Learning fair graph representation through graph information disentanglement.

Abstract

Graph Neural Networks (GNNs) are widely applied to graph-structured data, but they often suffer from fairness concerns, as inherent biases in node attributes and graph topology can result in discriminatory predictions. Existing approaches typically attempt to mitigate all sources of bias within a single, entangled representation, thereby limiting the effectiveness of debiasing. To overcome this limitation, we propose FairGID, a novel framework for fair graph representation learning that enhances fairness by separating topology from node attributes and disentangling node representations. Specifically, FairGID first learns attribute-only and structure-only representations independently. It then further disentangles the attribute representations into multiple latent factors and applies sensitive attribute masking to suppress bias-related information. Finally, an adversarial fusion module integrates the attribute and structural representations into a unified embedding that is both informative and fair. Extensive experiments on five real-world datasets demonstrate that FairGID achieves a superior accuracy-fairness trade-off compared with state-of-the-art baselines, highlighting its potential as an effective solution for fair graph representation learning.