Unsupervised feature selection with joint self-expression and spectral analysis via adaptive graph constraints.

Autor: You, Mengbo, Ban, Lujie, Wang, Yuhan, Kang, Juan, Wang, Guorui, Yuan, Aihong
Předmět:
Zdroj: Multimedia Tools & Applications; Feb2023, Vol. 82 Issue 4, p5879-5898, 20p
Abstrakt: Unsupervised feature selection (UFS) plays a critical role in the maintenance of representative feature subset from high dimensional data. Both the spectral analysis model and the self-expression model are effective in selecting important features. However, few alternative methods embed these two models into a joint FS framework. To address this problem, we propose a novel UFS method that simultaneously selects the most representative feature subset and makes the selected feature subset discriminative by mapping the original features into the label space. Specifically, both the self-expression and spectral analysis are introduced into our method and the self-expression matrix is used as the FS matrix. The two modules are not simply added together, but interact with each other through two graph constraints which preserve the local structure and the manifold structure of the original data, respectively. Furthermore, this paper proposes an alternative iterative algorithm to solve the four matrices involved in the proposed method. To verify the effectiveness of our method, extensive experiments are implemented, and the experimental results prove that the proposed method achieves the best performance among the current state-of-the-art UFS methods. Moreover, an ablation study is performed to show the effectiveness of each part of the proposed method. [ABSTRACT FROM AUTHOR]
Databáze: Complementary Index
Externí odkaz: