Improving Fairness in Graph Neural Networks via Counterfactual Debiasing

2025-08-20

AI Summary

Graph Neural Networks (GNNs) often yield unfair predictions due to biases in sensitive attributes (e.g., race, gender) and graph structural biases. Existing debiasing methods typically prune edges or mask features to suppress sensitive information, yet this compromises non-sensitive feature integrity and degrades the accuracy–fairness trade-off. This paper proposes Fair-ICD, a novel end-to-end debiasing framework that performs counterfactual data augmentation *before* message passing: it generates diverse counterfactual neighbors to encourage learning of sensitive-attribute-invariant node representations. Fair-ICD integrates an adversarial discriminator with mainstream GNN backbones without explicitly removing structural components or masking features. Evaluated on multiple benchmark datasets, Fair-ICD significantly improves fairness metrics—e.g., reducing Equalized Odds difference by 32%–58%—while preserving predictive accuracy. Its core innovation lies in being the first method to embed counterfactual augmentation at the *front end* of GNN message passing, thereby simultaneously ensuring fairness guarantees and representation fidelity.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Computer Vision: Bias, Fairness & Privacy
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Social Networks and Social Media: Fairness and bias in social network and social media analysis

Abstract

Graph Neural Networks (GNNs) have been successful in modeling graph-structured data. However, similar to other machine learning models, GNNs can exhibit bias in predictions based on attributes like race and gender. Moreover, bias in GNNs can be exacerbated by the graph structure and message-passing mechanisms. Recent cutting-edge methods propose mitigating bias by filtering out sensitive information from input or representations, like edge dropping or feature masking. Yet, we argue that such strategies may unintentionally eliminate non-sensitive features, leading to a compromised balance between predictive accuracy and fairness. To tackle this challenge, we present a novel approach utilizing counterfactual data augmentation for bias mitigation. This method involves creating diverse neighborhoods using counterfactuals before message passing, facilitating unbiased node representations learning from the augmented graph. Subsequently, an adversarial discriminator is employed to diminish bias in predictions by conventional GNN classifiers. Our proposed technique, Fair-ICD, ensures the fairness of GNNs under moderate conditions. Experiments on standard datasets using three GNN backbones demonstrate that Fair-ICD notably enhances fairness metrics while preserving high predictive performance.

Problem

• Mitigating bias in Graph Neural Networks predictions
• Addressing fairness-accuracy trade-off in GNN classifiers
• Reducing sensitive attribute discrimination in graph data

Innovation

• Counterfactual data augmentation for bias mitigation
• Creating diverse neighborhoods before message passing
• Adversarial discriminator to diminish prediction bias