Research on personalized federated graph learning based on embedding similarity in recommendation systems

With the widespread adoption of recommendation systems, efficiently utilizing data distributed across different communities or edge devices has become a core challenge. Traditional centralized learning methods struggle to meet the demands of data privacy and transmission constraints, while Federated Learning (FL) offers a viable solution for distributed collaborative learning. However, existing FL methods (e.g., FedGCN) face limitations in handling data heterogeneity and achieving personalized models. These challenges are particularly pronounced in non-independent and identically distributed (non-IID) data scenarios, where knowledge collapse is prone to occur. This study proposes a personalized federated learning method based on FedGCN to optimize recommendation systems. By calculating the similarity of embedding vectors between clients, we dynamically adjust aggregation weights during the model aggregation process, thereby generating personalized models that adapt to heterogeneous data distributions. Experimental results demonstrate that, compared to existing methods, our approach achieves significant improvements in both convergence speed and prediction accuracy, particularly excelling in scenarios with significant data heterogeneity across communities.