Mitigating bias for unseen demographic groups in graph neural networks

Fairness is an important factor to consider in graph neural networks (GNNs) as biases in the graph data can be amplified through the link structure. While many fairness-aware GNN methods have been proposed, most assume that the sensitive attribute values for all demographic groups are available during training. To overcome this limitation, we propose FairGRUNT+, a novel framework that integrates novelty detection with disentangled representation learning to jointly learn fair node embeddings for classification while identifying previously unseen demographic groups. Trained exclusively on samples from observed demographic groups, FairGRUNT+ learns a nonlinear decision boundary that encloses nodes from known groups, allowing it to identify nodes falling outside the boundary as belonging to novel (unseen) groups. To promote fairness, FairGRUNT+ further employs disentangled representation learning to decouple node embeddings used for class prediction from those encoding demographic information. Experimental results on real-world datasets demonstrate that FairGRUNT+ outperforms existing fairness-aware GNN methods by effectively reducing bias in node classification while maintaining strong predictive performance.