the law of diminishing returns or, more specifically,
the fact that revenues and costs depend on the category size in different
ways. The general tendency is that consumer buying capacity comes to
saturation at some point, whereas costs continue to grow because of
the increasing selling area and other operational costs, as shown in
Figure 6.27. This tendency leads to the category optimization problem.
It is a very challenging problem because it requires the modeling of an
entire category with the interdependencies between the products in it
accounted for. However, despite these challenges, a practically feasible
assortment optimization model has been developed and evaluated at
Albert Heijn, a supermarket chain in the Netherlands [Kök and Fisher,

2007]. We spend the rest of this section studying their solution.
Figure 6.27: Diminishing returns in category management.
Consider a supermarket store chain that operates multiple stores.

Each store sells many categories of products, but, as we discussed
above, demands are assumed to be dependent only within a category
and categories are considered to be independent. Consequently, a re-
tailer solves an assortment optimization problem for each category in
each store independently. Let us first introduce the following notation
that applies to a single category in a single store:
• N t1, 2, . . . , Ju: the maximal set of products in a category that
a retailer offers to its customers, that is, the full assortment.
• fi P t0, 1, 2, . . . u: the stock level for product j. A retailer optimizes

its assortment by choosing f to be zero (product is not present in
the assortment) or non-zero. Let us also denote the vector of stock
levels for all J products as f pf1 , . . . , fJ q.
• F0 : the total inventory capacity measured in the same units as

stock levels. It is assumed that the sum of stock levels for all
products cannot exceed F0 . The total capacity can be constrained
by warehouse space or available shelf space in a store.
• Nh N: the assortment in store h, a subset of the full assort-

ment.
• dj : the original demand rate for product j, that is, the number
of customers who would select product j if presented with full
assortment N. We also denote the vector of demands for all prod-
ucts as d pd1 , . . . , dJ q.
• Dj : the observed demand rate for the products, that is, the actual
number of customers per day who selected product j because
of their original intention or substitution. The observed demand
for a given product depends on the original demand and the
availability of other products because of the substitution effect,
so it can be thought as function Dj pf, dq.
With the above notation, the assortment optimization problem for a
given store and a given category can be specified as follows:
¸
max Gj fj , Dj pf, dq
f
P
j N
¸ (6.124)
subject to fj ¤ F0
j
in which Gj is a function that describes the profit for a given product

and corresponding observed demand. This function heavily depends
on a retailer’s business model, so we can outline a few generic tem-
plates that can be customized for practical usage. The simplest way
to model profit is to multiply the observed demand by the product
margin m:
Gj pfj , Dj q mj Dj (6.125)
Equation 6.125 implicitly assumes perfect replenishment and the ab-

sence of stockouts. This might be the case for fast-moving consumer
goods like groceries, but other retail domains such as apparel should
probably take stockouts into account by taking the minimum of the
demand and the actual stock levels:
Gj pfj , Dj q mj minpDj , fj q (6.126)
Retailers of perishable goods should also take into account the losses
owing to disposed-of inventory, which can be modeled by introducing
a per-unit disposal loss L that applies to unsold inventory:

Gj pfj , Dj q mj minpDj , fj q Li fi minpDj , fj q (6.127)
For the sake of brevity, we hereafter assume that all products are
perfectly replenished, so stockouts are not possible or are negligible.
This allows us to treat fj P t0, 1u as a binary variable that indicates the

presence of a product in the assortment.
To solve optimization problem 6.124, one needs to define the ob-
served demand function. With the stockout-free assumption that we
made above, the demand function can be modeled as follows:
¸
Dj pf, dq dj αkÑj dk (6.128)

k:fk 0
in which αkÑj is the probability of substitution of product k by prod-

uct j. Formula 6.128 is relatively straightforward: the first term is the
original demand and the second term corresponds to the cumulative
substitution effect from all products that are evicted from the assort-
ment set.
Equation 6.128 requires the estimation of the substitution probabili-
ties αkÑj and original demand rates dj . In order to do this estimation,
let us assume that the following variables are known (we already dis-
cussed demand prediction in Section 6.6):
• Qjh , j P Nh : the demand for product j per customer at store
h. If Kh is the number of customers who visit store h during the
day, Dj Kh Qjh .
• Q0jh , j P N: the demand for product j per customer at store h

with a full assortment (let us assume that stores with full assort-
ments exist). Q0jh corresponds to the original demand because no
substitution happens at stores with the full assortment.
Substitution rates αkÑj are challenging to estimate because up to J2
different rates can exist for the assortment of J products. It is unlikely
that a retailer has enough data to estimate this variety of rates reli-
ably. However, there is empirical evidence that the following simplistic
model of customer behavior is sufficiently accurate in practice and re-
quires the estimation of just one variable instead of J2 : if product k is
not available, customers either select their second-choice product j as a
substitution with the probability δ, which is the same for all products
in a category, or no purchase takes place, with the probability p1 δq.
This model leads to the following simple equation for the substitution
rate:
αkÑj δ |N1 | (6.129)
In order to estimate δ, let us define the total demand at a given store

as the sum of Qjh values, which can be estimated from the data:
¸
Sh Qjh (6.130)
P
j Nh
On the other hand, the same value can be estimated according to

expression 6.128 as follows:

¸ ¸
ph pδq
S Q0 αkÑ 0
j Qkh
jh
P
j Nh P z
k N Nh
(6.131)
¸ ¸ ¸
Q0jh
δ 0
|N| Qkh
P
j Nh P
j Nh k N NhP z
Now δ can be estimated by solving the following optimization prob-
lem, which minimizes the discrepancy between the observed and pre-
dicted values of the total demand:
¸ 2
δ0 argmax ph pδq Sh
S (6.132)
¤¤
0 δ 1 h
The next step in solving optimization problem 6.124 is to compute

the original demand rates that are used in equation 6.128. We first note
that the total demand for all products in N at store h can be computed
as follows:
¸
Th Vh Q0jh
Sh
j NP ph δ0
S p q (6.133)
in which Vh is the total number of customers visiting store h per

day. In equation 6.133, the sum of all Q0jh multiplied by Vh represents
the total demand given a full assortment. However, values Q0jh are esti-
mated for stores with a full assortment, so the specifics of a given store
h (e. g., location, store size in square feet, etc.) are not modeled. This
is compensated for by scaling the ratio of estimated category demand
from equation 6.130 to the predicted demand from equation 6.131.
In a store with a restricted assortment, the total demand Th is the
sum of two components: the demand that comes from the products
included in the assortment of a given store and the demand for other
products in N. The ratio between these two components can be ex-
pressed via Q0jh as follows:
°
rh °
P Q0jh
j Nh
(6.134)
jPN Qjh
0
Consequently, Th rh represents the fraction of the demand at-

tributed to the products in the assortment, and Th p1 rh q represents
the remaining fraction attributed to the products that are not in the
assortment. Finally, we compute the demand for a single product as a
fraction of the total demand proportional to the estimated per-product

demand:
$
'
'
'
'
& Th rh °
'
Qjh
, if j P Nh
djh ' Q P
j Nh jh
(6.135)
' Q0jh
'
'
'
% Th p1 rh q ° , if j R Nh
P z
j N Nh Q0jh
All coefficients in equations 6.135 and 6.132 can be estimated from

