Subgraph Filtering Enhances Fairness in Graph Neural Networks

Graph Neural Networks (GNNs) are powerful tools, but they can inadvertently perpetuate or even amplify unfair biases, even when sensitive attributes are explicitly excluded from node features. This bias often arises because graph topology and the message-passing mechanism can propagate group-correlated signals, especially under conditions of "sensitive homophily" where similar individuals tend to connect.Existing fairness-aware GNN methods typically apply global constraints on representations or prediction distributions, which may not directly address the local structural pathways through which bias propagates. This paper proposes Subgraph Filtering for Fair Graph Neural Networks (SF-GNN), a novel and lightweight framework designed to tackle structural bias at its source.SF-GNN identifies edges prone to bias by combining sensitive homophily with structural propagation amplifiers like hub participation and triadic closure. It then incorporates stochastic edge filtering into each message-passing step, selectively downweighting or removing these problematic edges while preserving the overall graph structure. Training also includes a statistical-parity regularizer with a warm-up schedule to stabilize optimization. Experiments on five benchmark datasets demonstrate that SF-GNN consistently improves fairness while maintaining competitive predictive performance, achieving a superior fairness-accuracy trade-off compared to other fairness-aware GNN baselines.