Front. Comput. Sci., 2012, 6(2): 197–208 DOI 10.1007/s11704-012-2871-7 Two of a kind or the ratings game? Adaptive pairwise preferences and latent factor models Suhrid BALAKRISHNAN , Sumit CHOPRA AT&T Labs-Research, Florham Park, NJ 07932, USA c  Higher Education Press and Springer-Verlag Berlin Heidelberg 2012 Abstract Latent factor models have become a workhorse for a large number of recommender systems. While these sys- tems are built using ratings data, which is typically assumed static, the ability to incorporate different kinds of subsequent user feedback is an important asset. For instance, the user might want to provide additional information to the system in order to improve his personal recommendations. To this end, we examine a novel scheme for efficiently learning (or refining) user parameters from such feedback. We propose a scheme where users are presented with a sequence of pair- wise preference questions: “Do you prefer item A over B?” User parameters are updated based on their response, and subsequent questions are chosen adaptively after incorporat- ing the feedback. We operate in a Bayesian framework and the choice of questions is based on an information gain cri- terion. We validate the scheme on the Netflix movie ratings data set and a proprietary television viewership data set. A user study and automated experiments validate our findings. Keywords recommender systems, latent factor models, pairwise preferences, active learning 1 Introduction Recommender systems have been widely deployed in the past decade, and have been extremely successful in a large number of applications. This is reasonable given the rather limitless Received June 14, 2011; accepted October 18, 2011 E-mail: suhrid@research.att.com array of items a user has to choose from in more and more applications, and that recommender systems have all the while gained acceptance and exposure. Examples of appli- cation domains are very diverse, and include movies (Net- flix), products (Amazon.com), music (Pandora, Last.fm, and iTunes Genius), online videos (YouTube’s Recommended For You), social networking (Facebook’s Other People You May Know), short text messages [1], and books (What Should I Read Next). Many such systems operate in a collaborative filtering paradigm [2]. Essentially, collaborative filtering approaches try to extrapolate unobserved user-item preferences by using the information contained in user provided ratings on items. Of the approaches in the collaborative filtering paradigm, la- tent factor models are one of the most successful and popular [3–5]. The defining characteristic of latent factor models is the assumption that there are unobserved user-specific and item- specific latent (hidden) parameters whose combination deter- mines the preference the user will have for the item. These user and item-specific parameters are what need to be esti- mated using an observed sparse ratings matrix. In the case of the Netflix Prize contest 1) for instance, the items were about eighteen thousand movies, and the training ratings data set included from approximately 100 million numeric “star” rat- ings (from 1 = Hate it, to 5 = Love it). This ratings data is collected on about five hundred thousand users, resulting in a sparse user-movie matrix with about 1 percent of entries observed. Typically, latent factor models are estimated using such 1) http://www.netflixprize.com/