A Stochastic Interpolants Method for Unifying Individual and Counterfactual Fairness in Graphs

Jiayi Fan, Zheyun Qin, Hangyuan Du, Xiaoming Xi, Haoliang Sun, Zhongyi Han, Yilong Yin

Keywords: Counterfactual Fairness, Fairness in Graph Neural Networks, Individual Fairness, Optimal Transport, Stochastic Interpolants

TL;DR: SI-Fair unifies individual and counterfactual fairness in graph neural networks through stochastic interpolants theory.

Abstract: The fairness problem in graph neural networks faces a serious conflict between individual fairness and counterfactual fairness. Strict adherence to individual fairness perpetuates structural bias, whereas excessive pursuit of counterfactual fairness may overlook genuine structural differences. This paper proposes the first framework that applies stochastic interpolants to graph fairness problems. Unlike optimal transport methods based on the Wasserstein distance or the Sinkhorn algorithm, our framework precisely controls noise levels in the transport path, allowing dynamic adjustment of emphasis on the two fairness criteria at different stages. Our method overcomes the binary choice limitation in traditional fairness approaches and achieves a continuous trade-off between the two criteria. Specifically, we design a structure-attribute disentanglement representation method that decomposes node representations into bias-carrying features and unbiased structural attributes. Through dynamic noise-level adjustment during transport, we achieve gradual integration of the two criteria. Theoretical analysis proves the upper bound of Kullback-Leibler divergence between the model and the ideal fair distribution. This bound decomposes into the realization levels of individual fairness and counterfactual fairness. Experiments on multiple datasets, including Pokec, Facebook100, Credit-Default, and COMPAS, show that, compared with existing methods, our framework significantly improves both fairness metrics while maintaining model performance.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 3422