Federated Learning-Based Social Recommendation with Social Denoising

In social recommendation (SR), users' social connections and behavioral data often contain sensitive information, which might lead to a deep concern about privacy leakage. In the case where raw data is absent, federated learnings social recommendation (FLSR) provide an effective framework that can not only preserve a reliable performance but also ensure the security of privacy data through collaboratively training a global model. However, existing results on FLSR often directly employ the raw interaction data without considering the impact of noise or unreliable links in social network, which may significantly decrease the effectiveness of recommendation algorithms. To address such an issue, we propose a federated learning-based social de noising recommendation (FLSDR) method consisting of three modules: graph de noising, heterogeneous graph en-coder and privacy protector. Specifically, the graph de noising module incorporates a user preference consistency function and a deletion rate to remove low-quality social connections, achieving the goal of denoising. The heterogeneous graph encoder module ensures that user embed dings are updated locally, assigning different weights to different neighbors. The privacy protection module is designed to further safeguard privacy and prevent gradient leakage of user information. Experiments over two data sets shows that our model has achieved a 4.18% improvement in performance. Ablation studies demonstrate the necessities of each component.