Balancing Fairness and Accuracy in Graph Learning via Fairness-Constrained Rewiring

Abstract

Graph rewiring can improve graph neural network accuracy by mitigating topological pathologies such as oversmoothing and oversquashing, but those same changes may amplify fairness issues. This work studies the fairness-accuracy tradeoff in graph rewiring and introduces topological bias as a fairness metric for evaluating relational bias separately from feature bias. It then proposes a fairness constraint that can be incorporated into classical rewiring methods to reduce topological bias. Experiments show that fairness-constrained rewiring can balance downstream accuracy with fairness in graph learning tasks.

Publication
NeurIPS Workshop on Symmetry and Geometry in Neural Representations (Best Paper Award)