FacGNN: Multi-faceted Fairness Enhancement for GNN through Adversarial and Contrastive Learning

Albeit presenting great capabilities for graph node representations, Graph neural networks (GNNs) face biases and discrimination issues. Existing solutions mostly focus on one aspect of fairness, failing to thoroughly consider the multiple dimensions of fairness. To break this limitation, we propose a novel fairness-aware framework, FacGNN, which establishes connections between group fairness, counterfactual fairness, and stability fairness for the first time. Specifically, FacGNN enhances both group fairness and counterfactual fairness while extending stability metrics beyond prior works. FacGNN employs a phased nested iterative training process with self-supervised contrastive learning, to model fairness relationships and improve discriminatory detection in adversarial frameworks. Experimental results show that FacGNN outperforms previous methods across group fairness, counterfactual fairness, and model stability while maintaining model performance.