Grad2Fair Achieves Graph Fairness Without Demographics

Graph Neural Networks (GNNs) often exhibit group fairness issues, leading to biased predictions against specific demographic groups. Most existing solutions rely on the strong assumption of fully available sensitive demographic attributes, which is often not met in real-world scenarios. Attempts to use predicted demographics as proxies have proven unreliable due to potential inaccuracies. This research introduces Grad2Fair, a groundbreaking approach that tackles graph fairness without needing explicit demographic information. The method is based on the observation that gradient distributions of misclassified nodes implicitly encode demographic biases. Grad2Fair first proposes GradDist, a gradient-based metric to quantify bias by measuring distances between local modes within these distributions. To mitigate this identified bias, Grad2Fair directly leverages these gradients to debias the GNN, completely bypassing the need for demographic prediction. This direct gradient-guided debiasing leads to more stable fairness performance. Experiments on several real-world datasets confirm Grad2Fair's effectiveness, consistently outperforming existing baselines in achieving group fairness.