[3K5-OS-2b] OS-2

[3K5-OS-2b-02] A Benchmarking Study of Group-Fair Graph Neural Networks

〇Joyce Guo², Yuya Sasaki¹ (1. Osaka university, 2. University of California, Berkley)

[[Online]]

Keywords:Fair Graph Neural Networks, Benchmarking, Group fairness

Studies in fair graph neural networks (Fair GNNs) is expanding, aiming to mitigate discriminating outputs. However, the lack of a comprehensive benchmark for fair GNN methods prevent progress assessment and future improvements, which leads to inconsistent experimental settings, difficulty in reproducing, and comparing results. In this paper, we conduct a benchmark that evaluates various fair GNN methods on five datasets.
We compare results of three state-of-the-art fair GNN methods (FairGNN, BIND, and NIFTY) on three GNN backbone models (GCN, GAT, and GraphSAGE).
Through experiments and analysis of both fair and regular GNN methods, our findings provide valuable insights and establish a baseline for future advancements in fair machine learning.