Significant Features Determination for ATS Drug Identification

Abstract

Laboratory testing for ATS drug identification is a costly and lengthy process. In this paper, we propose a computational analysis approach as an alternative solution in identifying the ATS drugs. High dimensional dataset is one of the key challenges for computational analysis. This paper will investigate the effectiveness of several feature selection algorithms in identify the significant features and filter out the irrelevant features in the dataset. Specifically, four filters feature selection techniques (Information Gain (IG), Gain Ratio (GR), Symmetrical Uncertainty (SU), and ReliefF) and two embedded feature selection techniques (Support Vector Machine based Recursive Elimination Method (SVM-RFE) and Variable Importance based Random Forest (VIRF)) have been explored. The main fundamental perspective that is taken into consideration in performance analysis is to identify which feature selection technique can return minimal features while achieving a higher identification performance. The experimental evaluation on the ATS drugs 3D molecular structure representation dataset is performed using five classifiers, which are Random Forest (RF), Naïve Bayes (NB), IBK, SMO and J48 decision trees. The findings show that ReliefF and VIRF can select a smaller feature subset with the highest identification accuracy than the other feature selection techniques.