Efficient Feature Selection Algorithm for Gene Classification

Abstract

Microarray technology was evolved as one of the authoritative mechanisms for an organism to analysis of gene expression level. The microarray gene expression datasets contain a considerably large number (in terms of thousands) of features (genes) and a comparatively small number (in terms of hundreds) of samples. Because of these characteristics, microarray gene expression data analysis is complex. Therefore, efficient feature selection is the immediate requirement. The essential aspects of microarray gene expression data analysis are feature selection and classification. Although many feature selection methods were developed, the SVM, along with recursive component reduced termed as SVM-RFE, was tested to be a promising method. The genes are ranked during SVM classification model training, and critical features are selected with a combination of recursive feature elimination (RFE). The SVM-RFE main drawback was a significant amount of time consumption in the process. Therefore, efficient deployment of linear Support Vector Machine was introduced to overcome this issue. At the same time, Recursive Feature Elimination (RFE) was improvised with the technique known as the variable step size. Along with this, an effective resampling technique was proposed to preprocess the datasets in order to overcome the class imbalance problem. By using this method, the sample became balance from the same distribution that provides better classification result. The recursive feature elimination with variable step size (RFEVSS) with an effective resampling method was used in order to achieve better performance of the classifier that has been presented in this work. The class imbalance problem was addressed by implementation the effective resampling method described in this work. The large-scale linear support vector machine (LLSVM) has also been implemented effectively in order to increase efficiency. The detailed experiments were conducted to test the result with three classifiers on four benchmark microarray gene expression datasets. The results were presented in graphical form for better understanding.