SMSBPR: A symmetric multi-pairwise preferences and similarity based BPR method for recommendation with implicit feedback

The Bayesian Personalized Ranking (BPR) method is a classical pairwise preference learning algorithm designed for implicit feedback. However, it assumes uniform user preferences for all non-interacted items, overlooking the diversity of user preferences. To address this, the Multi-pairwise Item preferences and Similarity based Bayesian Personalized Ranking (MISBPR) method (Zeng et al., 2023) further subdivides user preferences for items and proposes a multiple pairwise item preference assumption, which learns the diversity of user preferences for items. However, users and items are two entities that naturally appear in pairs in recommendation systems, solely considering preference relationships among items is insufficient. Therefore, we classify users into different preference levels for each given item, thereby capturing more granular user-item interactions. Based on this, we propose a Symmetric Multi-pairwise preference and Similarity based BPR (SMSBPR) method. SMSBPR further subdivides the preferences of users and items through an improved similarity computation method. By integrating our refined multiple pairwise preference assumption, SMSBPR delicately learns the user-item preference relationships and captures more preference diversity. Additionally, we develop an efficient learning algorithm based on stochastic gradient descent to optimize the proposed method. Empirical studies on six real-world datasets demonstrate that SMSBPR generally surpasses eleven state-of-the-art methods in terms of six evaluation metrics, highlighting its superiority in learning greater user/item preference diversity. The source code for implementing our method can be accessed publicly at: https://github.com/wjy/smsbpr.