PaSta Improves Node Classification with Noisy Labels

Node classification in real-world graph-based systems often suffers from noisy or unreliable labels, leading to models that overfit or accumulate errors during pseudo-labeling. This new research proposes PaSta, a Partial label-based Self-training framework, to tackle these challenges. PaSta begins by training multiple "annotators" to capture diverse class distributions, then aggregates their predictions to create high-quality partial labels for nodes. The framework then uses these partial labels to guide a classification model, employing two specialized loss functions for both label and representation spaces. To further enhance robustness, PaSta incorporates an iterative self-training loop where the refined partial labels continuously optimize the annotators. Experiments on five datasets demonstrate that PaSta consistently improves classification performance, averaging a 1.1% gain over state-of-the-art methods under various noise conditions.