Causal Hyperbolic Encoder with Unseen Prototype Offset for federated graph learning

Graph Federated Learning (GFL) faces two major challenges: structural heterogeneity and class imbalance. Structural heterogeneity, arising from substantial discrepancies in graph structures across clients, can distort information propagation and hinder global generalization. Class imbalance occurs when certain classes are underrepresented or absent on local clients, thereby impairing the model’s ability to recognize locally unseen classes. To address these challenges, we propose FedCHE, a federated graph learning framework that integrates a Causal Hyperbolic Encoder (CHE) with a learnable Unseen Prototype Offset (UPO) mechanism. Specifically, FedCHE maps node features into the Poincaré ball model of hyperbolic space to better capture hierarchical graph structures. To further enhance information propagation, CHE incorporates a causal attention mechanism that selectively identifies reliable neighbors, suppresses noisy interactions, and facilitates the transfer of informative structural signals. In addition, FedCHE alleviates class imbalance through the UPO mechanism, which generates auxiliary prototypes for unseen classes, thereby bridging semantic gaps and improving the recognition of locally missing classes. By combining global prototype aggregation with a cross-domain alignment loss, FedCHE promotes semantic consistency across clients and enables effective knowledge sharing without exchanging raw graph data. Extensive experiments on multiple benchmark datasets demonstrate that FedCHE consistently outperforms state-of-the-art baselines, including Federated Learning for Local Generalization (FedLoG). The results further show that FedCHE achieves superior generalization on seen data, robust performance in scenarios involving unseen nodes and unseen classes, and rapid adaptability to newly introduced clients.