Neighborhood Constraints Based Bayesian Personalized Ranking for Explainable Recommendation

Abstract

Model-based Collaborative Filtering (CF) methods such as Matrix Factorization (MF) have achieved promising ranking performance in recent years. However, these methods tend to be a black box that cant not provide any explanation for users. To obtain the trust of users and improve the transparency, recent research starts to focus on the explanation of recommendations. Explainable Bayesian Personalized Ranking (EBPR) leverages the relevant item to provide intuitive explanations for unexplainable model-based CF methods. It relies solely on feedback data and can be trained efficiently. However, EBPR ignores the neighborhood information of users and items in the latent space. In addition, the explainability of EBPR suffers from exposure bias. To address these issues, we propose a novel explainbale loss function and a corresponding Matrix Factorization-based model called Constrained Explainable Bayesian Personalized Ranking (CEBPR), which introduces neighborhood information in the latent space to enhance ranking performance and explainability. Furthermore, we analyze the impact of exposure bias on explainability and propose CEBPR+ to mitigate the bad effect. Finally, We conduct empirical experiments on three real-world datasets that demonstrate the advantages of our proposed methods in terms of accuracy and explainability.