SMIGNN: social recommendation with multi-intention knowledge distillation based on graph neural network

Social recommendation based on user preference aims to make use of user interaction information and mine user's fine-grained preferences for item prediction. However, existing methods have not explored the differences in intentions between various user interactions, and they do not sufficiently model the user social graph and user item graph. In order to solve the above problems, this paper proposes a novel social recommendation model. It utilizes graph neural network to construct separate models for user social model and user item model, capturing multiple intents that drive user preferences. We use node attention and intention attention to calculate feature weights separately to obtain rich features of users and items. To increase the information dissemination between models and improve their overall predictive ability, we introduce knowledge distillation technology. The specific design is to combine the user's dual graph as the teacher model and transfer knowledge to the student model of a single user graph separately. Subsequently, the model is optimized for item recommendation prediction by calculating both the basic loss and distillation loss using the cross-entropy function. Experiments are conducted on Yelp and Ciao datasets, and the results achieve good predictive performance, which has demonstrated the effectiveness of the proposed method.