Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks

TL;DR

This work tackles fairness in GNNs by explicitly disentangling attribute, structure, and potential biases into separate embeddings. It introduces DAB-GNN, which amplifies each bias during a dedicated disentanglement phase and then debiases via two regularizers that enforce separation and reduce sensitive information leakage, optimizing a multi-objective loss $L_{total} = L_{primary} + \alpha L_{fh} + \beta L_{bco}$. Extensive experiments on five real-world datasets against twelve baselines demonstrate improved accuracy–fairness trade-offs, with evidence of effective bias isolation in embedding space and near-linear training cost. The approach offers a practical path to fair graph representations, supported by ablation and hyperparameter analyses and provided code for reproducibility.

Abstract

Graph Neural Networks (GNNs) have become essential tools for graph representation learning in various domains, such as social media and healthcare. However, they often suffer from fairness issues due to inherent biases in node attributes and graph structure, leading to unfair predictions. To address these challenges, we propose a novel GNN framework, DAB-GNN, that Disentangles, Amplifies, and deBiases attribute, structure, and potential biases in the GNN mechanism. DAB-GNN employs a disentanglement and amplification module that isolates and amplifies each type of bias through specialized disentanglers, followed by a debiasing module that minimizes the distance between subgroup distributions. Extensive experiments on five datasets demonstrate that DAB-GNN significantly outperforms ten state-of-the-art competitors in terms of achieving an optimal balance between accuracy and fairness. The codebase of DAB-GNN is available at https://github.com/Bigdasgit/DAB-GNN

Figures (3)

• Figure 1: Overview of DAB-GNN, which consists of (M1) disentanglement and amplification module, and (M2) debiasing module.
• Figure 2: Visualization of disentangled node embeddings by using t-SNE: AbEmb, SbEmb, and PbEmb.
• Figure 3: The effects of $\alpha$ and $\beta$ on AUC ($\uparrow$) and EO ($\downarrow$).