Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Homophily-aware Supervised Contrastive Counterfactual Augmented Fair Graph Neural Network

Mahdi Tavassoli Kejani, Fadi Dornaika, Charlotte Laclau, Jean-Michel Loubes

Abstract

In recent years, Graph Neural Networks (GNNs) have achieved remarkable success in tasks such as node classification, link prediction, and graph representation learning. However, they remain susceptible to biases that can arise not only from node attributes but also from the graph structure itself. Addressing fairness in GNNs has therefore emerged as a critical research challenge. In this work, we propose a novel model for training fairness-aware GNNs by improving the counterfactual augmented fair graph neural network framework (CAF). Specifically, our approach introduces a two-phase training strategy: in the first phase, we edit the graph to increase homophily ratio with respect to class labels while reducing homophily ratio with respect to sensitive attribute labels; in the second phase, we integrate a modified supervised contrastive loss and environmental loss into the optimization process, enabling the model to jointly improve predictive performance and fairness. Experiments on five real-world datasets demonstrate that our model outperforms CAF and several state-of-the-art graph-based learning methods in both classification accuracy and fairness metrics.

Homophily-aware Supervised Contrastive Counterfactual Augmented Fair Graph Neural Network

Abstract

In recent years, Graph Neural Networks (GNNs) have achieved remarkable success in tasks such as node classification, link prediction, and graph representation learning. However, they remain susceptible to biases that can arise not only from node attributes but also from the graph structure itself. Addressing fairness in GNNs has therefore emerged as a critical research challenge. In this work, we propose a novel model for training fairness-aware GNNs by improving the counterfactual augmented fair graph neural network framework (CAF). Specifically, our approach introduces a two-phase training strategy: in the first phase, we edit the graph to increase homophily ratio with respect to class labels while reducing homophily ratio with respect to sensitive attribute labels; in the second phase, we integrate a modified supervised contrastive loss and environmental loss into the optimization process, enabling the model to jointly improve predictive performance and fairness. Experiments on five real-world datasets demonstrate that our model outperforms CAF and several state-of-the-art graph-based learning methods in both classification accuracy and fairness metrics.

Paper Structure

This paper contains 38 sections, 3 theorems, 30 equations, 5 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Problem Statement and Background Knowledge
  4. Problem statement
  5. Notations.
  6. From tabular data to graphs.
  7. Objective
  8. Fairness In Graph Neural Networks
  9. Prediction-level (Statistical Parity (SP) dai2021say)
  10. Prediction-level (Equal Opportunity (EO) dai2021say)
  11. Graph-level Homophily Ratio chen2023fairness
  12. The Counterfactual Augmented Fair Graph Neural Network (CAF)
  13. Training loss of CAF
  14. Limitations of CAF
  15. Our Contribution.
  16. ...and 23 more sections

Key Result

Theorem 1

For any $K$ consisting only of Type III edges: Moreover, the changes satisfy the exact identities with equality if and only if $|K|=0$ or $N_s=m$. $\blacktriangleleft$$\blacktriangleleft$

Figures (5)

  • Figure 1: Overview of HSCCAF. The framework extends CAF with (i) a fairness-aware graph editing phase that edits the input graph to adjust homophily, and (ii) two additional loss terms that explicitly regularize the content and environmental representations.
  • Figure 2: Hyper-parameter $\omega$ study on the German (two plots on the right) and Bail (two plots on the left) datasets.
  • Figure 3: Hyper-parameter ( $\eta$) study on the German (two plots on the right) and Bail (two plots on the left) dataset.
  • Figure 4: Projection of the embeddings in the Environment component using TSNE on the German dataset.
  • Figure 5: Projection of the embeddings in the Content component using TSNE on the German dataset.

Theorems & Definitions (6)

  • Theorem 1: Monotonic effect of removing Type III edges
  • proof
  • Proposition 1: Single-edge type optimality under deletions
  • proof
  • Corollary 1: Minimal deletions to reach targets
  • proof