GCARe : mitigating subgroup unfairness in graph condensation through adversarial regularization

Abstract

Training Graph Neural Networks (GNNs) on large-scale graphs in the deep learning era can be expensive. While graph condensation has recently emerged as a promising approach through which to reduce training cost by compressing large graphs into smaller ones and for preserving most knowledge, its capability in treating different node subgroups fairly during compression remains unexplored. In this paper, we investigate current graph condensation techniques from a perspective of fairness, and show that they bear severe disparate impact toward node subgroups. Specifically, GNNs trained on condensed graphs become more biased than those trained on original graphs. Since the condensed graphs comprise synthetic nodes, which are absent of explicit group IDs, the current algorithms used to train fair GNNs fail in this case. To address this issue, we propose Graph Condensation with Adversarial Regularization (GCARe), which is a method that directly regularizes the condensation process to distill the knowledge of different subgroups fairly into resulting graphs. A comprehensive series of experiments substantiated that our method enhances the fairness in condensed graphs without compromising accuracy, thus yielding more equitable GNN models. Additionally, our discoveries underscore the significance of incorporating fairness considerations in data condensation, and offer invaluable guidance for future inquiries in this domain.