the data, so we can roll up all formulas to original optimization prob-
lem 6.124, which can be solved by using numerical methods.
Equation 6.124 will produce a set of presumably optimal stock levels
fj for all products. These levels can be used to adjust inventory and
optimize shelf layout. It is important to note that the model enables
a retailer to perform what-if analyses to evaluate how changes in as-
sortment and stock levels might impact the gross margin. In particular,
a retailer can plot curves that show the expected gross margin as a
function of stock levels for a given product or group of products. Such
curves are especially descriptive for perishable products because the
gross margin is a convex function that is zero when the stock level is
zero and also zero when the stock level is too high, which leads to
losses from expired products, with a maximum in between these two
extremes.
6.10 architecture of price management systems
Although the design and implementation of algorithmic price manage-

ment systems can vary significantly across industries, price manage-
ment typically includes several principal processes that can be thought
of as functional components in the reference logical architecture. A
high-level diagram that depicts these key components and their rela-
tionships is presented in Figure 6.28 and includes three major subsys-
tems.
transactional system The purpose of the transactional system is

to execute the pricing decisions received from the optimization
system. In certain environments, the optimized prices and pric-
ing rules can be loaded into multiple transactional systems that
calculate final prices independently. For example, a retailer can
refresh prices in store point of sales (POS) nightly or weekly, as
well as updating prices in its eCommerce platform. Each store
or eCommerce platform then operates independently. In many
other environments, a single transactional system, which can be
6.10 architecture of price management systems 467
Figure 6.28: High-level architecture of a price management system.
centralized or distributed in technical terms, processes real-time

requests for prices or resource quotes. This approach is gener-
ally preferable because complex pricing decisions can be made
in a consistent way. For example, a retailer often calculates the
final price for every transaction based on the list prices, active
discounts, customer loyalty number, entered or scanned promo-
tion codes, and other factors. The reservation systems in service
industries, such as airlines or hotels, are even more interactive be-
cause they not only apply the pricing rules but also keep track of
booked resources and update records in the operational database
after each transaction.
A transactional system can be owned by a seller or can be a
shared resource provided a third-party. For instance, airline tick-
ets are typically reserved through global distribution systems
(GDS) operated by dedicated companies, and the airlines set the

booking limits directly in the GDS.
analytics Demand prediction plays a mission-critical role in price

management because it enables quantitative evaluation of pric-
ing and assortment decisions. This requires the creation of an
analytical infrastructure that collects and stores historical data
and supports training and evaluation of predictive models. We
have seen that demand prediction models can use quite a wide
range of information, including product data, discounts, compet-
itive prices, weather data, and holiday calendars. Many of these
data elements can be pulled from the enterprise resource man-
agement systems and third-party data providers.
The demand model training pipeline is often automated and
works in a loop or incrementally to refresh the model as new data
arrive. This ensures that models are up to date with changes in
the business and competition landscapes.
optimization system The main component of the optimization

system is a solver, which typically uses numerical optimiza-
tion methods to find the price schedule or booking levels
that maximize the revenue. The solver is configured from
the administration component to set optimization parameters
and constraints. In certain applications, such as airlines, the
optimization process can be administered by a team of analysts
who monitor system performance and can correct the decisions
in special cases, such as large public events.
Similarly to the demand modeling process, optimization is also
executed in a loop to recalculate prices or booking limits as the
sales and inventory data are updated. The optimization process,
however, is typically executed more frequently than analytics to
catch up with rapidly decreasing capacity levels. This continual
optimization is almost always used for dynamic pricing and re-
source allocation, as well as other price optimization methods
that may not model temporal changes directly but that contin-
ually adjust prices when applied repeatedly and, thus, become
dynamic.
6.11 summary
• Pricing decisions are extremely important for firm competitiveness

and profitability. Improvements in pricing often have much more
impact on profits than comparable improvements in quantities sold
6.11 summary 469
(advertising and sale channels), variable costs (supply and produc-

tion), or fixed costs (operations and asset management).
• Price is a monetary measure of value. Boundaries for a good price

can be determined by estimating consumer utility and comparable
alternatives. The perceived surplus of a product or service, however,
depends on how exactly the product information and price is com-
municated. In particular, price drops and multiple unbundled gains
are known to be efficient. This provides fundamental justification for
markdowns and discounts.
• Demand is determined by customers’ willingness to pay. Different

distributions of willingness to pay result in different demand curves.
The basic demand curves include linear, constant-elasticity, and logit
functions.
• The basic price structures include unit price, segmented price, two-
part tariffs, tying arrangements, and bundling. All of these struc-
tures can be optimized if a global demand curve is accurately esti-
mated.
• The basic demand curves do not account for seasonality, compe-

tition, and product properties. Algorithmic price optimization re-
quires more powerful demand models to be developed. A significant
number of such models can be found in the literature for different
industries.
• Price optimization typically requires the solution of a numerical op-

timization problem that is tailored for specific structural, supply,
and demand constraints. The two major business cases for price op-
timization are price differentiation, which aims to optimize prices
for multiple market segments, and dynamic pricing, which can also
be considered as a market segmentation technique. The main con-
straints accounted for in price optimization include interdependent
demand functions, limited capacity, and perishable inventories.
• Service industries, such as airlines, hotels, freight transport, and

sports, can reserve fractions of their capacity for different fare classes.
This resource allocation approach can be considered as an alterna-
tive to dynamic pricing. The most basic resource allocation methods
allow booking levels to be optimized for individual units of con-
straint capacity, such as seats in an airplane, but a large number of
more advanced methods exist that are able to optimize networks of
resources and handle booking cancellations.
• Assortment optimization is closely related to price management be-

cause both problems are based on demand prediction. Assortment
optimization focuses on modeling dependencies between demands
on different products and categories, thereby enabling the analy-
sis of assortment changes and enterprise resource reallocation that
could impact profits.
• Price, booking limits, and assortment can be considered as different

controls that a seller can use to execute pricing decisions. All opti-
mization methods related to these controls use demand prediction
as a basic building block but perform different business actions to
accommodate this prediction.
A
APPENDIX: DIRICHLET DISTRIBUTION
Dirichlet distribution is a relatively advanced topic, especially in the

context of marketing applications, so we will provide a brief introduc-
tion here. We use the Dirichlet distribution in Chapter 3 where we
discuss observational studies and Chapter 4 where the Dirichlet distri-
bution is used in topic modeling.
In most marketing applications, we deal with probability distribu-
tions of random variables, such as events or counts. Although the
Dirichlet distribution can be viewed in this way, it also has a sub-
stantially different meaning that is important for our purposes. Let
us consider a simplistic example that illustrates this aspect [Frigyik
et al., 2010]. A six-sided die can be viewed as a discrete probability
distribution that generates numbers from one to six. With a perfect die,
all of the numbers have the same probability, equal to one sixth. The
probability distribution of a real die, however, would deviate from the
uniform one because of imperfect manufacturing and other physical
factors. If we take a bag of 100 dice, each die corresponds to its own
probability mass function (PMF), and the bag of dice corresponds to the
distribution of PMFs. The properties of this distribution depend on
the quality of the dice: the PMFs can substantially deviate from the
uniform distribution in the case of low-quality dice or can be almost
identical to it in the case of precisely manufactured ones. This distribu-
tion of PMFs can be described by using the Dirichlet distribution.
A more practical and relevant example is a collection of text docu-
ments. Given that documents contain m distinct words in total, each
document can be viewed as a PMF that can be estimated by counting
frequencies for each word in the document. A collection of documents
is then a collection of PMFs, and we can choose the parameters of a
Dirichlet distribution to fit this collection. More formally, each docu-
ment d can be modeled as a vector of m word probabilities that must
add up to one:
pd1 ... pdm 1, pdi P r0, 1s (A.1)
471
472 appendix: dirichlet distribution
Geometrically, this equation describes an pm 1q-dimensional

