Towards Federated Long-Tailed Graph Learning: An Energy-Guided Dual Decoupling Approach

In federated graph learning (FGL), data often follows a long-tailed distribution where majority classes dominate the structure. This leads to two critical problems: minority nodes are structurally isolated and submerged in heterophilic (label-disparate) neighborhoods, and standard statistical adjustments for imbalance often cause the model to overfit structural noise from majority classes. The authors investigate how to resolve this misalignment between statistical frequency and topological structure without compromising privacy or model performance.

The authors propose FedEPD, a framework based on a dual decoupling paradigm. First, it performs topological purification using a distribution-aware Dirichlet energy pruning mechanism. This filters heterophilic edges in the spatial domain, ensuring the graph structure is cleaned before representation learning begins. Second, it employs a server-assisted local consensus mechanism to extract robust global prototypes from topologically central nodes. These prototypes are injected into the low-frequency semantic components of local nodes using a spatial low-pass injection technique. Finally, a two-stage alternating optimization strategy is used to protect majority decision boundaries while specifically improving the accuracy of minority classes.

FedEPD significantly outperforms existing federated long-tailed learning methods across diverse benchmarks. By separating topological denoising from semantic recalibration, the model avoids the common pitfall of overfitting structural noise. Experimental results show that FedEPD maintains high classification stability for head and medium classes while achieving substantial accuracy improvements for tail categories, with absolute gains of up to 11.89% in some scenarios. The approach effectively reconciles the optimization conflict between majority and minority classes without introducing significant computational overhead.