Migrate demographic group for fair Graph Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph Neural Networks (GNNs) demonstrate superior performance in graph learning tasks.
• Existing fairness techniques in GNNs often rely on fixed sensitive attributes, failing to address underlying biases.
• Biased information in training data can lead to unfair outcomes for specific demographic groups.

Purpose of the Study:

• To address the limitations of existing fair GNN techniques.
• To propose a novel framework, FairMigration, for dynamic demographic group adjustment.
• To improve the trade-off between model performance and fairness in GNNs.

Main Methods:

• FairMigration employs a two-stage training process.
• Stage 1: Initial GNN optimization with personalized self-supervised learning and dynamic demographic group adjustment.
• Stage 2: Supervised learning with frozen demographic groups, incorporating adversarial training.

Main Results:

• FairMigration effectively migrates demographic groups dynamically.
• The framework achieves a favorable balance between model performance and fairness.
• Extensive experiments validate the efficacy of the proposed approach.

Conclusions:

• FairMigration offers a new paradigm for achieving fairness in GNNs.
• Dynamic adjustment of demographic groups is crucial for mitigating bias.
• The framework shows promise for developing more equitable AI systems.