Abstract
Local matrix factorization (LMF) methods have been shown to yield competitive performance in rating prediction. The main idea is to leverage the ensemble of submatrices for better low-rank approximation. However, the generated submatrices and recommendation results in the existing methods are usually hard to interpret. To address this issue, we adopt a probabilistic approach to enhance model interpretability of LMF methods by leveraging user reviews. In specific, we incorporate item-topics to construct meaningful “local clusters”, and further associate them with opinionated word-topics to explain the corresponding semantics and sentiments of users’ ratings. The proposed approach is a joint model which characterizes both ratings and review text. Extensive experiments on real-world datasets demonstrate the effectiveness of our proposed model compared with several state-of-art methods. More importantly, the produced results provide meaningful explanations to understand users’ ratings and sentiments.