simplex in an m-dimensional space. For example, a collection of doc-
uments with three distinct words corresponds to a two-dimensional
triangle (simplex) in a three-dimensional space, as illustrated in
Figure A.1. Each point on the simplex corresponds to a valid PMF,
whereas all other points in the space do not match any valid PMF.
We can generate a collection of m documents by specifying a dis-
tribution over the simplex, drawing m PMFs from this distribution,
and then generating the m-th document by drawing terms from the
corresponding PMF.
Pr(t1)
1.0
B
A
Pr(t2)
1.0
1.0
Pr(t3)
Figure A.1: A three-dimensional probability simplex. Point A corresponds to a

document with a uniform distribution of terms; point B corresponds
to a document where term t1 is much more probable than the other
two terms.
The Dirichlet distribution is the probability distribution over the sim-

plex. Assuming m dimensions, each instance drawn from the Dirichlet
distribution is an m-component probability mass function:
¸
p pp1 , . . . , pm q , pi 1 (A.2)
i
The distribution itself is specified by a vector of m parameters:
α pα1 , . . . , αm q , αi ¡0 (A.3)
in which each parameter can be thought of as a weighting of the corre-

sponding component. The probability density function of the Dirichlet
distribution is then defined as
1 ¹ αi 1
m
Dirpαq
Bpαq
pi (A.4)

i 1
appendix: dirichlet distribution 473
in which Bpαq, the normalization constant, is given by

±m
Bpαq p q
i°1 Γ αi
(A.5)

m
Γ i 1 αi
If all of the elements making up the parameter vector have the same
value, the distribution is completely specified by this single value,
called the concentration parameter. Probability density functions for the
Dirichlet distribution in a three-dimensional space are visualized for
different parameter values in Figure A.2.
0.0
0.0
0.5 0.5
1.0 1.0
0.0 0.0
α = (2.0, 2.0, 2.0) α = (5.0, 5.0, 5.0)

1.0 1.0
0.0 0.0
0.5 0.5
1.0 1.0
0.0 0.0
α = (2.0, 5.0, 10.0) α = (0.9, 0.9, 0.9)

1.0 1.0
Figure A.2: Density plots for the Dirichlet distribution over the probability sim-
plex in a three-dimensional space.
The density function is completely flat when α p1, 1, 1q. The den-
sity function is bell-shaped and symmetric about the center of the
simplex when the parameter vector is flat. If the parameter vector is
not flat, the bell is shifted in the direction of the parameters with the
biggest magnitudes. Finally, the important thing to note is that the
Dirichlet distribution is sparse for the parameter elements with small
magnitudes, in the sense that the density is concentrated in the corners,
and, consequently, the PMFs drawn from such a distribution tend to
have a strong bias towards a small subset of terms [Telgarsky, 2013].
INDEX
Acquisition, 84 Canned search results, 191

Adstock model, 139 Carpet bombing, 149
Decay parameter, 139 Cascading hybrid
Advertiser, 5, 144 recommender, 357
Algorithmic attribution, 154 Catalog coverage, 287
Algorithmic marketing, 4 Causal inference, 163
Analogy, 247 Censoring, 59, 110
Association rule, 371 Channel, 145
Attribution, 146 Channel contribution, 140
Attrition, 84
Channel mix modeling, 138
Auction, 68
Churn, 84
Backpropagation, 243 Classification, 25
Bag of words, 201, 296 Classification tree, 40
Bayes rule, 327 Cluster, 49
Beta distribution, 159 Clustering model, 108
Bid, 68 Cold-start problem, 293, 304,
Bid price, 153 347
Big data, 15 Collaborative filtering, 303
Bigram, 205 Competing products, 441
Binomial distribution, 158 Concept, 224
BOGO promotion, 123 Concept search, 219, 253
Booking request, 448 Concept space, 332
Boolean query, 195 Confidence, 371
Boost and bury, 190 Conjoint analysis, 389
Brand, 77, 144 Conjugate prior, 159
Brand proximity, 151
Consumer, 77
Brand safety, 147
Consumer choice theory, 52
Breslow estimator, 67
Consumer packaged goods, 76
Broken assortment effect, 421
Content analyzer, 269, 378
Budget allocation, 86
Business objective, 21 Content filtering, 291
Business process, 3 Content-boosted collaborative
Business strategy, 3 filtering, 354
Business to consumer, 11 Contextual recommendations,
368
Campaign response, 88 Continuous bag of words, 240
Campaign template, 87, 122 Control group, 93, 173
475
476 INDEX
Controlled precision Document field, 207

reduction, 253, 361 Document frequency, 202
Conversion, 146
Conversion rate, 157, 192 Economic model, 21
Coordination factor, 203 Elasticity of demand, 394
Cosine similarity, 200 Email fatigue, 82
Cost per acquisition, 147 Ensemble learning, 346
Cost per action, 148 Enterprise search, 181
Cost per click, 148 Entrance fee, 406
Cost per mile, 6, 148 Error decorrelation, 349
Cox model, 65 Euclidean norm, 199
CPA-LT model, 149 Expectation–maximization
Credible interval, 159 algorithm, 235
Cross-sell, 84, 458 Expected loss, 135
Cumulative gain, 189 Expected Marginal Seat
Customer, 77 Revenue, 455
Customer equity, 113 Expected relevance, 265
Customer experience, 121 Expert search, 181
Customer journey, 80
Customer journey map, 121 Fare class, 447
Customer life cycle, 79, 83 Fast-moving consumer goods,
Customer profile, 96 402
Customer ratings, 276 Feature, 13, 24
Feature augmentation, 354
Data point, 23 Feature engineering, 25
Decorrelation, 42 Feature mapping, 37
Defection, 84 Feature preparation, 175
Demand curve, 394 Flash sales, 422
Demand prediction, 414 Fragmented features, 213
Demand shifting, 428 Free-standing inserts, 123
Demand uncertainty, 434 Frequency capping, 126
Demand unconstraining, 422
Descriptive analytics, 19 Genre expansion, 220
Design matrix, 24 Gibbs sampling, 169
Dimensionality reduction, 46 Global distribution system,
Diminishing returns, 141, 461 467
Dirichlet distribution, 171, 238, Gradient descent, 335
471
Concentration parameter, Hard targeting, 124
473 Hazard function, 62
Discounted cumulative gain, Hazard risk ratio, 65
189, 285 Hierarchical aggregation, 359,
Discounted similarity, 312, 351 361
Distributed lag, 142 Hybrid recommender, 346
INDEX 477
Imperfect segmentation, 405 Look-alike modeling, 100, 149

