Semi-disentangled non-negative matrix factorization for rating prediction

Abstract

Rating predictions have been extensively used in evaluation websites for recommendation tasks in recent past. Due to the single data category and the simple way of feature interaction, traditional rating prediction methods do not portray the actual scoring properly, and lack of interpretation at the same time, which problem is particularly acute in latent factor model. Thanks to the excellent interpretability and generalizability of disentangled representations, in this paper, we propose a Semi-Disentangled Non-negative Matrix Factorization (SDNMF) by splitting user- and item-latent embedding into two blocks with an improved regularizer: the whiten block and the blacked block. Specifically, the whiten block uses external prior knowledge instead of latent factors to improve the internal interpretability of SDNMF, whereas the blacked block consists of those latent factors different from the prior knowledge to promote more feature interaction and preserve algorithm performance, and the regularizer is defined by Kullback–Leibler divergence to characterize the relationship between whiten and black blocks. By disentangling user–item relations to obtain disentangled representation, we want to be able to on this basis, the non-negative matrix factorization for a certain degree of interpretation. Extensive experiments conducted on four real-world data sets (Amazon, Yelp, Yelp10, Dianping) indicate that the proposed SDNMF not only is the state-of-the-art rating prediction method compared with other existing methods, but also preserves the ability to be interpreted.