Utility-Preserving Federated Graph Learning with Dual-Perspective Fairness

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Federated graph neural networks (FedGNNs) face challenges in achieving fairness from both server (global) and client (local) perspectives simultaneously.
• Existing fairness-aware methods often sacrifice model utility (performance) to achieve fairness in distributed learning settings.
• The utility sacrifice is particularly pronounced in federated frameworks, exacerbating fairness challenges.

Purpose of the Study:

• To propose F3GL, a dual-perspective fairness federated graph learning method.
• To enhance both global and local fairness in FedGNNs while preserving model utility.
• To provide theoretical insights into fairness preservation using spectral graph theory.

Main Methods:

• Developed F3GL, a novel dual-perspective fairness federated graph learning approach.
• Conducted theoretical analysis using spectral graph theory to understand feature similarity after convolution.
• Identified the principal eigenvalue's role in enhancing feature similarity and applied a specialized eigenvalue selection strategy.

Main Results:

• Demonstrated that the principal eigenvalue is key to enhancing similarity between original and convolved sensitive features.
• Showcased that the theoretical findings are universally applicable to both clients and servers in federated learning.
• Experimental results on real-world datasets confirm F3GL's superiority over existing methods in improving dual-perspective fairness without utility loss.

Conclusions:

• F3GL effectively enhances dual-perspective fairness (global and local) in federated graph learning.
• The proposed method preserves utility, overcoming a key limitation of prior fairness-aware approaches.
• Spectral graph theory provides a robust theoretical foundation for achieving fairness in FedGNNs.