SeBPR: Semantics Enhanced Bayesian Personalized Ranking with Comparable Item Pairs

Abstract

Learning user preferences from their implicit feedbacks is crucial to enable recommendations in various online applications. The Bayesian personalized ranking (BPR) with pairwise preference learning has been reported as one of the most promising algorithms for this problem. It follows a fundamental assumption that a user prefers interacted items to the unobserved items, which feedbacks have not happened. Then in the item pair generation, it either uniformly samples negative samples from unobserved items or incorporate some heuristics for an observed item. This paradigm would obtain noisy item pairs thus lead to a biased solution and a prolonged training period. In this paper, we attempt to enhance BPR with semantics. We first introduce semantically comparable item pairs and elaborate why we should adopt them into pairwise learning. We then present what semantics can be utilized and how can they be modeled. Furthermore, we propose a new method named Semantics enhanced Bayesian Personalized Ranking (SeBPR) to incorporate semantically comparable item pairs into the BPR learning framework. Finally, experimental results demonstrate that the proposed methods could reduce the noisy relationships for learning and thus improve the recommendation accuracy.