Towards heterogeneous federated graph learning via structural entropy and prototype aggregation

Abstract

In today's data-driven landscape, Federated Graph Learning (FGL) facilitates collaborative training between distributed data while providing robust privacy protections. However, FGL faces significant challenges in practical application: data heterogeneity owing to divergent node distributions and graph structures across clients, coupled with model heterogeneity caused by heterogeneous GNN architectures, substantially impedes the aggregation efficacy and generalization capabilities of global models. Existing FGL frameworks often overlook the unique impact of graph topology, inherent to graph data, between data and model heterogeneity. We propose an innovative framework called Structural Entropy Federated Graph Learning (SEFGL) that leverages structural entropy to simultaneously address data and model heterogeneity. At the client level, structural entropy-based virtual node generation and graph reconstruction techniques are applied to strengthen minority class node representations and optimize local graph topology while maintaining the original data distribution. At the server level, a prototype learning approach based on structural entropy aggregates data from clients with similar entropy characteristics. This enables each client to acquire a more diverse global representation, fostering the development of a personalized and robust prototype. Experiments conducted on three graph datasets demonstrate that the SEFGL framework achieves superior performance in terms of generalizability, efficiency, and effectiveness in high-heterogeneity scenarios.