RankMBPR: Rank-Aware Mutual Bayesian Personalized Ranking for Item Recommendationl

Previous works indicated that pairwise methods are state-of-the-art approaches to fit users' taste from implicit feedback. In this paper, we argue that constructing item pairwise samples for a fixed user is insufficient, because taste differences between two users with respect to a same item can not be explicitly distinguished. Moreover, the rank position of positive items are not used as a metric to measure the learning magnitude in the next step. Therefore, we firstly define a confidence function to dynamically control the learning step-size for updating model parameters. Sequently, we introduce a generic way to construct mutual pairwise loss from both users' and items' perspective. Instead of user-oriented pairwise sampling strategy alone, we incorporate item pairwise samples into a popular pairwise learning framework, bayesian personalized ranking (BPR), and propose mutual bayesian personalized ranking (MBPR) method. In addition, a rank-aware adaptively sampling strategy is proposed to come up with the final approach, called RankMBPR. Empirical studies are carried out on four real-world datasets, and experimental results in several metrics demonstrate the efficiency and effectiveness of our proposed method, comparing with other baseline algorithms.