Adaptive Feedback Graph-Enhanced Network ForSocial Recommendation

Social recommendation systems leverage social network data and graph models to enhance recommendation performance. Recent studies have highlighted the effectiveness of integrating social recommender systems with graph neural networks (GNNs). However, several critical challenges still remain: (i) most existing models tend to overlook previous mistakes, leading to repeated errors in future predictions, potentially causing local optima and preventing optimal performance; and (ii) basic sampling methods fail to capture the structural characteristics of graph data effectively, often leading to samples that are inadequate for model training needs. To address these challenges, we propose an A daptive F eedback G raph-Enhanced N etwork (AFGN) for social recommendation. While our approach is inspired by reinforcement learning, it differs by emphasizing the penalization of errors instead of relying on a reward function to reinforce correct behavior. This error-driven correction mechanism allows the model to learn from past mistakes and improve its predictive accuracy. Additionally, we introduce a novel and efficient, structure-aware graph-enhanced negative sampling method, which enhances the model's ability to capture the graph structure between users and items. Experiments on real-world datasets show that our method achieves significant improvements in recommendation accuracy over strong baselines.