Semi-supervised multi-label feature selection via partial label correlation and feature self-representation

Abstract

In the field of multi-label feature selection (MLFS), semi-supervised learning can effectively reduce the labeling cost and alleviate the negative impacts caused by labeling noise. However, there are complex inherent correlations among labels in multi-label data. Existing semi-supervised MLFS methods fail to fully exploit the limited label information to assist the learning process of pseudo-labels, limiting the accuracy and reliability of pseudo-labels during model training. To address this issue, we design a manifold regularization term based on partial label correlations and integrate it with the instance manifold to jointly guide the learning process of pseudo-labels. In addition, we develop a sparse formulation for feature self-representation to capture dynamic feature correlations. Moreover, we introduce latent representation learning to explore the latent supervisory information within these dynamic feature correlations. Combining all these ingredients, we propose a novel semi-supervised MLFS method named PLCFS (Semi-supervised MLFS via partial label correlation and feature self-representation). Moreover, we theoretically demonstrate the convergence of PLCFS. Finally, extensive experimental results on multiple datasets show that, when 20% of the training samples are labeled, compared with existing advanced methods, PLCFS has achieved an overall performance improvement of 1.06%–4.15% in terms of the average precision metric.