Towards Fair Graph Neural Networks via Graph Counterfactual without Sensitive Attributes

TL;DR

This paper addresses the challenge of fairness in graph neural networks when sensitive attributes are unavailable. It proposes Fairwos, a framework that constructs pseudo-sensitive attributes through an encoder and generates graph counterfactuals guided by these proxies, aligning representations between original and counterfactual graphs to reduce bias. A dynamic weight update mechanism balances the contribution of each pseudo-sensitive attribute to fairness and utility, with theoretical guarantees on fairness bounds and convergence. Empirical results across six real-world datasets demonstrate that Fairwos achieves superior fairness-utility trade-offs compared to state-of-the-art baselines, while maintaining competitive predictive performance.

Abstract

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

Figures (8)

• Figure 1: Running example of node classification in a loan approval scenario: users are classified based on their features (race, income, zip code) as well as their relationships with others to predict loan approval.
• Figure 2: An illustration of our proposed framework
• Figure 3: A causal relationship exists between sensitive attributes and predictions, where pseudo-sensitive attributes are low-dimensional representations of non-sensitive attributes and graph structure. (a) Applying regularization to each pseudo-sensitive attribute ensures that sensitive attributes do not influence predictions; (b) To balance the impact of pseudo-sensitive attributes on utility and fairness, a weight-update mechanism is used to adjust the strength of fairness enhancement.
• Figure 4: Comparisons between Fairwos and its variants on NBA and Bail dataset
• Figure 5: Impacts of the dimension of the encoder to Fairwos on GCN

Theorems & Definitions (7)

• Theorem 1
• proof
• Theorem 2
• proof
• Lemma 1
• Theorem 3
• proof