Flexible multi-view feature selection with semi-supervised label semantic alignment

Semi-supervised multi-view feature selection is an important pretext task that optimizes features in different views using a few annotations. Most semi-supervised feature selection optimizes features of labeled and unlabeled samples via supervised regression and unsupervised regularization objectives. However, the label semantic gap between supervised and unsupervised modules is easily neglected. Although semi-supervised label propagation obtains consistent labels, it is hard to incorporate it with feature selection. Moreover, the plain manner in feature selection that puts ℓ_2,1-norm of different views together overlooks the diversity of views towards multi-view data. To address them, we investigate a novel label semantic learning framework for semi-supervised multi-view feature selection. Specifically, the semantic labels of the entire data are required by the semi-supervised regression module and the unsupervised clustering module simultaneously, and two kinds of semantic labels are aligned by a coordinate rotation loss, enabling the supervised labels and unsupervised labels to couple with each other. Meanwhile, we tactfully design an elastic ℓ_δ-norm sparse regularization that collaborates ℓ_2,1-norm and ℓ_2,0-norm, and it could distinguish different views in multi-view feature selection via the view-specific parameter δ_v. Extensive results in evaluating the optimized labels and features demonstrate the superiority of the proposed method compared to the state-of-the-art and representative algorithms. The code will be released at https://github.com/zhanghan9937/Multi-view-Feature-Selection-with-Label-Alignment.