Attentive Autoencoder Matrix Factorization for Recommender Systems

摘要

In real applications, sparseness of user-item rating data significantly causes degrading in recommendation performance. There are several effective techniques which reduce prediction error by auxiliary information. However, these works focus on how to increase the ability of extracting features from a particular part rather than how to fuse features from different parts effectively. And the difference of Gaussian noise is ignored in modeling latent factors of both users and items. In this paper, to address these issues, we propose attentive autoencoder(AAE) to effectively extract features from several different information sources, which is integrated into improved probabilistic matrix factorization (PMF). The improved PMF consider the difference of Gaussian noise between both users and items. We perform extensive experiments on three real-world datasets, and the result shows that our AAEMF significantly outperforms the state-of-the-art models.