Federated Non-IID Graph Learning Based on Graph Optimization

Federated graph learning (FGL) enables distributed GNN training and protects data privacy, where the data nonIID problem is one of the major challenges. To address this problem, recent research shares only structural information rather than model parameters. This is because parameters inevitably contain information about heterogeneous graph features, while some structural properties are universally shared. However, real-world graphs are often noisy or incomplete, making it difficult to obtain optimal or even accurate structural information. Furthermore, we find structural heterogeneity among graph datasets in both cross-datasets and cross-domain settings. To overcome these limitations, we propose FedGO, an FGL framework for non-IID graphs based on graph optimization and weighted averaging for structural information aggregation. In particular, to mitigate the effects of suboptimal graph structure, FedGO trains a graph optimizer and a GNN on the client side, which are used to optimize the graph structure and learn better node embeddings, respectively. The GNN is learned in a personalized way, while the graph optimizer is shared among clients as the optimizer contains graph structural information. In addition, considering the possible structural heterogeneity between different graphs, FedGO utilizes graph structure embeddings to compute similarities between clients and uses the similarities to perform weighted averaging for server-side aggregation. We conduct experiments in multiple non-IID settings to demonstrate the superiority of FedGO. Furthermore, we simulate possible attacks on real-world FGL scenarios to demonstrate the robustness of FedGO to adversarial graphs.