MLL-NDLSR: A Unified Multi-Label Learning Method for Missing Labels
Kaixiang YANG , Haolin SHEN , Guojie LI , Guoxian YU , Zhiwen YU , C. L. Philip Chen
In the field of multi-label learning with missing labels (MLML), the incomplete supervision caused by missing labels can significantly degrade the model’s performance. Most existing multi-label learning methods extract useful supervisory information from the original feature space and primarily focus on exploring the linear relationship between features and labels. However, these approaches may overlook the non-linear relationship between features and labels, which can result in the insufficient utilization of supervisory information. Additionally, researchers often fail to account for potential noise in the features. To address these issues, we propose a unified multi-label learning method for missing labels based on noise feature decomposition and label secondary reconstruction, called MLL-NDLSR. Specifically, we first apply the kernel method to transform the original feature matrix into a kernel matrix, treating the kernel matrix as a new representation of the instances’ features. Then, we explore the non-linear relationship between features and labels in the kernel matrix, decompose the noise features, and use a non-linear function to normalize the classifier’s output, thereby enhancing the model’s ability to express non-linear relationships and increasing its robustness. Furthermore, we introduce the label secondary reconstruction method to restore missing labels, ensuring that supervisory information is recovered. Finally, we consider feature correlation and label correlation, designing a manifold learning term to guide the model’s learning process. Extensive experiments demonstrate the superiority of the proposed MLL-NDLSR over other advanced models.
multi-label learning / missing label / feature decomposition / label reconstruction / label correlation
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(1297KB)
