Class-Imbalanced Graph Classification by using Federated Graph Learning

Federated graph learning (FGL) on graphs enables clients to train graph neural networks collaboratively without sharing private edge structures for graph tasks such as node classification and link prediction. However, existing methods are mainly proposed for tackling balanced graph datasets, neglecting the imbalance of graph classes and the heterogeneity of edge structures and node features. To address this issue, we propose a stable FGL framework (called SFGL) for class-imbalanced federated graph classification. Specifically, SFGL first decouples the feature and structure modules of the encoder, uploads only the structure encoder to the server for aggregation to learn cross-domain shared structural knowledge, and keeps the feature encoder on each client to capture domain-specific information, thereby mitigating structural and feature heterogeneity. Then, before federated aggregation, SFGL calibrates the gradients of local structure encoders to mitigate the influence of extreme gradients from majority-dominated clients. Finally, SFGL introduces a similarity-guided graph augmentation method, namely SGBS, to synthesize high-quality minority-class graphs locally based on graph-level and node-level similarities, which alleviates statistical heterogeneity and leads to better minority-class decision boundaries. Extensive experiments on four cross-domain federated graph classification tasks demonstrate that SFGL consistently outperforms state-of-the-art baselines.