Implicit feedback, 278, 343, Low-rank approximation, 47
370 Loyalty ID, 78
Implicit feedback matrix, 344 Loyalty–monetary
Impression, 146 segmentation, 97
Independence assumption, 35
Independence of irrelevant Manufacturer, 77
alternatives, 55 Manufacturer-sponsored
Indexer, 269 campaign, 80
Information need, 185 Market basket analysis, 372,
Information retrieval, 180 458
Inverse document frequency, Market cannibalization, 405
202 Market segmentation, 401
Inverse user frequency, 312 Marketer, 144
Item bias, 306 Marketing channel, 14, 78
Itemset, 371 Marketing mix, 3, 11
Marketing mix modeling, 138
Kaplan–Meier estimator, 61 Marketing science, 10
Kernel function, 38 Marketplace, 7
Kernel trick, 39 Markov chain, 116
Kullback-Leibler divergence, Matching, 193
28 Matrix completion, 281, 304
Maximization, 84
Laplace estimator, 328 McRank algorithm, 265
Last-touch attribution, 148 Mean average precision, 187
Latent concepts, 224 Mean squared error, 225, 282
Latent Dirichlet allocation, 237 Measurements, 15
Latent factors, 166, 332 Meta-features, 351
Latent semantic analysis, 224 Metered price, 406
Latent semantic indexing, 226 Micro-decisioning, 1
Latent semantic space, 225 Micro-moment behavior, 2
Latent variable, 231 Mixed bundling, 410
Latent variable model, 233 Mixed hybrid recommender,
Learning rate, 335 356
Learning to rank, 260 Mixture models, 51
Lift, 128, 458 Model, 14
Lift chart, 128 Model fitting, 26
Likelihood, 28 Model-based collaborative
Linear separability, 29 filtering, 324
Listwise ranking, 264 Monte Carlo simulation, 119,
Littlewood’s rule, 451 160
Logistic function, 32 Multinomial logit, 397
Logistic regression, 32, 103 Multinomial logit model, 54,
Long tail, 273 418
478 INDEX
Multipart pricing, 406 Pressure rule, 126

Price bundling, 410
N-gram, 205 Price discrimination, 425
Naive Bayes classifier, 34 Price–response function, 393
Naive Bayes recommender, Principal component analysis,
354 42
Natural buyer, 149 Principal components, 44
Natural trier, 102 Private label, 77, 81
Nearest neighbor algorithm, Probabilistic latent semantic
26 analysis , 232
Nested booking limits, 449 Probabilistic topic modeling,
Net present value, 114 231
Network management, 448 Probability mass function, 471
Non-negative matrix Product collection, 255
factorization, 342 Product discovery, 75, 179
Normalized discounted Product information
cumulative gain, 189 management, 270
Profile model, 292
Operations research, 10
Profit equation, 387
Ordinal ratings, 277
Programmatic, 4
Organic relevance, 181
Projected gradient descent,
Overbooking, 449
340
Promotion, 78
PageRank algorithm, 181
Promotion distribution
Pairwise ranking, 264
network, 78
Partial likelihood, 66
Propensity, 85, 99
Pay per click, 6
Proportional hazard, 64
Pearson correlation coefficient,
Prospect theory, 390
311, 316
Perishability, 425, 435 Protection level, 449
Persona, 108, 114 Psychographic features, 301
Phrase query, 196 Public service announcements,
Pilot campaign, 105 163
Pixel, 146 Publisher, 5, 145
Planner, 175 Purchasing cycle, 137
Pointwise ranking, 263 Purchasing intent, 179
Polysemy, 218, 237 Pure bundling, 410
Porter stemmer, 197
Post-view action, 147 Quadratic assignment
Potential outcome, 164 problem, 459
Precision, 185 Query chain, 266
Precision–recall curve, 187 Query engine, 269
Predictive analytics, 19 Query folding, 226
Prescriptive analytics, 20 Query preprocessor, 270
INDEX 479
Random forest, 41 Semantic space, 223

Random utility model, 53 Shingle, 205
Randomized experiment, 161 Showroom behavior, 405
Randomized Signal, 14
recommendations, Signal disbalance, 207
379 Singular value decomposition,
Rank, 260 46, 225, 333
Ranking, 193 Soft targeting, 124
Rating baseline estimate, 305, Stacking, 348
364 Standard score, 313
Rating biases, 306 Stemming, 197
Rating matrix, 277 Stochastic gradient descent,
Rating sparsity, 278 335
Real-time bidding, 145 Stock keeping unit, 255, 419
Recall, 185 Stockout event, 421
Recommendation context, 357 Stockout price, 431
Redemption, 78 Stop word, 197
Regression, 25 Strategic pricing, 383
Relevance, 185 Sum of squared errors, 30
Relevance grade, 260 Supervised learning, 24, 25
Relevance signal, 193 Support, 371, 458
Relevance tuning, 258 Support vector machines, 39
Replenishment notifications, Survival analysis, 58
137 Survival curve, 62
Representation learning, 25, 41 Survival function, 60, 120
Reservation system, 448 Survival time, 59
Residual errors, 350 SVD model, 343
Resource allocation, 15 SVD++ model, 345, 367
Response label, 24 Switching recommender, 346
Result postprocessor, 270 Synonym contraction, 220
Retailer, 77 Synonym expansion, 220
Retention, 84 Synonymy, 218
Return on investment, 80 Syntactic search, 218
Revenue management, 9, 383
RFM analysis, 98, 116 Tapestry recommender
Ridge regression, 348 system, 303
Risk set, 66 Targeted campaign, 79
Root mean squared error, 283 Targeting, 75
Targeting model, 94
Savability, 136 Targeting system, 78
Search intent, 185 Term, 196
Search result specification, 215 Term frequency, 201
Segmentation, 49, 107 Term hierarchy, 220
Semantic search, 219 Term–document matrix, 224
480 INDEX
Term-centric scoring, 214 Uplift modeling, 91, 93, 147

Test group, 93 User, 146
TFIDF , 369 User bias, 306
TFIDF model, 202 Utility, 52, 388
Thesaurus, 219
TimeSVD++ model, 367 Value exchange model, 389
Tokenization, 194 Variable demand, 434
Topic model, 238 Vector space model, 200
Treatment, 78 Vickrey auction, 69
Two-part tariffs, 406
Tying arrangements, 406 Web search, 180
Willingness to pay, 393
Unary ratings, 278 Word embedding, 222, 224
Unit price, 399
Unsupervised learning, 42 Yield management, 383
Up-sell, 84
Uplift, 136, 161 Z-score, 313
BIBLIOGRAPHY
Adomavicius, G. and Tuzhilin, A. (2008). Context-aware recommender

systems. In Proceedings of the 2008 ACM Conference on Recommender
Systems, RecSys ’08, pages 335–336, New York, NY, USA. ACM.
Aggarwal, C. C. (2016). Recommender Systems: The Textbook. Springer

