Fair Graph Representation Learning via Sensitive Attribute Disentanglement

TL;DR

This work tackles group fairness in graph neural networks by reframing fairness as preserving task-related information while reducing sensitivity to protected attributes. It introduces FairSAD, a two-component framework combining Sensitive Attribute Disentanglement (SAD) and Sensitive Attribute Masking to separate the sensitive attribute into an independent channel and decorrelate it from downstream predictions. The method leverages a neighbor assigner and multi-channel disentangled layers to learn factorized representations, with a distance-covariance based macro/micro-disentanglement objective and a learnable mask guided by covariance with the sensitive attribute. Empirical results on five real-world datasets show that FairSAD achieves superior fairness with competitive or improved utility compared to state-of-the-art baselines, highlighting the practical value of disentanglement for fair graph representation learning.

Abstract

Group fairness for Graph Neural Networks (GNNs), which emphasizes algorithmic decisions neither favoring nor harming certain groups defined by sensitive attributes (e.g., race and gender), has gained considerable attention. In particular, the objective of group fairness is to ensure that the decisions made by GNNs are independent of the sensitive attribute. To achieve this objective, most existing approaches involve eliminating sensitive attribute information in node representations or algorithmic decisions. However, such ways may also eliminate task-related information due to its inherent correlation with the sensitive attribute, leading to a sacrifice in utility. In this work, we focus on improving the fairness of GNNs while preserving task-related information and propose a fair GNN framework named FairSAD. Instead of eliminating sensitive attribute information, FairSAD enhances the fairness of GNNs via Sensitive Attribute Disentanglement (SAD), which separates the sensitive attribute-related information into an independent component to mitigate its impact. Additionally, FairSAD utilizes a channel masking mechanism to adaptively identify the sensitive attribute-related component and subsequently decorrelates it. Overall, FairSAD minimizes the impact of the sensitive attribute on GNN outcomes rather than eliminating sensitive attributes, thereby preserving task-related information associated with the sensitive attribute. Furthermore, experiments conducted on several real-world datasets demonstrate that FairSAD outperforms other state-of-the-art methods by a significant margin in terms of both fairness and utility performance. Our source code is available at https://github.com/ZzoomD/FairSAD.

Figures (7)

• Figure 1: Comparison of FairSAD with existing works. Existing works inevitably eliminate the task-related information due to its correlations with the sensitive attribute.
• Figure 2: Overview of FairSAD. Disentangled layers in this example have three channels due to assuming three latent factors.
• Figure 3: Performance of state-of-the-art fairness methods. Despite significant progress in fairness, these methods inevitably suffer from utility performance degradation.
• Figure 4: Ablation analysis w.r.t. micro- and macro-disentanglement on Credit, Pokec-z, and Pokec-n.
• Figure 5: Parameters sensitivity results on two datasets.

Theorems & Definitions (2)

• lemma 1
• proposition 1