.png){kind=icon}
.png){kind=icon}
Return
Nonlinear perturbed label-based graph regularized multi-label learning with missing labels
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.115093.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Most existing approaches to multi-label learning with incomplete labels impose manifold regularization on classifier weights rather than directly on the label structure. This guidance influences missing-label recovery indirectly through the learned feature-label mapping, often leading to inaccurate imputations, especially under high missing-label rates. To address this issue, we propose NPLGR (Nonlinear Perturbed Label-based Graph Regularization), a synchronous recovery framework driven by label-structure-based manifold regularization. NPLGR directly regularizes the label correlation matrix using a label similarity graph and incorporates a perturbation strategy to mitigate the instability caused by sparse co-occurrences. Moreover, instance-level correlations and smooth nonlinear transformations are integrated to capture instance similarities and complex feature-label dependencies. Extensive experiments on benchmark datasets demonstrate that NPLGR consistently outperforms state-of-the-art methods. This improvement is especially pronounced in scenarios with high label incompleteness, offering enhanced robustness and semantic accuracy in multi-label learning tasks.
Keywords:
Multi-label learning
Incomplete labels
Label-structure regularization
Nonlinear transformation
Perturbation strategy