New Features Selection Method for Multi-label Classification based on the Positive Dependencies among Labels

Authors

  • Mo'ath Al-luwaici, Ahmad Kadri Junoh, Wael Ahmad AlZoubi, Raed Alazaidah, Wais Al-luwaici

Abstract

The problem of high dimensionality badly affects the accuracy of the classification task in both Single Label Classification (SLC) and Multi-label Classification (MLC). In fact, the problem is more worst in MLC; since the features in any instance are associated with different labels at the same time. Also, most datasets in MLC usually suffer from the high dimensionality problem, especially in the large-sized datasets where a dataset could have hundreds if not thousands of features. Therefore, this research is more interested in proposing a novel features selection method that is designed specifically to handle the problem of MLC. The proposed features selection method highly considers the existing dependencies among the class labels. Also, the proposed features selection method depends on the most frequent and strong labels combinations; since most instances in the datasets are linked to combinations of labels, and not to one class labels as have been assumed by the conventional features selection methods. An extensive evaluations of the proposed features selection method considering four different multi-label datasets and five popular Fuzzy-based classifiers revealed the superiority of the proposed features selection method and the significance of considering the frequent label combinations instead of the original class labels when selecting the features to be used in the classification step.

Published

2020-12-01

Issue

Section

Articles