Identifying and Correcting Label Noise for Robust GNNs via Influence Contradiction
Researchers have developed a new method called ICGNN to improve the robustness of Graph Neural Networks (GNNs) when dealing with noisy labels. The approach uses a novel noise indicator, the influence contradiction score (ICS), derived from graph diffusion, to identify potentially mislabeled nodes. A Gaussian mixture model is then employed for precise noise detection, followed by a soft strategy to correct these labels using neighboring node predictions. The method also incorporates pseudo-labeling to enhance supervision signals for unlabeled nodes, with experiments demonstrating its effectiveness over existing baselines. AI
IMPACT Improves the reliability of GNNs in real-world applications by addressing data quality issues.