Adversarial Social Recommendations With Capturing Multi-Modal Views Of Social Friends

Abstract

Social recommendation aims to collect social relationship information to promote user preference learning, which is an important branch of the recommendation system. Although the existing works using graph neural network have been fruitful, they still face the following three challenges: (1) insufficient capture of user's fine-grained preferences at aspect level; (2) unable to capture the unique multi-modal distribution of user and friend aspect-level preferences; (3) lack the ability to weigh the importance of aspect information adaptively. To fill this gap, we propose a social recommendation model called AMMSR based on adversarial multi-modal viewpoint capturing. In order to achieve fine-grained modeling, AMMSR firstly projects candidate item into multiple latent aspect spaces to realize one-to-one query of user preferences at aspect level. Secondly, the adversarial variational bayes technique is introduced to capture the multi-modal preferences of users and their friends in latent aspects. Finally, two self-attention modules are set up on the user end and the commodity end to achieve the purpose of personalized information fusion. Extensive experiments on three public datasets show that AMMSR is generally superior to the existing methods, and several visual case studies verify that our model can effectively capture multi-modal viewpoints.