Sparsity-aware social recommendation via information graph bottleneck

With the widespread adoption of online social networks, social recommender systems have become a key paradigm for mitigating data sparsity and cold-start issues in traditional collaborative filtering. However, existing graph neural network (GNN) based recommendation models face two fundamental structural bottlenecks when dealing with real-world social data: social noise and over-smoothing. Invalid or weak social connections dilute preference signals, while representation homogenization induced by deep graph convolution undermines personalized recommendation. To address these challenges, this report proposes a new framework named SRIGB (Social Recommendation via Information Graph Bottleneck). SRIGB introduces the Information Bottleneck principle as a principled denoising mechanism by maximizing the mutual information between the denoised graph and interaction labels while minimizing redundant information with respect to the original noisy graph, thereby extracting task-relevant social structure more precisely. In addition, SRIGB incorporates a contrastive learning paradigm to counteract the smoothing effect of graph convolution, enabling the model to preserve high-order social influence while maintaining discriminative and independent user representations. This report presents the theoretical foundations, model architecture, and experimental design of SRIGB, aiming to provide both theoretical support and a practical technical pathway for building the next generation of robust social recommender systems.