New Approach for Federated Long-Tailed Graph Learning Introduced

Federated Graph Learning (FGL) allows multiple clients to collaboratively build graph models while maintaining data privacy. However, real-world data often exhibits long-tailed distributions, where a few classes are abundant and many are rare. This imbalance severely impacts FGL performance by biasing global models towards majority classes and isolating minority nodes within heterophilic neighborhoods. Existing solutions often struggle with data scarcity, overfitting noise from dominant classes rather than effectively recovering tail nodes. To overcome these limitations, researchers propose FedEPD, a novel framework based on a dual decoupling paradigm. FedEPD first purifies the graph topology and then recalibrates semantics. Specifically, FedEPD employs distribution-aware Dirichlet energy pruning to filter out irrelevant or misleading heterophilic edges. It then addresses Non-IID data shifts by extracting robust global prototypes from topologically central nodes, which are subsequently injected into local representations. An alternating optimization strategy ensures that majority decision boundaries are protected while minority accuracy is significantly improved, achieving state-of-the-art results on various long-tailed benchmarks.