Publishing Company, Incorporated.
Almon, S. (1965). The distributed lag between capital appropriations

and expenditures. Econometrica, 33(1):178–196.
Anderson, C. (2008). The Long Tail: Why the Future of Business is Selling
Less of More. Hyperion.
Anupindi, R., Dada, M., and Gupta, S. (1998). Estimation of consumer

demand with stock-out based substitution: An application to vend-
ing machine products. Marketing Science, 17(4):406–423.
Artun, O. and Levin, D. (2015). Predictive Marketing: Easy Ways Every

Marketer Can Use Customer Analytics and Big Data. Wiley.
Asuncion, A., Welling, M., Smyth, P., and Teh, Y. W. (2009). On smooth-
ing and inference for topic models. In Proceedings of the Twenty-Fifth
Conference on Uncertainty in Artificial Intelligence, UAI ’09, pages 27–
34. AUAI Press.
Bell, R. M. and Koren, Y. (2007). Scalable collaborative filtering with

jointly derived neighborhood interpolation weights. In Proceedings of
the 2007 Seventh IEEE International Conference on Data Mining, ICDM
’07, pages 43–52. IEEE Computer Society.
Bellogín, A., Castells, P., and Cantador, I. (2014). Neighbor selection

and weighting in user-based collaborative filtering: A performance
prediction approach. ACM Trans. Web, 8(2):12:1–12:30.
Belobaba, P. (1987). Air travel demand and airline seat inventory man-
agement. Technical report, Cambridge, MA: Flight Transportation
Laboratory, Massachusetts Institute of Technology.
Belobaba, P. P. (1989). Application of a probabilistic decision model to

airline seat inventory control. Operations Research, 37(2):183–197.
Bergamaschi, S., Po, L., and Sorrentino, S. (2014). Comparing topic

models for a movie recommendation system.
481
482 bibliography
Berger, P. and Nasr, N. (1998). Customer Lifetime Value: Marketing Models

and Applications, volume 12.
Berry, M. (2009). Differential Response or Uplift Modeling.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer-

Verlag New York, Inc., Secaucus, NJ, USA.
Blattberg, R. C. and Deighton, J. A. (1996). Manage Marketing by the

Customer Equity Test, volume 74.
Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent dirichlet alloca-
tion. J. Mach. Learn. Res., 3:993–1022.
Bradford, R. (2008). An empirical study of required dimensionality for

large-scale latent semantic indexing applications. In Proceedings of
the 17th ACM Conference on Information and Knowledge Management.
ACM.
Breese, J. S., Heckerman, D., and Kadie, C. (1998). Empirical Analysis of

Predictive Algorithms for Collaborative Filtering. UAI’98. Morgan Kauf-
mann Publishers Inc., San Francisco, CA, USA.
Breslow, N. E. (1972). Discussion following “regression models and life

tables”. Journal of the Royal Statistical Society, 34:187–220.
Breslow, N. E. (1974). Covariance analysis of censored survival data.

Biometrics, 30(1):89–99.
Broadbent, S. (1979). One way tv advertisements work. Journal of the

Market Research Society, 23(3).
Brumelle, S. L. and McGill, J. I. (1993). Airline Seat Allocation with Mul-

tiple Nested Fare Classes, volume 41.
Burges, C., Shaked, T., Renshaw, E., Lazier, A., Deeds, M., Hamilton, N.,
and Hullender, G. (2005). Learning to rank using gradient descent.
In Proceedings of the 22Nd International Conference on Machine Learning,
ICML ’05.
Burges, C. J. C. (2010). From RankNet to LambdaRank to Lamb-

daMART: An overview. Technical report, Microsoft Research.
Burke, R. (2002). Hybrid recommender systems: Survey and experi-

ments. User Modeling and User-Adapted Interaction, 12(4):331–370.
bibliography 483
Cao, Z., Qin, T., Liu, T.-Y., Tsai, M.-F., and Li, H. (2007). Learning to
rank: From pairwise approach to listwise approach. In Proceedings of
the 24th International Conference on Machine Learning, pages 129–136.
ACM.
Caro, F. and Gallien, J. (2012). Clearance Pricing Optimization for a Fast-

Fashion Retailer, volume 60.
Carpenter, G. S. and Shankar, V. (2013). Handbook of Marketing Strategy.

Elgar original reference. Elgar.
Catalina Marketing (2014). Catalina Category Marketing.
Chalasani, P. and Sriharsha, R. (2016). Monte Carlo Simulations in Ad

Lift Measurement Using Spark.
Chapelle, O. and Chang, Y. (2011). Yahoo! learning to rank challenge

overview. Journal of Machine Learning Research - Proceedings Track,
14:1–24.
Chickering, D. M. and Pearl, J. (1996). A Clinician’s Tool for Analyzing

Non-compliance.
Cossock, D. and Zhang, T. (2006). Subset ranking using regression.

In Proceedings of the 19th Annual Conference on Learning Theory, pages
605–619. Springer-Verlag.
Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal

Statistical Society, 34(2):187–220.
Cox, D. R. (1975). Partial likelihood. Biometrika, 62(2):269–276.
Crammer, K. and Singer, Y. (2001). Pranking with ranking. In Advances

in Neural Information Processing Systems 14, pages 641–647. MIT Press.
Cremonesi, P., Koren, Y., and Turrin, R. (2010). Performance of recom-

mender algorithms on top-n recommendation tasks. In Proceedings
of the Fourth ACM Conference on Recommender Systems, pages 39–46.
ACM.
Crocker, C., Kulick, A., and Ram, B. (2012). Real user monitoring at
walmart.
Cuellar, S. S. and Brunamonti, M. (2013). Retail Channel Price Discrimi-

nation, volume 21.
Dalessandro, B., Perlich, C., Hook, R., Stitelman, O., Raeder, T., and
Provost, F. (2012a). Bid Optimizing and Inventory Scoring in Targeted
Online Advertising.
484 bibliography
Dalessandro, B., Perlich, C., Stitelman, O., and Provost, F. (2012b).

Causally Motivated Attribution for Online Advertising.
Debreu, G. (1960). Review of R. D. Luce, Individual Choice Behavior: A

Theoretical Analysis.
Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer, T. K., and Harsh-
man, R. (1990). Indexing by latent semantic analysis. Journal of the
American Society for Information Science, 41(6).
Devooght, R., Kourtellis, N., and Mantrach, A. (2015). Dynamic matrix

factorization with priors on unknown values. In Proceedings of the
21th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, pages 189–198. ACM.
Ding, Y. and Li, X. (2005). Time weight collaborative filtering. In Pro-

ceedings of the 14th ACM International Conference on Information and
Knowledge Management, CIKM ’05, pages 485–492, New York, NY,
USA. ACM.
Duhigg, C. (2012). How Companies Learn Your Secrets.
Efron, B. (1977). The efficiency of cox’s likelihood function for censored

data. Journal of the American Statistical Association, 72(359):557–565.
Ekstrand, M. D., Riedl, J. T., and Konstan, J. A. (2011). Collaborative

filtering recommender systems. Found. Trends Hum.-Comput. Interact.,
4(2):81–173.
Falk, K. (2017). Practical Recommender Systems. Manning Publications

