Abstract: Graph Neural Network (GNN) have been widely applied in numerous real-world scenarios due to their powerful modeling capabilities for graph-structured data. Recent studies indicate that GNN may inherit and amplify biases inherent in training data, leading to unfair treatment of specific groups defined by sensitive attributes. To mitigate bias in graph data, this paper proposes FairDC, a fair graph learning framework based on view decoupling and counterfactual augmentation, which addresses feature bias and structural bias separately. The framework first decouples raw graph data into feature views and structural views, then introduces counterfactual views. Finally, a multi-view fusion strategy is employed to learn fair node representations. Experimental results on multiple benchmark datasets demonstrate that FairDC maintains stable prediction performance while reducing fairness metrics DP and EO by 32% and 36%, respectively, compared with the state-of-the-art baseline DAB. This validates the proposed method’s effective trade-off between utility and fairness.

Key words: graph neural network, fairness, disentanglement, counterfactual augmentation, multi-view fusion