Get started
Work

Self-taught graph-based semi-supervised learning for multi-view image data

Year: 2026
Type: article
Abstract: In recent years, graph-based semi-supervised learning and pseudo-labeling have attracted significant attention for their effectiveness in addressing the challenges of expensive and time-consuming data annotations. Pseudo-labeling improves model training by using predictions from unlabeled data, while graph-based methods excel with graph-structured data. However, the absence of explicit graph structures in image data and the complexities of multi-view features pose considerable challenges. This paper introduces the Self-taught Graph-based Semi-supervised Learning for Multi-view Data (SGSLM), a novel approach that combines linear feature transformation and multi-view graph fusion within a Graph Convolutional Network (GCN) framework. Our method dynamically integrates pseudo-labels into the GCN’s loss function, allowing the model to leverage pseudo-labeling benefits for multi-view data classification. This adaptive pseudo-label loss mechanism offers several advantages: (i) it integrates seamlessly into the existing framework without requiring extra iterations or models; (ii) all predictions are considered pseudo-labels, with weights adjusted based on confidence levels, eliminating the need for threshold settings; and (iii) it reduces the impact of misleading predictions during early training phases, minimizing the risk of negative influences on the model. Additionally, our model includes a Dice loss for label alignment and an unsupervised smoothing loss, which together enhance overall performance. The entire framework is implemented within a single lightweight GCN framework, optimizing both time and space efficiency. Experimental results demonstrate its effectiveness compared to state-of-the-art methods. 1
DOI: 10.1016/j.patcog.2026.114731
Language: English
Open Access status: hybrid
Estimated APC paid: $2,590
OpenAlex ID: W7207559599
FWCI:
Cites: 25
Cited by:
Related to: 10