Company.
Ferreira, K. J., Lee, B. H. A., and Simchi-Levi, D. (2016). Analytics for an

Online Retailer: Demand Forecasting and Price Optimization, volume 18.
Freund, Y., Iyer, R., Schapire, R. E., and Singer, Y. (2003). An efficient
boosting algorithm for combining preferences. J. Mach. Learn. Res.,
4:933–969.
Frigyik, B., Kapila, A., and Maya, G. (2010). Introduction to the dirich-
let distribution and related processes. Technical report, University
of Washington.
Funk, S. (2016). Netflix update: Try this at home.
Ge, M., Delgado-Battenfeld, C., and Jannach, D. (2010). Beyond accu-

racy: Evaluating recommender systems by coverage and serendipity.
In Proceedings of the Fourth ACM Conference on Recommender Systems,
pages 257–260. ACM.
bibliography 485
Geman, S. and Geman, D. (1984). Stochastic Relaxation, Gibbs Distribu-

tions, and the Bayesian Restoration of Images, volume 6. IEEE Computer
Society, Washington, DC, USA.
Geng, L. and Hamilton, H. J. (2006). Interestingness measures for data

mining: A survey. ACM Comput. Surv., 38(3).
Ghani, R. and Fano, A. (2002). Building recommender systems using a

knowledge base of product semantics. In In 2nd International Confer-
ence on Adaptive Hypermedia and Adaptive Web Based Systems, Malaga.
Ghani, R., Probst, K., Liu, Y., Krema, M., and Fano, A. (2006). Text min-
ing for product attribute extraction. SIGKDD Explor. Newsl., pages
41–48.
Giunchiglia, F., Kharkevich, U., and Zaihrayeu, I. (2009). Concept

search. In Proceedings of the 6th European Semantic Web Conference on
The Semantic Web: Research and Applications, ESWC 2009 Heraklion.
Springer-Verlag.
Goldberg, D., Nichols, D., Oki, B. M., and Terry, D. (1992). Using col-
laborative filtering to weave an information tapestry. Commun. ACM,
35(12):61–70.
Google Inc. (2011). The Arrival of Real-Time Bidding.
Gormley, C. and Tong, Z. (2015). Elasticsearch: The Definitive Guide.

O’Reilly Media.
Green, P. E. and Srinivasan, V. (1978). Conjoint Analysis in Consumer

Research: Issues and Outlook, volume 5.
Grigsby, M. (2016). Advanced Customer Analytics: Targeting, Valuing,

Segmenting and Loyalty Techniques. Marketing Science Series. Kogan
Page, Limited.
Hall, R. E. (1967). Polynomial distributed lags.
Heckerman, D. and Shachter, R. D. (1995). Decision-Theoretic Founda-

tions for Causal Reasoning.
Herbrich, R., Graepel, T., and Obermayer, K. (2000). Large margin rank
boundaries for ordinal regression. In Advances in Large Margin Clas-
sifiers, pages 115–132. MIT Press.
Herlocker, J. L., Konstan, J. A., Borchers, A., and Riedl, J. (1999). An

algorithmic framework for performing collaborative filtering. In Pro-
ceedings of the 22Nd Annual International ACM SIGIR Conference on
486 bibliography
Research and Development in Information Retrieval, SIGIR ’99, pages

230–237, New York, NY, USA. ACM.
Herlocker, J. L., Konstan, J. A., Terveen, L. G., and Riedl, J. T. (2004).

Evaluating collaborative filtering recommender systems. pages 5–
53.
Hofmann, T. (1999). Probabilistic latent semantic indexing. In Pro-

ceedings of the 22Nd Annual International ACM SIGIR Conference on
Research and Development in Information Retrieval, SIGIR ’99, pages 50–
57, New York, NY, USA. ACM.
Hu, Y., Koren, Y., and Volinsky, C. (2008). Collaborative filtering for
implicit feedback datasets. In Proceedings of the 2008 Eighth IEEE In-
ternational Conference on Data Mining, ICDM ’08, pages 263–272. IEEE
Computer Society.
Hughes, S. (2015). Implementing conceptual search in Solr using LSA

and Word2Vec.
Jack, K., Ingold, E., and Hristakeva, M. (2016). Mendeley suggest archi-
tecture.
Jahrer, M., Töscher, A., and Legenstein, R. (2010). Combining predic-

tions for accurate recommender systems. In Proceedings of the 16th
ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, KDD ’10, pages 693–702, New York, NY, USA. ACM.
Jambor, T. and Wang, J. (2010). Optimizing multiple objectives in col-

laborative filtering. In Proceedings of the Fourth ACM Conference on
Recommender Systems, RecSys ’10, pages 55–62, New York, NY, USA.
ACM.
Järvelin, K. and Kekäläinen, J. (2000). IR evaluation methods for re-

trieving highly relevant documents. In Proceedings of the 23rd annual
international ACM SIGIR conference on Research and development in in-
formation retrieval.
Jo, B. (2002). Statistical Power in Randomized Intervention Studies With

Noncompliance.
Johnson, J., Tellis, G. J., and Ip, E. H. (2013). To Whom, When, and How
Much to Discount? A Constrained Optimization of Customized Temporal
Discounts, volume 89. Elsevier.
Ju, C., Bao, F., Xu, C., and Fu, X. (2015). A novel method of interest-
ingness measures for association rules mining based on profit. In
Discrete Dynamics in Nature and Society.
bibliography 487
Kahneman, D. and Tversky, A. (1979). Prospect theory: An analysis of

decisions under risk.
Kamotsky, D. and Vargas, M. (2014). System and method for perform-

ing a pattern matching search. US Patent App. 14/292,018.
Kane, K., Lo, V. S., and Zheng, J. X. (2014). Mining for the truly respon-
sive customers and prospects using true-lift modeling: Comparison
of new and existing methods. Journal of Marketing Analytics, 2(4):218–
238.
Kaplan, E. L. and Meier, P. (1958). Nonparametric estimation from

incomplete observations. Journal of the American Statistical Association,
53(282):457–481.
Khan, R. J. and Jain, D. C. (2005). An Empirical Analysis of Price Discrim-

ination Mechanisms and Retailer Profitability, volume 42.
Khludnev, M. (2013). Concept search for eCommerce with Solr.
Kleinberg, J., Papadimitriou, C., and Raghavan, P. (1998). A Microeco-

nomic View of Data Mining, volume 2. Kluwer Academic Publishers,
Hingham, MA, USA.
Kohavi, R. and Longbotham, R. (2007). Online experiments: Lessons

learned.
Kök, A. G. and Fisher, M. L. (2007). Demand Estimation and Assort-

ment Optimization Under Substitution: Methodology and Application, vol-
ume 55. INFORMS.
Koren, Y. (2007). How useful is a lower rmse?
Koren, Y. (2008). Factorization meets the neighborhood: A multi-

faceted collaborative filtering model. In Proceedings of the 14th ACM
SIGKDD International Conference on Knowledge Discovery and Data Min-
ing, pages 426–434, New York, NY, USA. ACM.
Koren, Y. (2009). The bellkor solution to the netflix grand prize.
Koren, Y. and Bell, R. M. (2011). Advances in collaborative filtering.

