Fairness-Aware Subgraph Diffusion for Debiasing Graph Neural Network Predictions
A generative debiasing procedure for graph learning that samples small subgraphs from a large input graph and passes them through a fairness-aware diffusion process defined by stochastic differential equations. Adversarial bias perturbations are injected during forward diffusion and predicted by score-based models that learn the bias dynamics; reverse diffusion then removes unfairness, and a standard graph neural network trained on the debiased subgraphs yields fair node predictions.
2501.00595
Graph Neural Networks (GNNs) amplify biases in training data through their message-passing mechanism, producing unfair predictions with respect to sensitive node attributes such as age and gender. Th…