Content-enhanced Bayesian Personalized Ranking

Abstract

With the popularization of Knowledge Payment Products (KPP), more accurate recommendations are in great need to alleviate information overload of users. Bayesian Personalized Ranking (BPR) is one of the most representative pairwise ranking methods for recommendation, the performances of which greatly depend on the selection of negative feedback. However, traditional negative samplers may suffer from bias and noises. Therefore, in this paper, we focus on improving negative sampling strategy of BPR by incorporating side information of the knowledge products. We locate negative samples by calculating the cosine similarity among items by the textual features of KPP, under the assumption that a user shall have similar perceptions on items with similar content. We union our sampler strategy and the original one with different ratios. Compared to the original BPR that applies a uniform sampler on all the products, the join of our content-based sampler enhances BPR with a relative improvement over 4% on the ZhiHu Live dataset, which demonstrates the effectiveness of considering side information when capturing users' preferences on different items.