Disentangled-feature and composite-prior VAE on social recommendation for new users

Social recommendation has been an effective approach to solve the new user recommendation problem based on user-item interactions and user-user social relations. Although lots of research has been done, it is still an emergent and challenging issue to predict the behaviors of new users without any historical interaction. Firstly, the previous methods fail to consider social structures and social semantics when looking for potential social neighbors for new users, resulting in inconsistent preferences of these neighbors. Secondly, existing methods employ deterministic modeling way to represent and aggregate neighbors, limiting the diversity and robustness of new user representations. Therefore, we present a novel new user preference uncertainty modeling framework, named Disentangled-feature and Composite-prior VAE(DC-VAE), to predict the behaviors of new users without any interaction. Concretely, a length-adaptive similarity metric considering the length of user behaviors and social relationships is designed for all users to choose more analogous neighbors, especially more effective for new users due to the metric incorporating the social structures and social semantics. Then the Neighbor-based Disentangled Features module is proposed to disentangle different types of neighbor characteristics and model more diversified new user representations. Next, unlike traditional Gaussian prior constraint, the Neighbor-based Composite prior module is proposed to fuse the priors of neighbors and obtain more expressive and robust new user representations. Finally, we theoretically prove the advantages of composite prior and disentangled features. Extensive experiments on three datasets demonstrate that our model DC-VAE is remarkably superior to other baselines.