Collaborative Multi-Auxiliary Information Variational Autoencoder for Recommender Systems

Abstract

Collaborative filtering is widely used in recommendation systems. Hybrid approach has been proposed since the collaborative-based method is susceptible to problems such as sparsity and cold start. Recently, the related methods have pointed out that it is very effective to alleviate the above problem by inferring the stochastic distribution of the latent variables for item's auxiliary information. Usually, item boasts more than one kind of auxiliary information. How do we infer the stochastic distribution of the latent variables with multiple auxiliary information? In this paper, we proposed a collaborative multi-auxiliary information autoencoder that can simultaneously consider multiple types of auxiliary information correspondingly. On the one hand, we can successfully accomplish the above issues via the improvement of variational autoencoder. On the other hand, we demonstrated the effectiveness of our method through experiments on real datasets.