Augmenting Probabilistic Matrix Factorization Models for Rare Users
Cody Severinski · TSpace (University of Toronto) · 2016
Recommender systems are used for user preference prediction in a variety of contexts. Most commonly known for movie suggestion from the Netflix competition, these systems have evolved to cover generic product recommendation, friend suggestion, and even online dating. Matrix Factorization is a common model employed for several reasons. Among them, they scale well, are easily learned, and can be adapted to different contexts. Many extensions of the baseline Probabilistic Matrix Factorization model have been proposed in the literature, and as expected, all perform better than the baseline with reported test results. We review several of these extensions, notably: constraints based on similar rating patterns among users, allowing for nonconstant variance / precision in the model, introducing personal information on the users as constraints, and including user networks in the model. These models are extended to the Bayesian framework where necessary. We illustrate how these extensions perform overall, and for sets of users defined by different number of ratings at training time. In particular, we highlight the benefit of many of these extensions for infrequent users (those with few or no ratings in the system). This is particularly important as these users are the most common in the recommendation framework. In the case of user networks, we additionally study the robustness of the model in the presence of random links. This reflects the true state of user networks in applications such as Facebook, where social ties may not convey similar taste in preferences. In addition, we provide the first direct comparison of the performance of the models learned from Gibbs sampling and variational inference. Limitations of the variational algorithm are outlined for multiple models, with proposals given for alleviating overfitting.