Automated Modeling of Influence Diversity with Graph Convolutional Network for Social Recommendation

Social recommendation leverages social connections among users to alleviate the data sparsity issue in conventional recommendation methods. However, existing social recommendation models fall short in capturing the diversity of social influence. Specifically, due to users individualized preferences, the influence of social connections varies among users. In other words, different users depend on their social connections to varying extents. To tackle above issues, we propose AutoSRec, a novel social recommendation model which automatedly modeling the diversity of social influence using Graph Neural Network. Firstly, we construct two graph perspectives: user-item interaction graph and user-user social graph. These graphs empower AutoSRec the ability of capturing both non-social and social preferences of users. Secondly, we design an automated graph aggregation mechanism to update embeddings. The tailored aggregation mechanism automatically estimates the importance of each graph perspective for each user, reflecting the degree to which each user is influenced by social connections. This mechanism can model influence diversity in a manner that is highly interpretable, less prone to randomness, differentiable and does not require manual efforts and expert knowledge. Finally, extensive experiments on public datasets demonstrate the effectiveness of AutoSRec over state-of-the-art methods.