Multi-view anchor graph regularization induced unbiased learning for positive and unlabeled problem

In contrast to expensive and mundane annotations, it is wise to collect data from diverse sources or present it in multiple ways. Consequently, the scarcity of labeled data and the richness of data representation foster advancements in multi-view semi-supervised learning. Nevertheless, numerous real...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Applied soft computing Ročník 180; s. 113287
Hlavní autoři: Zhao, Jie, Li, Xiao, Xu, Yitian, Yuan, Min
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 01.08.2025
Témata:
Anchor graph
Loss function factorization
Multi-view learning
Positive and unlabeled learning
Loss function factorization
Positive and unlabeled learning
Anchor graph
Multi-view learning
ISSN:1568-4946
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:In contrast to expensive and mundane annotations, it is wise to collect data from diverse sources or present it in multiple ways. Consequently, the scarcity of labeled data and the richness of data representation foster advancements in multi-view semi-supervised learning. Nevertheless, numerous real-world scenarios often present a formidable challenge, namely the Positive and Unlabeled (PU) problem, where negative label information is absent. Most semi-supervised algorithms struggle to effectively address PU problems, and a universal model for multi-view positive and unlabeled (PU) learning remains lacking. Hence, in this paper, we propose a Multi-view Anchor Graph Regularization-induced unbiased method for PU problems (MvAGRPU). In MvAGRPU, we transform multi-view PU problems into supervised problems with label noise, learning unbiased classifiers through loss factorization and centroid approximation. Additionally, we devise a regularization term tailored for PU problems, which integrates anchor graph, local, and semantic information to facilitate interplay among diverse views. Subsequently, we devise an efficient algorithm to tackle the resultant optimization problem. Finally, we compare MvAGRPU against five PU algorithms on multiple datasets, validating its superiority and stability. •MvAGRPU transforms multi-view PU to supervised with one-sided noise.•MvAGRPU eliminates noise impact for an unbiased estimate of empirical risk.•MvAGRPU Supports various loss functions for flexibility.•The proposed regularization integrates anchor, local and semantic information.
ISSN:1568-4946
DOI:10.1016/j.asoc.2025.113287