FedRGL: Robust Federated Graph Learning for Label Noise

TL;DR

FedRGL tackles label noise in federated graph learning on subgraphs by integrating dual-consistency noise filtering with class-aware thresholds and graph contrastive learning for high-quality supervision, followed by entropy-based robust aggregation to reweight client contributions. The method combines a global model view and a local structural view to precisely identify noisy nodes, augments training with high-confidence pseudo-labels, and reweights server aggregation using predictive entropy of unlabeled nodes to suppress noisy clients. Key contributions include the first robust approach for label-noise in federated subgraph learning, a class-aware dynamic-threshold noise filtering mechanism, and a server-side aggregation strategy that improves robustness without requiring clean-label priors. Empirical results across six real-world graphs and large-scale data demonstrate consistent, substantial improvements over 12 baselines under varying noise rates, types, and client counts, highlighting FedRGL’s practical impact for privacy-preserving graph learning.

Abstract

Federated Graph Learning (FGL) is a distributed machine learning paradigm based on graph neural networks, enabling secure and collaborative modeling of local graph data among clients. However, label noise can degrade the global model's generalization performance. Existing federated label noise learning methods, primarily focused on computer vision, often yield suboptimal results when applied to FGL. To address this, we propose a robust federated graph learning method with label noise, termed FedRGL. FedRGL introduces dual-perspective consistency noise node filtering, leveraging both the global model and subgraph structure under class-aware dynamic thresholds. To enhance client-side training, we incorporate graph contrastive learning, which improves encoder robustness and assigns high-confidence pseudo-labels to noisy nodes. Additionally, we measure model quality via predictive entropy of unlabeled nodes, enabling adaptive robust aggregation of the global model. Comparative experiments on multiple real-world graph datasets show that FedRGL outperforms 12 baseline methods across various noise rates, types, and numbers of clients.

Figures (5)

• Figure 1: Example of Federated Graph Learning with noisy labels. Subgraph data is divided among clients using the Louvain algorithm blondel2008fast, with varying noise levels. The collaboratively trained global model is tested for final performance. , , and represent training nodes, while represents test nodes. Testing accuracy of three FGL algorithms on Cora under uniform label noise at different rates (10 clients), showing lack of robustness to label noise.
• Figure 2: Overall framework diagram of the FedRGL method.
• Figure 3: Accuracy on Cora under different levels of label noise and numbers of clients.
• Figure 4: Visualization of the training curves and ablation experiment.
• Figure 5: Analysis on hyper-parameter in FedRGL.