FedSig: A Federated Graph Augmentation for Class-Imbalanced Node Classification

Federated Learning (FL) is a learning paradigm that collaboratively trains machine learning models among distributed clients without leaking data privacy. A key challenge in federated learning is to handle the class imbalance issue across clients. Although significant efforts have been dedicated to addressing this challenge, the effect of the imbalanced classification on graphs is still not satisfactory. In this paper, we propose FedSig, a novel Federated graph data augmentation that employs Synthetic minority oversampling techniques for Imbalanced Graphs. Specifically, we adopt a shared encoder and private decoder architecture to support data heterogeneity. Building upon FedSig, we propose a two-stage synthetic minority node mechanism that can effectively capture both local and global features for generating minority nodes. In addition, we propose an edge generator to simulate the relationship between synthetic minority nodes and real nodes to balance the original imbalanced graph. To mitigate drift between local and global models, we propose a novel regularization term that enforces proximity between the distribution of local minority node embeddings and that of global minority node embeddings. Extensive experiments on Pubmed, Github, and Amazon-cs demonstrate that FedSig outperforms state-of-the-art approaches for imbalanced node classification.