Model-Agnostic Fairness Regularization for GNNs with Incomplete Sensitive Information

TL;DR

This work tackles fairness in Graph Neural Networks under incomplete sensitive information by introducing EOSP, a model-agnostic regularization framework that jointly enforces equal opportunity and statistical parity as differentiable losses during training. By operating on partially labeled sensitive attributes, EOSP preserves predictive utility while reducing bias, as demonstrated across five real-world datasets and multiple GNN architectures. The approach combines two regularizers with a controllable trade-off tuned via Bayesian optimization, achieving favorable fairness-accuracy dynamics and practical training overhead. The results suggest EOSP is a robust, scalable solution for post-hoc fairness concerns in graph-based semi-supervised tasks with limited sensitive attribute access.

Abstract

Graph Neural Networks (GNNs) have demonstrated exceptional efficacy in relational learning tasks, including node classification and link prediction. However, their application raises significant fairness concerns, as GNNs can perpetuate and even amplify societal biases against protected groups defined by sensitive attributes such as race or gender. These biases are often inherent in the node features, structural topology, and message-passing mechanisms of the graph itself. A critical limitation of existing fairness-aware GNN methods is their reliance on the strong assumption that sensitive attributes are fully available for all nodes during training--a condition that poses a practical impediment due to privacy concerns and data collection constraints. To address this gap, we propose a novel, model-agnostic fairness regularization framework designed for the realistic scenario where sensitive attributes are only partially available. Our approach formalizes a fairness-aware objective function that integrates both equal opportunity and statistical parity as differentiable regularization terms. Through a comprehensive empirical evaluation across five real-world benchmark datasets, we demonstrate that the proposed method significantly mitigates bias across key fairness metrics while maintaining competitive node classification performance. Results show that our framework consistently outperforms baseline models in achieving a favorable fairness-accuracy trade-off, with minimal degradation in predictive accuracy. The datasets and source code will be publicly released at https://github.com/mtavassoli/GNN-FC.

Figures (5)

• Figure 1: Bayesian Optimization Convergence
• Figure 2: Bayesian optimization for SAGE-EOSP on the German dataset: (a) Balanced accuracy decreases slightly but stabilizes after 25 trials; (b) Statistical parity ($\Delta_{SP}$) shows significant improvement, decreasing from 11.63 to 8.37; (c) Equal opportunity ($\Delta_{EO}$) converges to stable values. All metrics reach convergence within 25 trials.
• Figure 3: Hyper-parameter analysis on the German dataset, evaluating the effect of varying $\alpha$ and $\beta$ on predictive performance and fairness metrics.
• Figure 4: Hyperparameter analysis on the German dataset using Bayesian optimization, evaluating the effects of the fairness coefficients $\alpha$ (equal opportunity) and $\beta$ (statistical parity) on the accuracy-fairness trade-off.
• Figure 5: Impact of labeled sensitive attribute proportion on GCN vs GCN-EOSP performance across three metrics.

Theorems & Definitions (1)

• Definition 1: Possibility of Fairness Combinations