FPGNN: Fair path graph neural network for mitigating discrimination

Abstract

Fairness is a key issue in many real decision-making applications. Existing Graph Neural Network (GNN) models, designed for making non-discrimination decisions, are dependent on the assumption that the high-degree node in data can significantly impact the network even if the high-degree node is not clearly representative of the whole network. However, they can still lead to discriminatory decisions because the sensitive information of high-degree nodes may be easily leaked to the low-degree nodes in the network during the process of feature propagation. To reduce the impact of sensitive information leakage, a novel deep model is proposed with an expandable random walk approach (i.e., fair path), referred to as FPGNN (Fair Path Graph Neural Network). In the FPGNN, a gradient penalty is employed in adversarial learning to enhance the learning power of the FPGNN algorithm for capturing the sensitive information of these high-degree nodes. During the learning process, the fair path approach is first utilized to discover the high-degree nodes that have a large effect on node fairness. And then, the sensitive information leakage of high-degree nodes is mitigated by modifying the structural information. The experimental results on three real-world datasets demonstrate that the FPGNN is the best in all GNN models for making fair decisions.