PaSta, a graph-learning approach, was reported to achieve the best performance across four datasets when labels were deliberately made noisy. Its average accuracy was reported as 1.1% higher than BO-NNC and 12.7% higher than GCN. Those figures are benchmark averages.
The work addresses noisy node classification: deciding which class a node belongs to when the labels used for learning are unreliable. The paper describes PaSta as a Partial label-based Self-training framework for noisy node classification.
A model built for uncertain labels
PaSta combines self-supervised ensemble annotation with dual-space partial-label learning. Its framework also uses self-training to refine the learning process.
The experiments compared versions with and without self-training and tested different combinations of the framework's components. A multi-annotator strategy was reported to improve accuracy by 7.5% over DGI used alone, 6.9% over GCA used alone and 6.1% over SUGRL used alone to generate partial labels.
Five datasets, two kinds of injected noise
The evaluation used five real-world graph datasets: Cora, CiteSeer, DBLP, Computers and Photo. The comparison set included GCN, GAT, DGI, GCA, SUGRL, JoCoR, NRGNN, MTS-GNN and BO-NNC.
Cora and CiteSeer used standard splits. DBLP, Computers and Photo used 5% of nodes for training, 10% for validation and 60% for testing, with uniform and pair noise injected at different rates.
That design tests two deliberately injected noise patterns. It does not settle how PaSta would perform with naturally occurring label errors or on substantially larger graphs, questions that remain outside the supplied evaluation.
Where the reported gains appeared
The main four-dataset comparison described PaSta as the top performer. Photo results were reported separately in Appendix A: there, PaSta exceeded BO-NNC by 2.27% under uniform noise and 2.57% under pair noise. The appendix also gives Photo classification results at different noise rates.
In component tests, the full configuration was reported to improve average performance by 4.2% over versions containing any two components. Adding the PaSim and VaCE training losses was reported to yield an average improvement of 1.7%.
The self-training comparison reported an average improvement of 9.6% against methods that did not use self-training. These are comparisons between tested configurations, so the figures show differences between setups but do not by themselves establish that one component caused the entire gap.
Performance was reported to rise as the iteration count increased and to stabilize beyond three iterations. Training was said to converge after approximately three iterations, with little change reported after that point in the tested settings.
What the benchmark cannot answer
The preprint does not report statistical significance tests or confidence intervals. It also does not define the variability measure shown in parentheses with the reported accuracies, and the averaging procedure for the ablation results is not given in the supplied material.
The percentage differences therefore cannot be judged here for statistical significance. Nor do the results establish that PaSta will lead across other graph sizes, noise mechanisms, classifier architectures or annotation processes; the evidence is limited to the named datasets and injected noise settings.
An early report
The authors state that source code is available at the listed GitHub repository. The document is marked as arXiv version 1 dated 26 August 2026, so the results are presented as an early research report.
Paper data and sources
Original title: PaSta: Noisy Node Classification with Partial Label Learning
Authors: Yujing Liu, Yixin Liu, Yu Zheng et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text