Optimizing Social Recommendations with GBSR: A Graph Bottleneck Approach for Reducing Noise

In social recommendation systems, redundant social relations often introduce noise, reducing the accuracy of recommendation tasks. This paper presents a novel Graph Bottleneck Social Recommendation (GBSR) framework designed to address this issue by refining social networks to minimize noise and enhance recommendation accuracy. The proposed GBSR framework integrates a Social Graph Refinement Module, Graph Convolutional Networks (GCN), and a Bottleneck Learning approach based on the Hilbert-Schmidt Independence Criterion (HSIC). Through comprehensive architectural design, data preprocessing, and extensive evaluation, our approach significantly outperforms existing models in various scenarios, demonstrating its effectiveness in reducing noise and improving recommendation accuracy. The GBSR framework showcases substantial improvements over baseline models, especially in key evaluation metrics such as Normalized Discounted Cumulative Gain (NDCG). This work contributes to the field by offering a robust solution to the challenges posed by noisy social networks in recommendation systems.