Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mitigating topology biases in Graph Diffusion via Counterfactual Intervention

Wendi Wang, Jiaxi Yang, Yongkang Du, Lu Lin

TL;DR

Fair Graph Diffusion Model (FairGDiff) is proposed, a counterfactual-based one-step solution that mitigates topology biases while balancing fairness and utility, and achieves a superior trade-off between fairness and utility.

Abstract

Graph diffusion models have gained significant attention in graph generation tasks, but they often inherit and amplify topology biases from sensitive attributes (e.g. gender, age, region), leading to unfair synthetic graphs. Existing fair graph generation using diffusion models is limited to specific graph-based applications with complete labels or requires simultaneous updates for graph structure and node attributes, making them unsuitable for general usage. To relax these limitations by applying the debiasing method directly on graph topology, we propose Fair Graph Diffusion Model (FairGDiff), a counterfactual-based one-step solution that mitigates topology biases while balancing fairness and utility. In detail, we construct a causal model to capture the relationship between sensitive attributes, biased link formation, and the generated graph structure. By answering the counterfactual question "Would the graph structure change if the sensitive attribute were different?", we estimate an unbiased treatment and incorporate it into the diffusion process. FairGDiff integrates counterfactual learning into both forward diffusion and backward denoising, ensuring that the generated graphs are independent of sensitive attributes while preserving structural integrity. Extensive experiments on real-world datasets demonstrate that FairGDiff achieves a superior trade-off between fairness and utility, outperforming existing fair graph generation methods while maintaining scalability.

Mitigating topology biases in Graph Diffusion via Counterfactual Intervention

TL;DR

Fair Graph Diffusion Model (FairGDiff) is proposed, a counterfactual-based one-step solution that mitigates topology biases while balancing fairness and utility, and achieves a superior trade-off between fairness and utility.

Abstract

Graph diffusion models have gained significant attention in graph generation tasks, but they often inherit and amplify topology biases from sensitive attributes (e.g. gender, age, region), leading to unfair synthetic graphs. Existing fair graph generation using diffusion models is limited to specific graph-based applications with complete labels or requires simultaneous updates for graph structure and node attributes, making them unsuitable for general usage. To relax these limitations by applying the debiasing method directly on graph topology, we propose Fair Graph Diffusion Model (FairGDiff), a counterfactual-based one-step solution that mitigates topology biases while balancing fairness and utility. In detail, we construct a causal model to capture the relationship between sensitive attributes, biased link formation, and the generated graph structure. By answering the counterfactual question "Would the graph structure change if the sensitive attribute were different?", we estimate an unbiased treatment and incorporate it into the diffusion process. FairGDiff integrates counterfactual learning into both forward diffusion and backward denoising, ensuring that the generated graphs are independent of sensitive attributes while preserving structural integrity. Extensive experiments on real-world datasets demonstrate that FairGDiff achieves a superior trade-off between fairness and utility, outperforming existing fair graph generation methods while maintaining scalability.
Paper Structure (34 sections, 19 equations, 6 figures, 10 tables)

This paper contains 34 sections, 19 equations, 6 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Notations and Preliminaries
  4. Notations
  5. Graph Diffusion Model
  6. Forward Diffusion Process (Noise Injection).
  7. Backward Denoising Process (Graph Generation).
  8. Causal Model
  9. Diffusion-based Fair Graph Generation
  10. Overview.
  11. Bias Analysis and Targeted Fairness
  12. Unbiased Counterfactual Treatment with Causal Model
  13. Construction of Causal Model.
  14. Treatment Variable.
  15. Counterfactual Treatment.
  16. ...and 19 more sections

Figures (6)

  • Figure 1: Graphs generated by our FairGDiff achieve the best utility-fairness tradeoff in both supervised and unsupervised settings. Left: supervised node classification task. Right: unsupervised contrastive learning evaluated by downstream link prediction task.
  • Figure 2: Construction of Causal Model for FairGDiff.
  • Figure 3: Overview of the proposed FairGDiff, which consists two steps of Constructing Counterfactual Treatment and FairGDiff training, illustrated with gender as the sensitive attribute.
  • Figure 4: Graph Property Comparisons on NBA Dataset dai2021say. Relative differences from input graph are calculated as follows as: $\frac{|M(G^F)-M(\hat{G})|}{|M(G^F)|}$, where $G^F$ is the input graph, $\hat{G}$ is the generated graph, and $M$ stands for a certain metric shown in Table \ref{['tab:metric_computations']}.
  • Figure 5: Parameter sensitivity analysis of FairGDiff on NBA.
  • ...and 1 more figures