In Ricci, F., Rokach, L., Shapira, B., and Kantor, P. B., editors, Recom-
mender Systems Handbook, pages 145–186. Springer.
Koyck, L. M. (1954). Distributed lags and investment analysis.
Lauterborn, B. (1990). New Marketing Litany: Four Ps Passe: C-Words Take

Over.
488 bibliography
Lee, D. D. and Seung, H. S. (2001). Algorithms for non-negative matrix

factorization. In Leen, T. K., Dietterich, T. G., and Tresp, V., editors,
Advances in Neural Information Processing Systems 13, pages 556–562.
MIT Press.
Li, P., Burges, C. J. C., and Wu, Q. (2007). Mcrank: Learning to rank
using multiple classification and gradient boosting. In NIPS, pages
897–904. Curran Associates, Inc.
Linden, G., Smith, B., and York, J. (2003). Amazon.com recommenda-

tions: Item-to-item collaborative filtering. IEEE Internet Computing,
7(1):76–80.
Littlewood, K. (1972). Forecasting and Control of Passenger Bookings.
Liu, T.-Y. (2009). Learning to rank for information retrieval. Foundations

and Trends in Information Retrieval, (3):225–331.
Liu, T.-Y. and Qin, T. (2010). Microsoft learning to rank datasets.
Lo, V. S. (2002). The true lift model: a novel data mining approach to re-
sponse modeling in database marketing. ACM SIGKDD Explorations
Newsletter, 4(2):78–86.
Manning, C. D., Raghavan, P., and Schütze, H. (2008). Introduction to

Information Retrieval. Cambridge University Press, New York, NY,
USA.
Marn, M. and Roseillo, R. (1992). Managing price, gaining profit.
McCarthy, E. J. (1960). Basic Marketing. A Managerial Approach.
Melville, P., Mooney, R. J., and Nagarajan, R. (2002). Content-boosted

collaborative filtering for improved recommendations. In Eighteenth
National Conference on Artificial Intelligence, pages 187–192. American
Association for Artificial Intelligence.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013a). Efficient esti-
mation of word representations in vector space.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., and Dean, J. (2013b).
Distributed representations of words and phrases and their compo-
sitionality. In Advances in Neural Information Processing Systems 26.
Curran Associates, Inc.
Miyahara, K. and Pazzani, M. J. (2000). Collaborative filtering with

the simple bayesian classifier. In Proceedings of the 6th Pacific Rim
International Conference on Artificial Intelligence, PRICAI’00, pages 679–
689. Springer-Verlag.
bibliography 489
Mobasher, B., Dai, H., Luo, T., and Nakagawa, M. (2001). Effective
personalization based on association rule discovery from web usage
data. In Proceedings of the 3rd International Workshop on Web Informa-
tion and Data Management, WIDM ’01, pages 9–15, New York, NY,
USA. ACM.
Mooney, R. J. and Roy, L. (1999). Content-based book recommending

using learning for text categorization. In Proceedings of the SIGIR-99
Workshop on Recommender Systems: Algorithms and Evaluation.
Murphy, K. P. (2012). Machine Learning: A Probabilistic Perspective. The

MIT Press.
Musalem, A., Olivares, M., Bradlow, E. T., Terwiesch, C., and Corsten,
D. (2010). Structural estimation of the effect of out-of-stocks. Man-
agement Science, 56(7):1180–1197.
Oi, W. Y. (1971). A Disneyland Dilemma: Two-Part Tariffs for a Mickey

Mouse Monopoly.
Oneata, D. (1999). Probabilistic latent semantic analysis. In Proceedings

of the Fifteenth conference on Uncertainty, pages 1–7.
Park, L. A. and Ramamohanarao, K. (2009). Efficient storage and re-

trieval of probabilistic latent semantic information for information
retrieval. The VLDB Journal, 18(1):141–155.
Pashigian, P. (1987). Demand Uncertainty and Sales: A Study of Fashion

and Markdown Pricing.
Pennington, J., Socher, R., and Manning, C. D. (2014). Glove: Global

vectors for word representation. In Conference on Empirical Methods
on Natural Language Processing, volume 14, pages 1532–1543.
Perlich, C., Dalessandro, B., Raeder, T., Stitelman, O., and Provost,
F. (2013). Machine Learning for Targeted Display Advertising: Transfer
Learning in Action.
Peter, G. and Eugene, S. (2015). Deep data at macys: Searching hierar-

chical documents for ecommerce merchandising.
Pfeifer, P. and Carraway, R. (2000). Modeling Customer Relationships as

Markov Chains.
Porter, M. F. (1980). An algorithm for suffix stripping. Program,

14(3):130–137.
490 bibliography
Radcliffe, N. J. and Simpson, R. (2007). Identifying who can be saved

and who will be driven away by retention activity.
Radcliffe, N. J. and Surry, P. (1999). Differential response analysis: Mod-

eling true response by isolating the effect of a single action. Credit
Scoring and Credit Control VI. Edinburgh, Scotland.
Radcliffe, N. J. and Surry, P. D. (2011). Real-world uplift modelling

with significance-based uplift trees. White Paper TR-2011-1, Stochastic
Solutions.
Radlinski, F. and Joachims, T. (2005). Query chains: Learning to rank

from implicit feedback. In Proceedings of the Eleventh ACM SIGKDD
International Conference on Knowledge Discovery in Data Mining, pages
239–248. ACM.
Rodriguez, M., Posse, C., and Zhang, E. (2012). Multiple objective op-
timization in recommender systems. In Proceedings of the Sixth ACM
Conference on Recommender Systems, RecSys ’12, pages 11–18, New
York, NY, USA. ACM.
Rohde, D. L. T., Gonnerman, L. M., and Plaut, D. C. (2006). An im-

proved model of semantic similarity based on lexical co-occurence.
Communications of the ACM.
Rong, X. (2014a). wevi: Word embedding visual inspector.
Rong, X. (2014b). Word2Vec parameter learning explained.
Rubin, D. (1974). Estimating causal effects of treatments in randomized and

nonrandomized studies.
Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2001). Item-based
collaborative filtering recommendation algorithms. In Proceedings of
the 10th International Conference on World Wide Web, pages 285–295.
ACM.
Shao, X. and Li, L. (2011). Data-driven Multi-touch Attribution Models.
Sill, J., Takacs, G., Mackey, L. W., and Lin, D. (2009). Feature-weighted
linear stacking.
Smith, B., Leimkuhler, J., and Darrow, R. (1992). Yield Management in

