Article: Fast trace ratio-based feature selection Journal: International Journal of Wireless and Mobile Computing (IJWMC) 2022 Vol.22 No.3/4 pp.265 - 273 Abstract: The quality of features has a great impact on machine learning tasks. Feature selection obtains a high-quality feature subset from data, which has been widely studied because of high interpretability. In this paper, we propose a novel feature selection algorithm called trace Ratio-Based Feature Selection (RBFS), which first defines the distance of different classes and the same classes for a given sample and then projects these distances into the subspace. The margin is defined by the trace ratio of these two distances. The objective function is formulated by maximising the margin. To avoid a trivial solution, the orthogonal subspace and the L2,1 norm are incorporated into the objective function. Then, theoretically, the rewritten objective function can obtain the optimal solution through alternating iterations. In addition, power iteration is introduced to reduce the computational cost. Comprehensive experiments are conducted to compare the performance of the proposed algorithm with six other state-of-the-art ones. Inderscience Publishers - linking academia, business and industry through research

Title: Fast trace ratio-based feature selection

Authors: Bo Liu; Ling Ling Tao; Xi Ping He

Addresses: Chongqing Engineering Laboratory for Detection, Control and Integrated System, Nan'an, Chongqing, China; School of Artificial Intelligence, Chongqing Technology and Business University, Nan'an, Chongqing, China ' Chongqing Engineering Laboratory for Detection, Control and Integrated System, Nan'an, Chongqing, China; School of Artificial Intelligence, Chongqing Technology and Business University, Nan'an, Chongqing, China ' Chongqing Engineering Laboratory for Detection, Control and Integrated System, Nan'an, Chongqing, China; School of Artificial Intelligence, Chongqing Technology and Business University, Nan'an, Chongqing, China

Abstract: The quality of features has a great impact on machine learning tasks. Feature selection obtains a high-quality feature subset from data, which has been widely studied because of high interpretability. In this paper, we propose a novel feature selection algorithm called trace Ratio-Based Feature Selection (RBFS), which first defines the distance of different classes and the same classes for a given sample and then projects these distances into the subspace. The margin is defined by the trace ratio of these two distances. The objective function is formulated by maximising the margin. To avoid a trivial solution, the orthogonal subspace and the L_2,1 norm are incorporated into the objective function. Then, theoretically, the rewritten objective function can obtain the optimal solution through alternating iterations. In addition, power iteration is introduced to reduce the computational cost. Comprehensive experiments are conducted to compare the performance of the proposed algorithm with six other state-of-the-art ones.

Keywords: feature selection; trace ratio criteria; large margin; power iteration.

DOI: 10.1504/IJWMC.2022.124817

International Journal of Wireless and Mobile Computing, 2022 Vol.22 No.3/4, pp.265 - 273

Received: 22 Sep 2021
Accepted: 26 Feb 2022
Published online: 09 Aug 2022 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Fast trace ratio-based feature selection

Keep up-to-date