Toward Fair Federated Graph Learning

As a privacy-preserving collaborative paradigm, federated graph learning (FGL) enables distributed training of graph neural networks (GNNs) without exposing raw graph data. Subgraph-FL has become the dominant FGL paradigm, yet most studies focus on overall node classification performance while overlooking fairness issues stemming from heterogeneous node profiles and graph topology. Specifically, they exhibit biased performance to nodes with disadvantageous properties, such as being minority-class within local subgraphs or heterophilous connections (i.e., neighboring nodes possess dissimilar labels and misleading features). This underexplored fairness challenge reveals the robustness concerns of current subgraph-FL methods: high accuracy conceals degraded performance on structurally or semantically marginalized node groups. To address this, we advocate for: 1) enhancing the representation of minority-class nodes for class-wise fairness and 2) mitigating topological biases arising from heterophilous connections for topology-aware fairness. In this context, we propose FairFGL, a novel framework that performs fine-grained mining of graph properties and orchestrates a collaborative learning paradigm to enhance fairness. Specifically, on the client side, the majority alignment module enhances local clients’ expertise for boosting efficient cross-client knowledge transfer. The gradient modification module and the history-preserving module infuse scarce local minority knowledge through cross-client collaboration and regulate local training from being over-fit to locally dominant distribution. On the server side, FairFGL only requires uploading the changed value of the most influential subset of locally trained parameters. Subsequently, a cluster-based aggregation strategy reconciles conflicting updates from heterogeneous data distribution across clients, and suppresses global majority dominance to the newly aggregated global model. Extensive evaluations on eight benchmark datasets show that FairFGL significantly improves performance for disadvantaged node groups, achieving up to 21.07% increase in Overall F1—while enhancing convergence efficiency over SOTA baselines.