Fair Graph Neural Network with Supervised Contrastive Regularization

TL;DR

This work tackles fairness in graph neural networks by extending the Counterfactual Augmented Fair Graph Neural Network Framework (CAF) with two new losses: a Supervised Contrastive Loss to align content representations for same-label nodes, and an Environmental Loss to separate environment information tied to sensitive attributes. The authors propose SCCAF, which partitions latent representations into content $C$ and environment $E$ and optimizes a combined objective that balances accuracy and fairness. Empirical results on German Credit, Bail, and Credit Defaulter datasets show that SCCAF improves AUC and F1 while reducing $\Delta_{SP}$ and $\Delta_{EO}$ compared with CAF and other baselines, demonstrating a better trade-off between predictive performance and fairness. The approach highlights the practical potential of leveraging contrastive supervision and environment-aware objectives to obtain fair, robust graph representations in semi-supervised settings.

Abstract

In recent years, Graph Neural Networks (GNNs) have made significant advancements, particularly in tasks such as node classification, link prediction, and graph representation. However, challenges arise from biases that can be hidden not only in the node attributes but also in the connections between entities. Therefore, ensuring fairness in graph neural network learning has become a critical problem. To address this issue, we propose a novel model for training fairness-aware GNN, which enhances the Counterfactual Augmented Fair Graph Neural Network Framework (CAF). Our approach integrates Supervised Contrastive Loss and Environmental Loss to enhance both accuracy and fairness. Experimental validation on three real datasets demonstrates the superiority of our proposed model over CAF and several other existing graph-based learning methods.

Figures (2)

• Figure 1: An illustration of our proposed framework
• Figure 2: $\Delta_{EQ}$, $\Delta_{SP}$, AUC, and F1 percentages of the proposed model SCCAF as functions of the parameter $\omega$ and $\eta$ in the Bail databases