Subgraph Federated Learning with Global Graph Reconstruction

Missing cross-subgraph information is a unique problem in Subgraph Federated Learning (SFL) and severely affects the performance of the learned model. Existing cutting-edge methods typically allow clients to exchange data with all other clients to predict missing neighbor nodes. However, such client-to-client data exchanges are highly complex and lead to expensive communication overhead. In this paper, we propose FedGGR: subgraph federated learning with global graph reconstruction. FedGGR is a practical and effective framework. Specifically, the core idea behind it is to directly learn a global graph on the server by a graph structure learning module instead of predicting the missing neighbors on each client. Compared to existing methods, FedGGR does not require any data exchange among clients and achieves remarkable enhancements in model performance. The experimental results on four benchmark datasets show that the proposed method excels with other state-of-the-art methods. We release our source code at https://github.com/poipoipoi233/FedGGR .