FairSEP: Harmonizing Multiple Fairness Metrics in Graph Neural Networks via Pareto Optimization
Abstract
Group fairness in Graph Neural Networks (GNNs) aims to mitigate performance disparities across demographic groups. While existing methods primarily focus on Demographic Parity and Equal Opportunity, Predictive Equality (ΔFPR) remains under-explored, potentially leading to disparate error risks. Furthermore, these methods often rely on a single surrogate loss of fairness, which is insufficient to achieve multiple fairness criteria simultaneously. Moreover, these methods often apply uniform debiasing across different label groups, making it difficult to separately address group disparities in the positive and negative classes. Consequently, achieving an optimal trade-off between utility and multiple fairness metrics remains challenging. To address these issues, we propose FairSEP, a unified Pareto-optimized framework. Specifically, we construct a sensitive subspace from per-class centroid differences between sensitive groups, debias embeddings by removing their subspace projections, and enforce prediction consistency between original and debiased representations. On this basis, we introduce a conditional Wasserstein loss that minimizes the Wasserstein distance between predictive distributions of sensitive groups conditioned on true labels, explicitly reducing disparate error rates across groups. Finally, we design a Pareto-based strategy that automatically adjusts the loss weights to resolve conflicts between different objectives. Experiments on five benchmark datasets demonstrate that FairSEP achieves a superior trade-off compared to state-of-the-art baselines.