Non-convex regularized self-representation for unsupervised feature selection

Feature selection aims to select a subset of features to decrease time complexity, reduce storage burden and improve the generalization ability of classification or clustering. For the countless unlabeled high dimensional data, unsupervised feature selection is effective in alleviating the curse of...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Image and vision computing Jg. 60; S. 22 - 29
Hauptverfasser: Zhu, Pengfei, Zhu, Wencheng, Wang, Weizhi, Zuo, Wangmeng, Hu, Qinghua
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier B.V 01.04.2017
Schlagworte:
Group sparsity
Self-representation
Sparse representation
Unsupervised feature selection
Group sparsity
Sparse representation
Unsupervised feature selection
Self-representation
ISSN:0262-8856, 1872-8138
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Feature selection aims to select a subset of features to decrease time complexity, reduce storage burden and improve the generalization ability of classification or clustering. For the countless unlabeled high dimensional data, unsupervised feature selection is effective in alleviating the curse of dimensionality and can find applications in various fields. In this paper, we propose a non-convex regularized self-representation (RSR) model where features can be represented by a linear combination of other features, and propose to impose L2,p-norm (0 ≤ p < 1) regularization on self-representation coefficients for unsupervised feature selection. Compared with the conventional L2,1-norm regularization, when p < 1, much sparser solution is obtained on the self-representation coefficients, and it is also more effective in selecting salient features. To solve the non-convex (0 <p < 1) RSR model, we further propose an efficient iterative reweighted least square (IRLS) algorithm with guaranteed convergence to a stationary point. When p=0, we exploit the augmented Lagrangian method (ALM) to solve the RSR model. Extensive experimental results on nine datasets show that our feature selection method with small p is more effective. It mostly outperforms RSR with p=1 and other state-of-the-art unsupervised feature selection methods in terms of classification accuracy and clustering performance. •A non-convex regularized self-representation model is proposed.•An iterative reweighted least square algorithm is developed to solve the non-convex (0 <p < 1) case.•An augmented Lagrange method is introduced to solve the non-convex and non-differentiable L20-norm regularized RSR model.
ISSN:0262-8856
1872-8138
DOI:10.1016/j.imavis.2016.11.014