A Hybrid Feature Selection Based on Fisher score and SVM-RFE for Microarray Data
Hind Hamla, Khadoudja Ghanem
Abstract
In the last two decades, analyzing microarray data plays a critical role in disease diagnosis and identification of different tumors. However, it is difficult to classify microarray data because of the curse of the dimensionality problem, in which the number of features is huge while the number of samples is small. Thus, dimension reduction techniques, such as feature selection methods, play a vital role in eliminating non-informative features and enhancing cancer classification. In this paper, we propose a Filter-embedded hybrid feature selection method for the gene selection problem. First, the proposed method selects the top-ranked features obtained from the Fisher score to provide a candidate subset for the embedded stage. Second, Support Vector Machine-Recursive Feature Elimination (SVM-RFE) applies to the candidate subset to find the optimal subset. We assess the performance of our proposed method over ten high-dimensional microarray datasets. The results reveal that the proposed method enhances the classification accuracy, reduces the number of selected features, and decreases computational time.