MRVAE: Variational Autoencoder with Multiple Relationships for Collaborative Filtering

Abstract

Variational Autoencoder (VAE)-based collaborative filtering (VAE-based CF) methods have shown their effectiveness in top-N recommendation. Mult-VAE is one of them that achieves state-of-the-art performance. Multinomial likelihood and additional hyperparameter $\beta$ on the KL divergence term controlling the strength of regularization make Mult-VAE a strong baseline. However, Mult-VAE uses non-linear MLPs as its encoder and decoder, which will boost the performance on the dense datasets but degrade the performance on the sparse datasets in our experiments. While recent studies shed light on the non-linearity for modeling the relationships between users and items, they ignore the importance of linearity between users and items, especially on the sparse datasets. To bridge the gap and consider both the linearity and non-linearity user-item relationships, we design a hybrid encoder that incorporates both linearity and non-linearity, and use a linear decoder for VAE-based CF, which can achieve competitive performance on both sparse and dense datasets. Moreover, most VAE-based CF methods only consider the relationships between users and items but ignore the relationships between items for improving the performance in collaborative filtering. To overcome this limitation, we try to incorporate item-item relationships into VAE-based CF with the help of cosine similarity between items. Unifying these relationships into VAE-based CF forms our proposed method, Variational Autoencoder with Multiple Relationships (MRVAE) for collaborative filtering. Extensive experiments on several dense and sparse datasets show the effectiveness of MRVAE.

This work is supported by the National Natural Science Foundation of China (U1911203, 61902439, 61902438, U1811264, U1811262), Guangdong Basic and Applied Basic Research Foundation (2021A1515011902, 2019A1515011159, 2019A1515011704), National Science Foundation for Post-Doctoral Scientists of China underGrant (2019M663237), Macao Young Scholars Program (UMMTP2020-MYSP-016), the Key-Area Research and Development Program of Guangdong Province (2020B0101100001).