Multi-channel Graph Neural Networks with Contrastive Learning for Social Recommendation

Social recommender systems based on graph neural networks can successfully model the social influence among users and effectively mitigate the issue of data sparsity. However, existing methods mostly focus on individual influence while ignoring group influence. The social influence exerted by different groups of users can vary, and the intensity can also vary. A user is often more susceptible to the influence of groups with preferences similar to his/her own. This paper proposes a social recommendation model called Multi-channel Graph neural network with Contrastive Learning (MGCL). MGCL classifies users into different groups according to their preferences, establishes separate social propagation channels for each group, and extracts the influence of different user groups to better represent a user. In addition, contrastive learning is utilized to improve the modeling of user preferences and social influence during model training. The superiority of the proposed MGCL over state-of-the-art baselines has been demonstrated through extensive experiments conducted on three benchmark datasets.