.png){kind=icon}
.png){kind=icon}
Return
Robust semantic reconstruction for weakly supervised multi-label learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.133422.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In weakly supervised multi-label learning tasks, only part of the training data has fully known label knowledge, while the remaining samples suffer from partially missing or even completely missing labels. Existing weakly supervised multi-label learning methods utilize self-representation learning of label correlation and label propagation strategies for graph embedding to achieve correction of partially missing labels and infer completely missing labels. However, these self-representation learning approaches ignore the dependencies between features and labels during the label reconstruction process. Moreover, most graph-embedded label propagation methods are built on the smoothness assumption, which assumes that the graph structures of feature and label samples are consistent. Such hard constraints, however, are detrimental to preserving the intrinsic manifold structure of the data and may result in low-quality graphs. We propose a robust weakly supervised multi-label learning method to address these issues. Specifically, firstly, we utilize semantic self-representation learning and feature-label dependency maximization metric strategies to achieve the correction of partially missing labels and employ an l2,p-norm regularization term to enhance its robustness. Secondly, we use residual terms to relax the graph structure consistency constraints and employ flexible sparse neighborhood graph learning to train sample nonlinear embeddings for label information propagation, thereby leading to inference of unknown labels. Finally, we obtain ground-truth labels within a unified learning framework and develop an alternating optimization solution in which the iterative optimization process promotes the interplay between self-representation learning and graph embedding processes. Extensive experiments on various real-world tasks show that the proposed method outperforms some state-of-the-art approaches.
Keywords:
weakly supervised multi-label learning
semantic self-representation
feature-label dependency
graph embedding
label propagation
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W