Interpreting Unfairness in Graph Neural Networks via Training Node Attribution

Graph Neural Networks (GNNs) have emerged as the leading paradigm for solving graph analytical problems in various real-world applications. Nevertheless, GNNs could potentially render biased predictions towards certain demographic subgroups. Understanding how bias arises is critical, it guides design of GNN debiasing mechanisms. However, most existing works overwhelmingly focus on debiasing, but fall short explaining such induced. In this paper, we study a novel problem interpreting unfairness through attributing to influence training nodes. Specifically, propose strategy named Probabilistic Distribution Disparity (PDD) measure exhibited GNNs, and develop an algorithm efficiently estimate each node bias. We verify validity PDD effectiveness estimation experiments datasets. Finally, also demonstrate proposed framework be used GNNs. Open-source code can found at https://github.com/yushundong/BIND.