FedEBM: Robust graph federated learning via energy-based model.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Graph Federated Learning (GFL) is crucial for real-world applications but struggles with noisy labels in distributed data.
• Existing centralized methods for noisy labels are ineffective in distributed GFL settings and large-scale datasets.

Purpose of the Study:

• To propose a novel method, FedEBM, for effectively handling noisy labels in Graph Federated Learning.
• To leverage Energy-Based Models (EBM) to improve GFL performance under data sparsity and imbalance.

Main Methods:

• Developed FedEBM, integrating an Energy-Based Model (EBM) into the GFL framework.
• The EBM discriminates between clean and noisy samples using energy scores, even with scarce clean data.
• The EBM avoids probability competition, enhancing sensitivity to minority class features.

Main Results:

• FedEBM significantly outperforms six baseline methods across various noise rates, types, and client numbers.
• Achieved an average performance improvement of 5.43% on small-scale and 12.58% on large-scale datasets compared to the second-best method.

Conclusions:

• FedEBM offers a robust solution for noisy label challenges in Graph Federated Learning.
• The proposed EBM-based approach enhances GFL's reliability and performance in real-world distributed scenarios.