Heterogeneous fairness-aware adaptive feature enhancement framework for graph neural networks learning

Current Graph Neural Networks (GNNs), which predominantly rely on the homophily assumption, often amplify biases and degrade performance for minority groups in real-world graphs characterized by prevalent heterophilic structures. Current methods face limitations due to either dependence on sensitive attributes or separate handling of biases in both data and models, restricting their utility in privacy-sensitive contexts. To address these limitations, we propose a heterogeneous fairness-aware adaptive feature enhancement framework, namely FairHAFE , for fair node classification in GNNs. Different from existing methods, FairHAFE does not require sensitive attribute labels and jointly mitigates data and model biases through heterogeneity-aware feature enhancement. The novelty of FairHAFE lies in its unified design: it requires no sensitive attribute labels and leverages local feature and structural incompatibility as a proxy signal to jointly mitigate data and model biases. The FairHAFE framework consists of two core modules: a Heterogeneity Fairness-Aware Module, which unsupervisedly quantifies and identifies heterophilic edges using information-theoretic-inspired proxy scores, and an Adaptive Feature Enhancement Module, which utilizes a trained estimator to predict nodes’ heterogeneous context for adaptive feature enhancement. Designed as a preprocessing framework, FairHAFE requires no modification to the backend GNN architecture. Experimental results on multiple real-world datasets show that FairHAFE not only enhances fairness but also preserves utility, outperforming existing baselines. FairHAFE leverages local feature and structural incompatibility as a proxy signal to guide adaptive feature compensation, providing a practical and privacy-preserving route toward fair graph learning.