Revisiting fairGNN-WOD: A Reproduction and Analysis of Fair Graph Learning Without Demographics

TMLR Paper9464 Authors

Abstract: Graph Neural Networks are widely used for high-impact prediction tasks where fairness is important. fairGNN-WOD (Wang et al., 2025b) proposes a two-stage framework to make fair predictions without relying on sensitive demographics information, which is often unavailable. In this study, we implement fairGNN-WOD from scratch due to the lack of a publicly available code base. Ultimately, while we reproduce utility results, we fail to reproduce the reported fairness improvements, which is the main contribution of the original paper, because baselines are substantially more fair than originally reported. Additionally, we find no measurable contribution of stage 1 of the original framework, which is a key architectural component of fairGNN-WOD.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Qi_CHEN6

Submission Number: 9464