.png){kind=icon}
.png){kind=icon}
Return
UniTrain: A universal iterative semi-supervised training framework for graph representation learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2026.108623.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph neural networks (GNNs) and graph transformers (GTs) perform well in graph-related tasks, but their potential is often limited in semi-supervised settings due to label scarcity. Although robust encoders and pre-training tasks enhance performance, GNNs or GTs remain prone to over-fitting and task gaps. To address these issues, we propose the Universal iterative semi-supervised Training framework (UniTrain), which generates high-quality pseudo-labels for unlabeled nodes in a multi-stage manner. Within UniTrain, a semantic graph is constructed using hidden vector representations from pre-training stage. Label knowledge propagation with uncertainty filtering is then applied to infer or refine labels for unlabeled nodes. By incorporating high-confidence pseudo-labels, the framework mitigates noise and compensates for the limited guidance of original labels. UniTrain enhances fine-tuning in existing self-supervised methods and is compatible with any GNN or GT encoder. Extensive experiments conduct on seven graph-based benchmarks (Cora, Citeseer, Pubmed, Actor, Cornell, Texas, and Wisconsin) demonstrate significant improvements in node classification performance when applying UniTrain to both GNNs and GTs. These results validate the effectiveness and generalization capability of our method.