Recommender systems have become an important area of research with numerous applications on e-commerce. This chapter introduces a joint statistical model for user preferences and item features that can serve as the basis for a recommendation about recently published scientific papers. The model is constructed using ideas from the literature on Bayesian nonparametric mixture modeling. More specifically, user preferences are modeled using an Infinite Relational Model (IRM) in which both users and items are independently partitioned into homogeneous groups, while item features are modeled using a topic model, which also partitions items into groups with homogenous features. Information is shared across both components of the model through a common partition of items. Hence, the model is a hybrid system that combines ideas from collaborative and content-based filtering. The chapter discusses three different computational strategies, including a Markov chain Monte Carlo algorithm for full posterior inference, an iterated conditional maximization algorithm, and a mean-field variational algorithm for point estimation and prediction in large datasets where Markov chain Monte Carlo approaches might not be practical. The model is illustrated through simulation studies and by analyzing data from CiteULike.
Top1. Introduction
Recommendation systems aim to predict user ratings for items such as a books, movies or songs they have not yet considered, and use those ratings to suggest new items to the user. Recommendation systems are widely used in a variety of e-commerce applications, with Websites such as amazon.com, pandora.com and netflix.com being well known examples of platforms that incorporate them into their business models. Furthermore, the underlying models and algorithms behind recommendation systems have applications in a number of scientific areas as diverse as genomics and political science. For recent reviews, see Adomavicius and Tuzhilin (2005), Jannach et al. (2010) or Shapira (2011).
Collaborative filtering approaches to recommendation systems (Schafer et al., 2007; Herlocker et al., 2000; Su & Khoshgoftaar, 2009) leverage information about the behaviors, activities or preferences of multiple users to predict what others will like based on how similar their overall preferences are. The item-to-item filtering algorithm used by amazon.com (Sarwar et al., 2001; Linden et al., 2003), which makes suggestions based on what other items have been bought by users that acquired the item currently being browsed, is a well known example of a recommendation system based on collaborative filtering. Collaborative filtering methods might use clustering tools to find items that are close to those each other in terms of user preferences (Ungar & Foster, 1998; Linden et al., 2003; Hofmann, 2003), or might rely on matrix factorization algorithms such as singular value or non-negative matrix factorizations (Paterek, 2007; Koren, 2008), or probabilistic factor models (Salakhutdinov & Mnih, 2008; Agarwal & Chen, 2009). Breese et al. (1998) and Herlocker et al. (2004) present comparative evaluations of a number of collaborative filtering algorithms.
In contrast to collaborative filtering, content-based filtering methods (Pazzani & Billsus, 2007; Van Meteren & Van Someren, 2000; Mooney & Roy, 2000) use information about the characteristics of the item to make recommendations. This requires that each item be characterized in terms of a set of measurable features. The recommendation engine behind pandora.com, which uses song attributes extracted from the Music Genome Project (John, 2006) to make suggestions starting with a user-provided seed, is a well-known example of a content-based filter. Content-based recommenders rely on classifiers such as support vector machines (Cortes & Vapnik, 1995) or classification trees (Breiman, 1993, 1996, 2001).