Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Equipping Federated Graph Neural Networks with Structure-aware Group Fairness

Nan Cui, Xiuling Wang, Wendy Hui Wang, Violet Chen, Yue Ning

TL;DR

This work tackles fairness in federated graph neural networks by examining how data bias in graph structure and learning bias from message passing affect group fairness. It introduces F^2GNN, a structure-aware framework that couples fairness-aware local updates with fairness-weighted global aggregation, linking edge balance to SP and EO via a point-biserial correlation analysis. The method jointly optimizes local fairness through JS-based interpolation and penalties, and global fairness via data-bias and model-fairness weights in aggregation, showing improved accuracy and notably reduced fairness gaps on Pokec-z, Pokec-n, and NBA datasets. The results demonstrate the feasibility of mitigating bias in federated GNNs without sacrificing performance, and point to future work on privacy-preserving reporting of balance statistics.

Abstract

Graph Neural Networks (GNNs) have been widely used for various types of graph data processing and analytical tasks in different domains. Training GNNs over centralized graph data can be infeasible due to privacy concerns and regulatory restrictions. Thus, federated learning (FL) becomes a trending solution to address this challenge in a distributed learning paradigm. However, as GNNs may inherit historical bias from training data and lead to discriminatory predictions, the bias of local models can be easily propagated to the global model in distributed settings. This poses a new challenge in mitigating bias in federated GNNs. To address this challenge, we propose $\text{F}^2$GNN, a Fair Federated Graph Neural Network, that enhances group fairness of federated GNNs. As bias can be sourced from both data and learning algorithms, $\text{F}^2$GNN aims to mitigate both types of bias under federated settings. First, we provide theoretical insights on the connection between data bias in a training graph and statistical fairness metrics of the trained GNN models. Based on the theoretical analysis, we design $\text{F}^2$GNN which contains two key components: a fairness-aware local model update scheme that enhances group fairness of the local models on the client side, and a fairness-weighted global model update scheme that takes both data bias and fairness metrics of local models into consideration in the aggregation process. We evaluate $\text{F}^2$GNN empirically versus a number of baseline methods, and demonstrate that $\text{F}^2$GNN outperforms these baselines in terms of both fairness and model accuracy.

Equipping Federated Graph Neural Networks with Structure-aware Group Fairness

TL;DR

This work tackles fairness in federated graph neural networks by examining how data bias in graph structure and learning bias from message passing affect group fairness. It introduces F^2GNN, a structure-aware framework that couples fairness-aware local updates with fairness-weighted global aggregation, linking edge balance to SP and EO via a point-biserial correlation analysis. The method jointly optimizes local fairness through JS-based interpolation and penalties, and global fairness via data-bias and model-fairness weights in aggregation, showing improved accuracy and notably reduced fairness gaps on Pokec-z, Pokec-n, and NBA datasets. The results demonstrate the feasibility of mitigating bias in federated GNNs without sacrificing performance, and point to future work on privacy-preserving reporting of balance statistics.

Abstract

Graph Neural Networks (GNNs) have been widely used for various types of graph data processing and analytical tasks in different domains. Training GNNs over centralized graph data can be infeasible due to privacy concerns and regulatory restrictions. Thus, federated learning (FL) becomes a trending solution to address this challenge in a distributed learning paradigm. However, as GNNs may inherit historical bias from training data and lead to discriminatory predictions, the bias of local models can be easily propagated to the global model in distributed settings. This poses a new challenge in mitigating bias in federated GNNs. To address this challenge, we propose GNN, a Fair Federated Graph Neural Network, that enhances group fairness of federated GNNs. As bias can be sourced from both data and learning algorithms, GNN aims to mitigate both types of bias under federated settings. First, we provide theoretical insights on the connection between data bias in a training graph and statistical fairness metrics of the trained GNN models. Based on the theoretical analysis, we design GNN which contains two key components: a fairness-aware local model update scheme that enhances group fairness of the local models on the client side, and a fairness-weighted global model update scheme that takes both data bias and fairness metrics of local models into consideration in the aggregation process. We evaluate GNN empirically versus a number of baseline methods, and demonstrate that GNN outperforms these baselines in terms of both fairness and model accuracy.
Paper Structure (24 sections, 4 theorems, 28 equations, 2 figures, 5 tables)

This paper contains 24 sections, 4 theorems, 28 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Federated Graph Learning
  4. Group Fairness and Data Bias
  5. Connection between Data Bias and Model Fairness
  6. The Proposed Method: $\text{F}^2$GNN
  7. Client-side Fairness-aware Local Model Update
  8. Server-side Fairness-weighted Global Model Update
  9. Experimental Evaluation
  10. Experimental Setup
  11. Performance Evaluation
  12. Ablation Study
  13. Conclusion
  14. Theoretical Analysis
  15. Calculation of $\mu_0$ and $\mu_1$
  16. ...and 9 more sections

Key Result

Lemma 5

Given a graph $\mathcal{G}$ and a GNN model that satisfies the data and model assumptions respectively, the Point-biserial correlation (denoted as $\rho_{\mathbf{Z},s}$) between the graph representation $\mathbf{Z}$ and the sensitive attribute $s$ is measured as the following: where $H_{\text{intra}}$ and $H_{\text{inter}}$ are defined by Definition def:edge_balance, $N_0$$(N_1)$ is the number of

Figures (2)

  • Figure 1: An overview of $\text{F}^2$GNN. In each iteration, client $C_i$ receives the global model, calculates the Jensen-Shannon (JS) divergence $\mathrm{js}^i_t$, and updates the local model $\hat{\omega}^i_t$. Local data is processed through a 2-layer GCNs, and the model is trained over several epochs. The updated model $\omega^i_t$ and the group balance score $B_i$ are uploaded to the server. The server then computes the data-bias and model-fairness weights, integrates them, and uses the combined weight to aggregate the local models and update the global model $\omega_t$.
  • Figure 2: Ablation study of the proposed method on three datasets.

Theorems & Definitions (10)

  • Definition 1
  • Definition 2: Statistical Parity
  • Definition 3: Equalized Odds
  • Definition 4: Point-Biserial Correlation Coefficient
  • Lemma 5
  • Theorem 6
  • Lemma
  • proof
  • Theorem
  • proof