Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FairWire: Fair Graph Generation

O. Deniz Kose, Yanning Shen

TL;DR

The work addresses structural bias in graph-based learning, including in synthetic graphs generated for privacy and scalability. It develops a theory-guided fairness regularizer $\mathcal{L}_{\text{FairWire}}$ and a diffusion-based FairWire framework to mitigate bias in both link prediction and graph generation, accommodating non-binary sensitive attributes. The approach leverages a batch-aware regularizer and a diffusion-based denoiser (MPNN) guided by synthetic sensitive attributes to prevent sensitive information leakage. Empirical results on real networks and synthetic graphs show improved fairness metrics $\Delta_{\mathrm{SP}}$ and $\Delta_{\mathrm{EO}}$ with little to no sacrifice in utility, demonstrating practical impact for fair graph analytics and privacy-preserving synthetic data generation.

Abstract

Machine learning over graphs has recently attracted growing attention due to its ability to analyze and learn complex relations within critical interconnected systems. However, the disparate impact that is amplified by the use of biased graph structures in these algorithms has raised significant concerns for the deployment of them in real-world decision systems. In addition, while synthetic graph generation has become pivotal for privacy and scalability considerations, the impact of generative learning algorithms on the structural bias has not yet been investigated. Motivated by this, this work focuses on the analysis and mitigation of structural bias for both real and synthetic graphs. Specifically, we first theoretically analyze the sources of structural bias that result in disparity for the predictions of dyadic relations. To alleviate the identified bias factors, we design a novel fairness regularizer that offers a versatile use. Faced with the bias amplification in graph generation models that is brought to light in this work, we further propose a fair graph generation framework, FairWire, by leveraging our fair regularizer design in a generative model. Experimental results on real-world networks validate that the proposed tools herein deliver effective structural bias mitigation for both real and synthetic graphs.

FairWire: Fair Graph Generation

TL;DR

The work addresses structural bias in graph-based learning, including in synthetic graphs generated for privacy and scalability. It develops a theory-guided fairness regularizer and a diffusion-based FairWire framework to mitigate bias in both link prediction and graph generation, accommodating non-binary sensitive attributes. The approach leverages a batch-aware regularizer and a diffusion-based denoiser (MPNN) guided by synthetic sensitive attributes to prevent sensitive information leakage. Empirical results on real networks and synthetic graphs show improved fairness metrics and with little to no sacrifice in utility, demonstrating practical impact for fair graph analytics and privacy-preserving synthetic data generation.

Abstract

Machine learning over graphs has recently attracted growing attention due to its ability to analyze and learn complex relations within critical interconnected systems. However, the disparate impact that is amplified by the use of biased graph structures in these algorithms has raised significant concerns for the deployment of them in real-world decision systems. In addition, while synthetic graph generation has become pivotal for privacy and scalability considerations, the impact of generative learning algorithms on the structural bias has not yet been investigated. Motivated by this, this work focuses on the analysis and mitigation of structural bias for both real and synthetic graphs. Specifically, we first theoretically analyze the sources of structural bias that result in disparity for the predictions of dyadic relations. To alleviate the identified bias factors, we design a novel fairness regularizer that offers a versatile use. Faced with the bias amplification in graph generation models that is brought to light in this work, we further propose a fair graph generation framework, FairWire, by leveraging our fair regularizer design in a generative model. Experimental results on real-world networks validate that the proposed tools herein deliver effective structural bias mitigation for both real and synthetic graphs.
Paper Structure (22 sections, 3 theorems, 25 equations, 1 figure, 7 tables)

This paper contains 22 sections, 3 theorems, 25 equations, 1 figure, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Inspection and Mitigation of Structural Bias
  5. Bias Analysis
  6. A Regularizer for Fair Connections
  7. Fair Graph Generation
  8. Diffusion Models and Structural Bias
  9. FairWire: A Fair Graph Generation Framework
  10. Experiments
  11. Datasets and Experimental Setup
  12. Link Prediction Results
  13. Results for Graph Generation
  14. Visualization of Synthetic Graphs
  15. Conclusion
  16. ...and 7 more sections

Key Result

Theorem 1

The disparity between the representations of nodes in a sensitive group $\mathcal{S}_{k}$ and the representations of the remaining nodes output by the $l$th GNN layer, $\delta_{h}^{(l+1)}$, can be upper bounded by: where $L$ is the Lipschitz constant of the activation function $\sigma (\cdot)$, $\left\|\mathbf{z}^{l+1}_i- {\rm mean}(\mathbf{z}^{l+1}_j \mid v_{j} \in \mathcal{V})\right\|_{\infty}

Figures (1)

  • Figure 1: Distribution of the intra-edges (blue) and inter-edges (red) in the synthetic graphs created for Cora dataset by GraphMaker li2023graphmaker (left) and FairWire (right).

Theorems & Definitions (3)

  • Theorem 1
  • Proposition 1
  • Corollary 2