Learning multiple models for exploiting predictive heterogeneity in recommender systems

Collaborative filtering approaches exploit information about historical affinities or ratings to predict unknown affinities between sets of "users" and "items" and make recommendations. However a model that also incorporates heterogeneous sources of information that may be available on the users and/or items can become a much more effective recommender, in terms of both increased relevance of the predictions as well as explainability of the results. In this paper, we propose a Bayesian approach that exploits not only such "side-information", but also a different kind of heterogeneity that captures the variations in the mapping from user/item attributes to the affinities of interest. Such predictive heterogeneity is likely to occur in large recommender systems that involve a diverse set of users, and can be mitigated by using multiple localized predictive models rather than a single global one that covers all user-item pairs. The scope or coverage of each local model is determined simultaneously with the model parameters. The proposed approach can incorporate different types of inputs to predict the preferences of diverse users and items. We compare it against well-known alternative approaches and analyze the results in terms of both accuracy and interpretability.