American Airlines.
Smith, T. (2012). Pricing Strategy. South-Western Cengage Learning.
Spangher, A. (2015). Building the next new york times recommenda-

tion engine.
bibliography 491
Su, X. and Khoshgoftaar, T. M. (2006). Collaborative filtering for multi-

class data using belief nets algorithms. In Proceedings of the 18th IEEE
International Conference on Tools with Artificial Intelligence, ICTAI ’06,
pages 497–504, Washington, DC, USA. IEEE Computer Society.
Su, X. and Khoshgoftaar, T. M. (2009). A survey of collaborative filter-

ing techniques. Adv. in Artif. Intell., pages 4:2–4:2.
Su, X., Khoshgoftaar, T. M., Zhu, X., and Greiner, R. (2008). Imputation-
boosted collaborative filtering using machine learning classifiers. In
Proceedings of the 2008 ACM Symposium on Applied Computing, SAC
’08, pages 949–950, New York, NY, USA. ACM.
Talluri, K. and Van Ryzin, G. (2004). The Theory and Practice of Revenue
Management. International Series in Operations Research & Manage-
ment Science. Springer.
Telgarsky, M. (2013). Dirichlet draws are sparse with high probability.
Terry, D. B. (1993). A tour through tapestry. In Proceedings of the Con-

ference on Organizational Computing Systems, COCS ’93, pages 21–30,
New York, NY, USA. ACM.
Töscher, A., Jahrer, M., and Bell, R. M. (2009). The bigchaos solution to
the netflix grand prize.
Train, K. (2003). Discrete Choice Methods with Simulation. SUNY-Oswego,

Department of Economics.
Turnbull, D. and Berryman, J. (2016). Relevant Search. With applications

for Solr and Elasticsearch. Manning Publications.
Vasigh, B., Tacker, T., and Fleming, M. (2013). Introduction to Air Trans-
port Economics: From Theory to Applications.
Vulcano, G. J., van Ryzin, G. J., and Ratliff, R. (2012). Estimating pri-
mary demand for substitutable products from sales transaction data.
Operations Research, 60(2):313–334.
Walker, R. (2009). The Song Decoders. The New York Times Magazine.
Wierenga, B. (2010). The interface of marketing and operations re-

search. In Kroon, L., Zuidwijk, R., and Li, T., editors, Liber Amicorum
in Memoriam Jo van Nunen.
Winston, W. L. (2014). Marketing Analytics: Data-Driven Techniques with

Microsoft Excel. Wiley Publishing, 1st edition.
492 bibliography
Xia, Z., Dong, Y., and Xing, G. (2006). Support vector machines for
collaborative filtering. In Proceedings of the 44th Annual Southeast Re-
gional Conference, ACM-SE 44, pages 169–174, New York, NY, USA.
ACM.
Xu, J. and Li, H. (2007). Adarank: A boosting algorithm for information

retrieval. In Proceedings of the 30th Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval, pages
391–398, New York, NY, USA. ACM.
Zaki, M. J. and Meira, W. (2014). Data Mining and Analysis: Fundamental

Concepts and Algorithms.
Zhang, A., Goyal, A., Kong, W., Deng, H., Dong, A., Chang, Y., Gunter,
C. A., and Han, J. (2015). adaqac: Adaptive query auto-completion
via implicit negative feedback. In Proceedings of the 38th International
ACM SIGIR Conference on Research and Development in Information Re-
trieval, pages 143–152. ACM.
Zhang, H. (2004). The optimality of naive bayes. In Barr, V. and Markov,

Z., editors, Proceedings of the Seventeenth International Florida Artificial
Intelligence Research Society Conference (FLAIRS 2004). AAAI Press.
Zhang, S., Wang, W., Ford, J., and Makedon, F. (1996). Learning from
incomplete ratings using non-negative matrix factorization. In In
Proc. of the 6th SIAM Conference on Data Mining, pages 549–553.

Algorithmic Marketing Ai For Marketing Operations r1 3g

Uploaded by

Copyright:

Available Formats

Algorithmic Marketing Ai For Marketing Operations r1 3g

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Algorithmic Marketing Ai For Marketing Operations r1 3g

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What are some of the endorsements for the book?

What are some of the endorsements for the book?

I LYA K AT S O V

Visit the book’s website at https://algorithmicweb.wordpress.com

Copyright © 2018 Ilya Katsov

—Ali Bouhouch, CTO, Sephora Americas

“This book is a live portrait of digital transformation in marketing. It shows

—Andrey Sebrant, Director of Strategic Marketing, Yandex

“This books delivers a complete end-to-end blueprint on how to fully digitize

—Victoria Livschitz, founder and CTO, Grid Dynamics

“This book provides a much-needed collection of recipes for marketing prac-

—Kira Makagon, Chief Innovation Officer, RingCentral;

“While virtually every business manager today grasps the conceptual

2.6.3 Auction Theory . . . . . . . . . . . . . . . . . . . . . . 68

3.8.4 Multi-Touch Attribution . . . . . . . . . . . . . . . . . 154

4.5.7 Word2Vec Model . . . . . . . . . . . . . . . . . . . . . 240

5.8.3.3 Advanced Latent Factor Models . . . . . . . . . . 342

6.7 Price Optimization . . . . . . . . . . . . . . . . . . . . . . . 424

In 1888, Vincent van Gogh, a little-known Dutch artist at that time,

science, and software engineering to realize the potential of the digi-

1.1 the subject of algorithmic marketing

One of the traditional definitions of marketing describes it as the activ-

be broken down into several streams that are typically described as

• Promotion – all methods of communication between the com-

• Price – pricing strategies, including posted prices, price dis-

• Place – historically, this refers to the process of making a prod-

1.2 the definition of algorithmic marketing

We define algorithmic marketing as a marketing process that is auto-

Figure 1.1: A conceptual view of the algorithmic marketing ecosystem.

Although it would be ideal for a programmatic system to be perfectly

applications, it is more appropriate to view programmatic systems as

1.3 historical backgrounds and context

There is no sharp boundary between algorithmic and non-algorithmic

1.3.1 Online Advertising: Services and Exchanges

The history of internet advertising can be traced back to May 3, 1978,

the measurements and adjustments could be done dynamically. How-

• Advertisers paid per click, not per impression.

revenue  bid price click-through rate

This was a programmatic self-learning technique that optimized the

Reflecting on the history of RTB, we can conclude that one of the

• Data services that collect and store information about consumers,

1.3.2 Airlines: Revenue Management

Figure 1.2: The marketplace model for data and services.

Figure 1.3: Revenue management as a counterpart of supply-chain manage-

1.3.3 Marketing Science

developed using advanced probabilistic and optimization techniques.

1.4 programmatic services

between them as well. Let us briefly review these relationships and

• The search and recommendation services solve the problem of

• The goal of pricing and assortment services is to determine and

Figure 1.4: Programmatic services.

• The raw input data often has to be transformed into well-defined

contexts, plays a critically important role in programmatic sys-

Figure 1.5: A high-level design framework for programmatic services.

• The most important step in the programmatic pipeline is to eval-

• The signals generated by models carry information about the

• A programmatic service communicates with the outside world

revenue bid price click-through rate

sopt argmax G ps, Dq (2.1)

Lpθq pmodel py | X, θq (2.16)

LLpθq Ex,ypdata r log pmodel py | x, θq s (2.19)

argmin KLppdata , pmodel q argmin LLpθq (2.21)

2σ12 SSEpwq n2 logp2πσ2 q

LLpwq py XwqT py Xwq

wML pXT Xq1 XT y (2.30)

Prpy 0 | xq Prpy 1 | xq (2.31)

Prpy 0 | xq 1 Prpy 1 | xq (2.36)

log Prpy 0 | xq wT0 x log Z