Self-supervised variational autoencoder towards recommendation by nested contrastive learning

Recommendation methods predicting potential items for user has evolved from linear factor models to non-linear factor deep learning models. Deep generative model, especially variational autoencoder(VAE), has been used in a wide range of recommendation systems such as MultVAE and RecVAE. Despite effectiveness, we argue that they suffer from two limitations: (1)sparse and noisy user-item interactions will affect the performance of VAE-based recommendation models; and (2)incorporating simple priors,e.g.,isotropic Gaussian, in VAE couldn't extract personalized user preference,as user's preference may be highly complex. In this paper, we propose a Nested Self-supervised Variational Autoencoder (NSVAE) model for recommendation to enhance generalization and accuracy of VAE-based recommendation models. Besides using VAE for predicting user interests, NSVAE supplements supervised task of recommendation with nested self-supervised task, which consider both partial and entire preference. Nested self-supervised task is composed of inside and outside pretext tasks. Inside pretext task aligns the representations learned from different views, where views contain user partial preference, and outside pretext task discriminates entire preference from other user. Recommendation task and pretext tasks can be seamlessly integrated and enhance each other. Extensive experiment results on three real-world benchmarks validate the superiority of our NSVAE model to state-of-the-art VAE-based recommendation models.