Abstract: With the rapid development of the Internet,the emergence of massive data makes recommender system become a research hotspot in the field of computer science.Variational autoencoders(VAE) have been successfully applied to the design of collaborative filtering methods and achieved excellent recommendation results. However,there are some defects in the previous VAE-based models,such as the problems of prior constraint and the “posterior collapse”,which essentially reduce their recommendation performance.To address this issue while enabling the latent variable model more suitable for the recommendation task,a novel collaborative filtering recommendation model based on latent vector quantization is proposed in this paper.By encoding the discrete vectors instead of directly sampling from the distribution of latent variables,our method can learn discrete representations that are consistent with the observed data,which greatly improves the capability of latent vector encoding and the learning ability of the model.Extensive evaluations conducted on three benchmark datasets demonstrate the effectiveness of the proposed model.Our model can significantly improve the recommendation performance compared with existing state-of-the-art methods while learning more expressive latent representations.

Key words: Recommender system, Collaborative filtering, Vector quantization coding, Variational autoencoder