Toward Fair Federated Graph Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Federated graph learning (FGL) enables privacy-preserving distributed training of graph neural networks (GNNs).
• Subgraph-FL, a dominant FGL paradigm, often overlooks fairness, leading to biased performance against minority nodes and those with heterophilous connections.
• Existing methods prioritize overall accuracy, masking degraded performance on marginalized node groups.

Purpose of the Study:

• To address fairness issues in subgraph-FGL by enhancing representation for minority nodes and mitigating topological biases.
• To propose FairFGL, a novel framework for fair federated graph learning.
• To improve robustness and performance for structurally or semantically marginalized nodes in FGL.

Main Methods:

• FairFGL employs client-side modules: majority alignment for efficient knowledge transfer, gradient modification, and history preservation to infuse minority knowledge and prevent overfitting.
• Server-side aggregation uses a cluster-based strategy to reconcile heterogeneous updates and suppress global majority dominance.
• The framework focuses on fine-grained graph property mining and collaborative learning.

Main Results:

• FairFGL significantly improves performance for disadvantaged node groups across eight benchmark datasets.
• Achieved up to a 21.07% increase in Overall F1-score for marginalized groups.
• Demonstrated enhanced convergence efficiency compared to state-of-the-art baselines.

Conclusions:

• FairFGL effectively enhances fairness in federated graph learning by addressing class-wise and topology-aware biases.
• The proposed framework improves model robustness and performance for underrepresented node groups.
• FairFGL offers a promising direction for developing equitable and high-performing distributed graph learning systems.