.png){kind=icon}
Return
GraphUAT: An uncertainty-driven pseudo-labeling method for class-imbalanced node classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.133450.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph Neural Networks (GNNs) have achieved remarkable success across various graph analysis tasks. However, in real-world scenarios, graphs typically exhibit a class-imbalanced distribution, where a few classes account for the majority of nodes. This distribution leads to a bias toward majority classes and degrades the model's discriminative ability on minority classes. It is a promising solution to leverage the numerous unlabeled samples to rebalance the class distribution. However, existing pseudo-labeling methods fail to select reliable unlabeled minority nodes because they directly rely on GNNs trained on imbalanced data. To address this issue, we propose GraphUAT, an uncertainty-driven pseudo-labeling method for class-imbalanced graphs. It incorporates uncertainty estimation with class-wise adaptive thresholds to select reliable pseudo-labeled minority samples. Specifically, we first quantify the prior misclassification risk of each node by jointly assessing uncertainty from the perspectives of feature and topology, thereby improving the reliability of pseudo-labels. Furthermore, we design a class-wise adaptive thresholding strategy that dynamically distributes an independent pseudo-label threshold for each class, enabling a biased pseudo-label sampling for the minority. Extensive experiments on six real-world datasets demonstrate that our method consistently outperforms existing baselines on the task of class-imbalanced node classification. Our codes are available at https://github.com/ZZY-GraphMiningLab/GraphUAT.
Keywords:
Class-imbalanced graph
Node classification
Adaptive threshold
Uncertainty
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
Organization
No organization information available
Cited Papers
No cited papers available