Accurate and Stable Feature Selection Powered by Iterative Backward Selection and Cumulative Ranking Score of Features

This paper focuses on a stable feature selection framework using Cross Validation technique and SVM-RFE. Though SVMRFE has outperformed many of its counterparts in feature subset selection for accurate cancer classification, its greediness in selecting optimal feature subset affect the stability of selection process in successive runs that brings down the confidence on the selected features. In this paper, we propose an iterative backward feature selection method using SVMRFE motivated by cross-validation technique. Cumulative Ranking Score (CRS) is a parameter formulated to determine the class discrimination ability of each feature. The proposed method is applied on the publically available breast cancer dataset and found top 10 highly discriminative genes. Later the SVM classifier is trained using the top 10 genes identified by the proposed method and the original SVM-RFE separately and tested. It is proved that the proposed method has improved the classification accuracy significantly compared to the original SVM-RFE.