Title: Enhancing GNN Robustness Through Influence Contradiction for Label Noise Identification and Correction

Abstract:

Graph Neural Networks (GNNs) have demonstrated exceptional proficiency in processing graph-structured data, enabling diverse applications ranging from social network analysis to bioinformatics. Nevertheless, the prevalence of label noise in practical settings presents a major obstacle to training robust GNNs. When graphs contain noisy labels—typically arising from annotation mistakes or inconsistencies—the performance of these models can degrade substantially. To mitigate this issue, this study introduces ICGNN, a novel methodology that utilizes graph structural information to counteract the adverse effects of noisy labels.

Our approach begins with the design of a distinctive noise indicator that calculates an influence contradiction score (ICS). Derived from the graph diffusion matrix, this metric evaluates the reliability of nodes possessing clean labels; specifically, nodes exhibiting higher ICS values are more probable candidates for containing noisy labels. Following this, a Gaussian mixture model is employed to accurately determine whether a specific node’s label is corrupted. Furthermore, we implement a soft aggregation strategy that integrates predictions from a node’s neighbors to rectify the identified noisy labels. Finally, the framework incorporates pseudo-labeling for the extensive set of unlabeled nodes, thereby supplying additional supervisory signals to steer model optimization. Empirical evaluations on standard benchmark datasets confirm that our proposed method outperforms existing competitive baselines in environments characterized by label noise.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC