Hierarchical Constrained Variational Autoencoder for interaction-sparse recommendations

Abstract

Predicting potential user behaviors is of great importance in recommendation systems. Existing behavior prediction works mainly aim to construct the user behavior preference based on a large amount of user interaction data, but the prediction performance deteriorates when facing the users with a few interactions. A common and effective approach to alleviate the interaction-sparse problem is to incorporate the auxiliary information as much as possible, in which the hybrid (Variational Autoencoder) VAE method reports the optimal performance with the advantages of non-linear modeling and comprehensive integration. However, the hybrid VAE only focuses on the attribute data, which limits the richness and robustness of the user preference representation in interaction-sparse scenario. In this paper, we propose a novel hybrid VAE model, named Hierarchical Constrained Variational AutoEncoder Jointly Attributes and Neighbors (HiCVAE), to tackle the above limitations. Specifically, HiCVAE additionally introduces fine-grained neighbors’ behaviors as auxiliary information to enhance the richness and robustness of the user preference representation on the basis of the attribute data, by learning a hierarchical prior instead of a standard Gaussian prior. The prior on the attribute data is responsible for the community preference relevant to a group with the same attributes. Further, the conditional prior over the neighborhood data can capture the personalized information and provide further constraints for the new user preference representation. In addition, a Preference Consistency Loss is proposed to restrict the reconstructed interest to be consistent with the user’s real interest and eliminate the irrelevant information on the neighborhood data. Thorough and extensive experiments conducted on two real-world datasets demonstrate that the HiCVAE significantly outperforms the state-of-the-art methods in various interaction-sparse scenarios (sparse, highly sparse and extremely sparse).