K-Core Structure Feature Encoding-Based Enhanced Federated Graph Learning Framework

Abstract

Federated Graph Learning (FGL) demonstrates tremendous potential in distributed graph data analysis and modeling. The rapid growth of graph data and the increasing awareness of privacy protection make FGL research highly valuable. However, its development faces two critical challenges: the non-IID problem in heterogeneous graphs and low communication efficiency. This study proposes an Enhanced FGL framework based on K-core Structure Feature Encoding (FedKcore) to utilize various heterogeneous graphs efficiently. The nested chain structure containing rich information and linear encoding time make K-core structural attributes highly suitable for graph enhancement and aggregate sharing on edge devices. Client personalization capabilities are enhanced by combining original features with K-core attributes for local training. To improve convergence speed and overcome the non-IID challenge, we aggregate and share only the learnable parameters related to K-core attributes. Upon this, the introduced Circle Loss function optimizes feature space and boundaries, enhancing the performance of K-core attributes. Extensive experiments on heterogeneous graphs show that, compared to the state-of-the-art FedStar, FedKcore improves accuracy by over 1.3% and speeds up convergence by 1.3 times.