ISSMLCF: an inductive semi-supervised multi-label learning algorithm with co-forest paradigm

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 55; no. 11; p. 808
• Main Authors: Liu, Wenhao, Duan, Jicong, Shao, Changbin, Yang, Xibei, Xu, Sen, Yu, Hualong
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.07.2025; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Computer Science; Confidence; Forests; Hypotheses; Labeling; Labels; Machine learning; Machines; Manufacturing; Mechanical Engineering; Prediction models; Processes; Propagation; Semi-supervised learning; Teaching methods; Multi-label learning; Unlabeled data; Semi-supervised learning; Co-forest; Thresholds calibration
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-025-06688-8

Multi-label learning aims at training accurate predictive models to recognize the instance with multiple potential class labels. In this scenario, it is more difficult to collect sufficient labeled instances as each instance requires being annotated with several labels. Considering it is possibly easy to collect lots of unlabeled instances, but labeling them is expensive, the semi-supervised learning can be combined with multi-label data to be a leverage between the quality of predictive models and labeling costs. In this study, a novel semi-supervised multi-label learning algorithm called the inductive semi-supervised multi-label learning algorithm based on co-forest paradigm (ISSMLCF) was proposed. Specifically, the proposed ISSMLCF algorithm uses both instance bootstrap and random feature split techniques to promote the diversity among base learners in co-forest, adopts thresholds calibration strategy to improve the predictive performance of base learning model, and integrates three confidence measures, namely prediction disagreement, label cardinality consistency, and label correlation consistency, to communicate confidential instance and label information among classifiers during the iterative process of co-forest. To guarantee the stability and avoid an excessive of error accumulation during semi-supervised learning, a n -Top communication strategy based on confidence threshold is used by the ISSMLCF algorithm. Experimental results conducted on eight benchmark multi-label datasets show that the proposed ISSMLCF algorithm can not only produce better classification performance, but also consume less training time than several SOTA algorithms which include both transductive and inductive ones.