logo
CanyamAI summaries for academic research
language
MobileMobile
  • Phone FeatureWeChat Mini Program
EnglishEnglish
noticeNotice
BlogBlog
delete
arrow
Return

Non-aligned multi-view partial multi-label learning via factor group-sparse regularization

delete2025-12-06
delete0
PRE
AI
Z
Ziyi Wang
Y
Yitian Xu *
DOI：10.1016/j.neucom.2025.132287delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
In complex real-world scenarios, imperfect data are pervasive, among which multi-view partial multi-label learning (MVPML) and non-aligned views represent two typical manifestations. In MVPML, each training sample described by multiple heterogeneous views is associated with a set of candidate labels, which includes an unknown number of ground-truth labels mixed with noisy labels. Although existing methods attempt to recover the ground-truth labels through disambiguation strategies, most of them rely heavily on the assumption of view alignment, limiting their applicability to heterogeneous and non-aligned data. To address the dual challenges of label ambiguity and view non-alignment, we propose a novel framework, termed nMVPML-FGS. At the feature level, the method employs kNN to capture manifold structures from non-aligned data and constructs feature-induced manifold structures in the label space, thereby alleviating label ambiguity caused by non-aligned view. Building upon this, the framework is reconstructed under the supervision guided by disambiguation results, preserving the feature diversity of each view while integrating complementarity and consistency across multi-view and exploring the latent low-rank correlations among labels. Moreover, to enhance low-rank modeling and shared information discovery across views, we incorporate a factor group sparse regularizer (FGSR). As a non-convex surrogate of the matrix rank function, FGSR enables more accurate characterization of latent low-rank structures and achieves superior computational efficiency. Finally, the method achieves joint optimization within a unified framework. Comprehensive experiments on five datasets demonstrate that nMVPML-FGS outperforms existing methods across multiple evaluation metrics, with its performance advantages being statistically significant.

Journal

Neurocomputing cover
Neurocomputing
IF:
6.5
Papers:
2.5W
Citations:
6.5W

Organization

No organization information available