Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FairGSE: Fairness-Aware Graph Neural Network without High False Positive Rates

Zhenqiang Ye, Jinjie Lu, Tianlong Gu, Fengrui Hao, Xuemin Wang

TL;DR

This work addresses fairness in GNNs by identifying and mitigating the high false positive rate (FPR) that can arise when optimizing traditional group-fairness metrics. It introduces two-dimensional structural entropy (2D-SE) as a principled objective to balance message aggregation across sensitive groups, and proposes FairGSE, a framework that jointly learns an edge-weighted graph structure and uses contrastive learning to preserve essential graph information. The approach yields competitive accuracy and fairness while substantially reducing FPR on high-risk datasets, demonstrating a meaningful improvement over state-of-the-art fairness-aware GNNs. The work offers practical impact for deploying fair GNNs in important domains like credit risk and social networks, where reliable negative predictions are crucial.

Abstract

Graph neural networks (GNNs) have emerged as the mainstream paradigm for graph representation learning due to their effective message aggregation. However, this advantage also amplifies biases inherent in graph topology, raising fairness concerns. Existing fairness-aware GNNs provide satisfactory performance on fairness metrics such as Statistical Parity and Equal Opportunity while maintaining acceptable accuracy trade-offs. Unfortunately, we observe that this pursuit of fairness metrics neglects the GNN's ability to predict negative labels, which renders their predictions with extremely high False Positive Rates (FPR), resulting in negative effects in high-risk scenarios. To this end, we advocate that classification performance should be carefully calibrated while improving fairness, rather than simply constraining accuracy loss. Furthermore, we propose Fair GNN via Structural Entropy (\textbf{FairGSE}), a novel framework that maximizes two-dimensional structural entropy (2D-SE) to improve fairness without neglecting false positives. Experiments on several real-world datasets show FairGSE reduces FPR by 39\% vs. state-of-the-art fairness-aware GNNs, with comparable fairness improvement.

FairGSE: Fairness-Aware Graph Neural Network without High False Positive Rates

TL;DR

This work addresses fairness in GNNs by identifying and mitigating the high false positive rate (FPR) that can arise when optimizing traditional group-fairness metrics. It introduces two-dimensional structural entropy (2D-SE) as a principled objective to balance message aggregation across sensitive groups, and proposes FairGSE, a framework that jointly learns an edge-weighted graph structure and uses contrastive learning to preserve essential graph information. The approach yields competitive accuracy and fairness while substantially reducing FPR on high-risk datasets, demonstrating a meaningful improvement over state-of-the-art fairness-aware GNNs. The work offers practical impact for deploying fair GNNs in important domains like credit risk and social networks, where reliable negative predictions are crucial.

Abstract

Graph neural networks (GNNs) have emerged as the mainstream paradigm for graph representation learning due to their effective message aggregation. However, this advantage also amplifies biases inherent in graph topology, raising fairness concerns. Existing fairness-aware GNNs provide satisfactory performance on fairness metrics such as Statistical Parity and Equal Opportunity while maintaining acceptable accuracy trade-offs. Unfortunately, we observe that this pursuit of fairness metrics neglects the GNN's ability to predict negative labels, which renders their predictions with extremely high False Positive Rates (FPR), resulting in negative effects in high-risk scenarios. To this end, we advocate that classification performance should be carefully calibrated while improving fairness, rather than simply constraining accuracy loss. Furthermore, we propose Fair GNN via Structural Entropy (\textbf{FairGSE}), a novel framework that maximizes two-dimensional structural entropy (2D-SE) to improve fairness without neglecting false positives. Experiments on several real-world datasets show FairGSE reduces FPR by 39\% vs. state-of-the-art fairness-aware GNNs, with comparable fairness improvement.

Paper Structure

This paper contains 29 sections, 6 theorems, 39 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fairness in Graph
  4. Structural Entropy
  5. Preliminaries
  6. Notation
  7. Fairness Metrics
  8. Theoretical Analysis
  9. The Proposed Framework: FairGSE
  10. Overview
  11. Graph Structure Learner
  12. Contrastive Learning Component
  13. Structure Bootstrapping Mechanism
  14. Training Objective
  15. Experiments
  16. ...and 14 more sections

Key Result

Lemma 1

Under two assumptions: (1) $\text{vol}(G) > 0$; (2) $\text{vol}(\mathcal{V}_{S_i}) > 0,\ \forall \mathcal{V}_{S_i} \in \mathcal{P}_S$. the 2D-SE $H^{\mathcal{P}_S}(G)$ is differentiable w.r.t. every edge $\mathbf{A}_{(i,j)}(>0)$, with:

Figures (4)

  • Figure 1: FPR, $\Delta_{SP}$ and $\Delta_{EO}$ of Fairness-aware GNNs on Credit dataset.
  • Figure 2: Overview of FairGSE, which consists of graph structure learner, contrastive learning component and structure bootstrapping mechanism.
  • Figure 3: Ablation study results for FairGSE on all datasets. Higher value indicating better performance for AUC, and lower values preferred for FPR, $\Delta_{\textit{SP}}$ and $\Delta_{\textit{EO}}$.
  • Figure 4: The hyperparameters study results on the Credit.

Theorems & Definitions (9)

  • Definition 1: 2D-SE w.r.t. sensitive attribute
  • Lemma 1: Optimization Feasibility
  • Theorem 1: Fairness via 2D-SE bound
  • Theorem 2: FPR via 2D-SE bound
  • Lemma 2: Optimization Feasibility
  • Theorem 3: Fairness via 2D-SE bound
  • Proof
  • Theorem 4: FPR via 2D-SE bound
  • Proof