Robust graph learning under label noise

Develop robust graph-learning methods that reduce the impact of noisy labels during training, thereby addressing the unresolved challenge of achieving reliable graph neural network performance under corrupted supervision.

Background

Graph neural networks, including graph convolutional networks, are vulnerable to label noise because message passing can propagate incorrect supervision through neighboring node representations. The paper situates its PCC+GCN framework within broader efforts to improve robustness through label correction, sample selection, modified training objectives, and topology-aware refinement.

The authors explicitly characterize robust graph learning under label noise as an unresolved challenge. Although the proposed PCC+GCN method improves performance in the reported experiments, the broader problem remains open beyond the evaluated datasets, noise models, and preprocessing strategy.