Statistical Methods for Recommender Systems

Deepak K Agarwal, Bee-Chung Chen

Published by
Cambridge University Press
ISBN 9781107036079
RRP £34.99
Reviewed by Patrick Hill BSc(Hons) MSc PhD CEng MBCS CITP
Score

10 out of 10

Recommender systems are a broad class of system whose function may be broadly described as identifying content that is most appropriate to users, based on a range of different criteria. On the web, recommender systems are ubiquitous, providing personalised content such as targeted advertising, news items, film suggestions and purchase recommendations. This book provides a comprehensive guide to state-of-the-art statistical techniques that are used to power recommender systems.

The book is divided into three sections. The first section considers the underlying principles of recommender systems. In this section, the authors describe approaches that may be used to generate feature vectors that represent users and content, and ways in which those vectors may be used to match content drawn from a pool to users in order to optimise click-through rate (CTR). The authors describe how the obvious solution to these problems, namely, by selecting the item with the best matching score, is often not applicable to web-based recommender systems, since the CTR is unknown at any point. Consequently, methods are required to produce good estimates of CTR. The book describes the use of explore-exploit techniques to help generate suitable estimates and also describes methods of model evaluation. 

The second section of the book describes the architecture of typical web-based recommender systems, and outlines the key issues faced by these systems, such as sparse user features, the requirement for off-line processing of recommendation models and subsequent updating of the live model, non-stationary CTR and the temporal nature of content such as news and product offers. The section provides an in-depth discussion and evaluation of various mathematical models that can be used to select content in order to maximise CTR.

The final section of the book describes advanced topics, including the use of context, such as what the user is currently reading, and the use of additional facets, such as post-read activities, in making optimal recommendations. While most of the book considers the optimisation of a single target, such as CTR, this section also describes the implementation of recommendation systems that aim to optimise multiple competing targets.

The text is authoritative and well written, with the authors drawing on their extensive experience of researching, implementing and evaluating real-world recommender systems. The book considers the underlying mathematics of the techniques it describes and, as such, is aimed at a readership with a strong background in statistics and cognate subjects. However, while readers without such a background are likely to find the mathematics somewhat challenging, the prose descriptions are highly readable and enable readers to understand the key principles and ideas which underpin the various approaches. This book should be of interest to those involved with recommender systems as well as to those with a broader interest in machine learning.

Further information: Cambridge University Press

March 2018