Interest-aware influence diffusion model for social recommendation

With the rapid development of social networks, the application of social relationships in recommendation systems has attracted more and more attention. The social recommendation system alleviates the problem of data sparsity by utilizing the preferences of users' friends, so as to better model the user embedding. The traditional social recommendation model optimizes the user embedding through the first-order neighbors' preference information, but it failed to model the social influence diffusion process from the global social network. Although the recently proposed DiffNet solves this problem to some extent, it ignores the fact that high-order adjacent users who have no common interests in the social network can also participate in graph convolution, which will make user embedding affected by negative information in social influence diffusion, thus affecting the performance of the recommendation system. Therefore, we propose a model named IDiffNet based on DiffNet. In IDiffNet, a self-supervised subgraph generation module is designed to identify users with similar interests according to user features. Thereby, the user social graph is decomposed into several user social subgraphs, and then user embedding is optimized through interest propagation on the subgraphs. Consequently, the IDiffNet can avoid the interaction of users with different interests in the user social network. Finally, we conducted a number of comparative experiments on two public datasets, and the results show that IDiffNet has better performance than current mainstream social recommendation models.