Rethinking Fair Graph Neural Networks from Re-balancing

TL;DR

This paper identifies group-level imbalance as a key source of unfairness in graph neural networks and shows that simple re-balancing can rival specialized fair GNNs. It proposes FairGB, combining Counterfactual Node Mixup (CNM) to generate balanced augmented graphs and Contribution Alignment Loss (CAL) to balance gradient contributions across groups. The authors provide causal and statistical analyses to justify debiasing and demonstrate state-of-the-art utility and fairness on German, Bail, and Credit datasets, with robust performance across different encoders and a single hyper-parameter controlling mix types. The work offers a practical, minimally tuned framework for fair graph learning and highlights the effectiveness of re-balancing perspectives in complex graph-structured data.

Abstract

Driven by the powerful representation ability of Graph Neural Networks (GNNs), plentiful GNN models have been widely deployed in many real-world applications. Nevertheless, due to distribution disparities between different demographic groups, fairness in high-stake decision-making systems is receiving increasing attention. Although lots of recent works devoted to improving the fairness of GNNs and achieved considerable success, they all require significant architectural changes or additional loss functions requiring more hyper-parameter tuning. Surprisingly, we find that simple re-balancing methods can easily match or surpass existing fair GNN methods. We claim that the imbalance across different demographic groups is a significant source of unfairness, resulting in imbalanced contributions from each group to the parameters updating. However, these simple re-balancing methods have their own shortcomings during training. In this paper, we propose FairGB, Fair Graph Neural Network via re-Balancing, which mitigates the unfairness of GNNs by group balancing. Technically, FairGB consists of two modules: counterfactual node mixup and contribution alignment loss. Firstly, we select counterfactual pairs across inter-domain and inter-class, and interpolate the ego-networks to generate new samples. Guided by analysis, we can reveal the debiasing mechanism of our model by the causal view and prove that our strategy can make sensitive attributes statistically independent from target labels. Secondly, we reweigh the contribution of each group according to gradients. By combining these two modules, they can mutually promote each other. Experimental results on benchmark datasets show that our method can achieve state-of-the-art results concerning both utility and fairness metrics. Code is available at https://github.com/ZhixunLEE/FairGB.

Figures (8)

• Figure 1: The F1-$\Delta_{sp}$ trade-off on German, Bail, and Credit datasets. "RW" denotes re-weighting and "OS" denotes over-sampling. The light green region indicates the model outperforms the vanilla model in both utility and fairness, while the light red region represents the opposite.
• Figure 2: The overview of proposed FairGB.
• Figure 4: Visualizations of node representation learned on the Bail dataset.
• Figure 5: Comparison of the utility and fairness performance with different graph encoders.
• Figure 6: Parameter sensitivity results w.r.t.$\eta$.

Theorems & Definitions (1)

• Theorem 1