GNN’s FAME: Fairness-Aware MEssages for Graph Neural Networks

This repository contains the code and experiments for the paper "GNN’s FAME: Fairness-Aware MEssages for Graph Neural Networks". The primary contribution of this work is the development of two novel in-processing and GNN-specific bias mitigation approaches, namely FAME (Fairness-Aware MEssages) and its variant A-FAME (Attention-based Fairness-Aware MEssages), designed for GCN-based and GAT-based architectures, respectively.
To read more on this, please refer to the short paper published during UMAP 2025 (or refer to the presentation slides).

Table of Contents

• Introduction
• Repository Structure
• Datasets

Introduction

Graph Neural Networks (GNNs) are powerful tools for learning representations of graph-structured data. However, GNNs are susceptible to biases that can arise from the underlying data, leading to unfair predictions. To address this issue, we propose two novel message-passing layers:

• FAME (Fairness-Aware Message Passing): This layer adjusts the messages during the aggregation phase based on the disparities in sensitive attributes of connected nodes.
• A-FAME (Attention Fairness-Aware Message Passing): This layer extends FAME by incorporating an attention mechanism to weigh the importance of node connections dynamically.

Note: enhanced_fame is a WIP extension of this repository.

These layers aim to ensure more equitable outcomes by mitigating bias propagation within GNNs.

Repository Structure

The repository contains:

• Dataset: Directory containing German, Credit, and Bail datasets.
• calculate_fairness.py: File containing the implementation of the fairness metrics calculation.
• fame.py: File containing the implementation of the FAME and A-FAME layers.
• main.py: File containing the implementation of the training and evaluation of the GNN models.
• model.py: File containing the implementation of the GNN models (GCN, GAT, FAME, A-FAME).
• preprocess_data.py: File containing the implementation of the data preprocessing.
• set_uid.py: File containing the implementation of the unique identifier setting for each node in the dataset.
• utils.py: File containing the implementation of the utility functions.

Datasets

Extract the datasets from the datasets folder.
Run the file preprocess_data.py to preprocess the datasets first, then run file set_uid.py to set the unique identifier for each node in the dataset.
The datasets adopted in the paper's evaluation can be found at the following links (or in the datasets folder - in the case of German and Credit):

• German
• Credit
• Pokec