Towards robust Graph Neural Networks: A comprehensive strategy against graph and label noise

Abstract

The increasing use of graph data in emerging domains has highlighted the importance of Graph Neural Networks (GNNs). However, traditional GNNs are highly susceptible to both graph and label noise, leading to degraded and unstable performance under noisy conditions. To address this challenge, we propose a R obust S tructure and L abel D enoising G raph N eural N etwork (RSLD-GNN) that jointly refines graph structure and node labels while decoupling their optimization. RSLD-GNN performs structural refinement by combining dynamic and static graph augmentation with knowledge distillation, enabling robust exploration and stabilization of the graph structure. For label denoising, a detached label head isolates label correction from the main computational graph, reducing the influence of structural noise on label refinement and generating more reliable pseudo-labels. These two refinement processes are further connected through mutual regularization: loss-based edge reweighting suppresses connections involving nodes with potentially noisy labels, while the resulting cleaner structure facilitates more reliable label propagation. All modules are pretrained under controlled noise to reduce sensitivity to noise during early optimization. Extensive experiments on five benchmark datasets under single-graph, single-label, and mixed-noise settings show that RSLD-GNN ranks first in 69 of 80 noisy evaluation settings while remaining competitive on clean datasets.