AI Learns to Remove Bias from Social Networks

Graph neural networks (GNNs) are powerful tools for analyzing connected data, like social networks or recommendation systems, but they often amplify existing biases, leading to unfair outcomes in critical areas such as loan approvals or criminal sentencing. A new study introduces FairMIB, a method that tackles this issue by decomposing graph data into distinct views to isolate and mitigate bias sources, achieving state-of-the-art results in both accuracy and fairness on real-world datasets.

The researchers found that FairMIB reduces bias significantly compared to standard GNNs. For example, on the German credit dataset, it cut demographic parity disparity by 98.8% and equal opportunity disparity by 99.3% relative to a vanilla graph convolutional network (GCN), while maintaining competitive predictive performance. This means the model makes decisions that are fairer across different demographic groups without sacrificing utility.

To achieve this, the method separates graph data into three independent views: a feature view focusing on node attributes, a structural view capturing pure connectivity patterns, and a diffusion view modeling how information spreads through the network. Each view is processed by dedicated encoders, and an inverse probability-weighted correction is applied to the diffusion view to prevent bias amplification during message passing. The approach uses a multi-view information bottleneck to compress these views into a fair representation, minimizing sensitive information leakage while preserving task-relevant data.

Experimental results on datasets like Bail, Credit, Pokec-z, and Pokec-n show that FairMIB outperforms seven state-of-the-art baselines. On the Bail dataset, it improved F1-score by 3.2% and reduced demographic parity disparity by 82.3% compared to GCN. Ablation studies confirmed that removing any component—such as the information compression or multi-view consistency—led to declines in fairness and utility, highlighting the necessity of the integrated design.

This advancement matters because it addresses real-world risks where biased AI can perpetuate discrimination in high-stakes applications. By ensuring fairer predictions, FairMIB could help build trust in AI systems used in finance, justice, and social media, promoting equitable outcomes for diverse populations. However, the study notes limitations, including that the current framework handles single sensitive attributes and may need extensions for multiple, intersecting biases or more complex graph scenarios.