GRAPHGINI: Fostering Individual and Group Fairness in Graph Neural Networks

TL;DR

GraphGini tackles fairness in graph neural networks by replacing the traditional Lipschitz-based, worst-case notion of individual fairness with a distributional Gini coefficient that accounts for the entire outcome spectrum. It introduces a differentiable upper bound $ ext{Gini}(\,\mathcal{V})$ through $ ext{Tr}(\mathbf{Z}^{T}\mathbf{L}\mathbf{Z})$ and uses Nash Social Welfare to achieve Pareto-optimal group fairness across sensitive-group partitions, all balanced by GradNorm to automate weight calibration. The approach yields significant improvements in individual fairness while preserving utility and achieving strong group fairness and equal-opportunity performance across multiple real-world datasets and GNN backbones. This work bridges economics-based fairness metrics with machine learning, offering a scalable, principled framework for safe, fair GNN deployments.

Abstract

Graph Neural Networks (GNNs) have demonstrated impressive performance across various tasks, leading to their increased adoption in high-stakes decision-making systems. However, concerns have arisen about GNNs potentially generating unfair decisions for underprivileged groups or individuals when lacking fairness constraints. This work addresses this issue by introducing GraphGini, a novel approach that incorporates the Gini coefficient to enhance both individual and group fairness within the GNN framework. We rigorously establish that the Gini coefficient offers greater robustness and promotes equal opportunity among GNN outcomes, advantages not afforded by the prevailing Lipschitz constant methodology. Additionally, we employ the Nash social welfare program to ensure our solution yields a Pareto optimal distribution of group fairness. Extensive experimentation on real-world datasets demonstrates GraphGini's efficacy in significantly improving individual fairness compared to state-of-the-art methods while maintaining utility and group fairness.

Figures (7)

• Figure 1: The figure illustrates the pipeline of GraphGini. The sequence of actions depicted in this figure is formally encapsulated in Alg. \ref{['alg:graphgini']} (in Appendix) and discussed in § \ref{['sec:framework']}.
• Figure 2: Impact of attention on individual fairness.
• Figure 3: Utility-Fairness Curve comparison of benchmarked algorithms across datasets using the GCN backbone architecture.
• Figure D: Lorenz curve
• Figure E: The Elbow plots show the optimal number of clusters in each dataset on K-means clustering.

Theorems & Definitions (19)

• Definition 1: Graph
• Definition 2: Graph Neural Network (GNN) and Node Embeddings
• Definition 3: Sensitive attributes
• Definition 4: User similarity matrix $\mathbf{S}$ and Laplacian $\mathbf{L}$
• Definition 5
• Definition 6: Gini Coefficient for Individual fairness
• Definition 7: Group Fairness
• Definition 8: Group Disparity
• Definition 9: Equal Opportunity hardt2016equality
• Proposition 1