A kernel-based approach to exploiting interaction-networks in heterogeneous information sources for improved recommender systems

Pairwise interaction networks capture inter-user dependencies (e.g. social networks) and inter-item dependencies (e.g item categories) that provide insight into user and item behavior. It is often assumed that such interaction information is informative for preference prediction. This may not be the case, as the some of the observed interactions may not be correlated with the preferences, and their use may negatively impact performance by introducing undesired noise. We propose an approach for weighting each interaction, such that we can determine the importance of each interaction to the preference prediction task. We model the preferences using kernel matrix factorization; where the kernels capture the weighted effects of the interactions. Our approach is validated on Last.fm and Movielens datasets; which include multiple sources of explicit and implicit inter-user and inter-item interactions. Our experiments suggest that learning the most important interactions can improve recommendation performance when compared to the standard matrix factorization approach.