Personalized federated few-shot node classification

Personalized federated learning (PFL) aims to train customized models for individual clients in a decentralized setting, with the account of non-independent and identically distributed data across clients. However, most PFL methods adopt uniform classification layers for diverse clients and give rise to error-prone predictions, due to the task heterogeneity notably prominent in decentralized graph data scenarios. Although some PFL solutions setup client-specific classification layers for each client and optimize them only locally, they are corrupted with limited local training data. We propose an innovative solution called federated parameter decoupling and node augmentation (FedPANO) to address these problems and to achieve personalized federated few-shot node classification, which is a prevalent and challenging but unexplored topic. Specifically, FedPANO first separates the local model into the GNN and classifier to handle unique client-specific task variations. The GNN is trained through federated learning to capture shared knowledge of graph nodes across clients, while the classifier is custom-designed and trained individually for each client. Additionally, a generic classifier shared among clients is adopted to encourage the GNN’s grasp of shared information. Then FedPANO further proposes the node generator along with its local and collaborative training strategies to deal with the node scarcity of clients. Extensive experimental results on benchmark datasets confirm that FedPANO outperforms eight competitive baselines across different settings.