Zone Oriented Binary Multi-Objective Charged System Search Based Feature Selection Approach for Multi-Label Classification

Abstract

Multi-label learning is used in situations when each instance has many labels. Due to the high-dimensional feature space and noise in multi-label datasets, multi-label learning algorithms face substantial problems. Researchers have researched multi-label FS techniques to minimise data dimensionality in multi-label classification (MLC) problems. Global optimization approaches, such as evolutionary algorithm (EA) optimizers, scale well to high-dimensional problems. This paper proposes a hybrid multi-objective FS approach based on the charged system search (CSS) and grey wolf optimization (GWO) methods for the MLC problem. The first objective is to minimise the hamming loss (HLoss) value, and the second objective is to minimise the features from the feature set. A novel concept feature zone based on informative and non-informative features has been added here. Here, we have added the Preference Ranking Organisation METHod for Enrichment of Evaluations (PROMETHEE) approach to the objective function in the FS approach. Here, we have added the new velocity equation for the updated charge particles in the CSS algorithm. The GWO property has been added to the new velocity equation to improve the exploration and exploiting property in the CSS algorithm. For experimental verification, we have utilised six publically accessible multi-label datasets: CAL500, Emotions, Medical, Enron, Scene, and the Yeast. The findings show that the proposed approach gets the best value regarding various performance metrics. The proposed method achieves optimal Jaccard Score (JC) and HLoss values of 0.4408 and 0.0645 for CAL500, 0.8169 and 0.0719 for Emotions, 0.9486 and 0.0019 for Medical, 0.5950 and 0.0205 for Enron, 0.7391 and 0.0495 for Scene, and 0.6452 and 0.0766 for Yeast datasets. In particular, according to empirical data on a popular six-label benchmark multi-label datasets, the proposed method obtains competitive performance when labels are constrained.