Towards Fair Graph Neural Networks via Graph Counterfactual without Sensitive Attributes

Graph-structured data is ubiquitous in today's connected world, driving\nextensive research in graph analysis. Graph Neural Networks (GNNs) have shown\ngreat success in this field, leading to growing interest in developing fair\nGNNs for critical applications. However, most existing fair GNNs focus on\nstatistical fairness notions, which may be insufficient when dealing with\nstatistical anomalies. Hence, motivated by causal theory, there has been\ngrowing attention to mitigating root causes of unfairness utilizing graph\ncounterfactuals. Unfortunately, existing methods for generating graph\ncounterfactuals invariably require the sensitive attribute. Nevertheless, in\nmany real-world applications, it is usually infeasible to obtain sensitive\nattributes due to privacy or legal issues, which challenge existing methods. In\nthis paper, we propose a framework named Fairwos (improving Fairness without\nsensitive attributes). In particular, we first propose a mechanism to generate\npseudo-sensitive attributes to remedy the problem of missing sensitive\nattributes, and then design a strategy for finding graph counterfactuals from\nthe real dataset. To train fair GNNs, we propose a method to ensure that the\nembeddings from the original data are consistent with those from the graph\ncounterfactuals, and dynamically adjust the weight of each pseudo-sensitive\nattribute to balance its contribution to fairness and utility. Furthermore, we\ntheoretically demonstrate that minimizing the relation between these\npseudo-sensitive attributes and the prediction can enable the fairness of GNNs.\nExperimental results on six real-world datasets show that our approach\noutperforms state-of-the-art methods in balancing utility and fairness.\n