Discriminative Gene Selection Employing Linear Regression Model
Microarray datasets enables the analysis of expression of thousands of genes across hundreds of samples. Usually classifiers do not perform well for large number of features (genes) as is the case of microarray datasets. That is why a small number of informative and discriminative features are always desirable for efficient classification. Many existing feature selection approaches have been proposed which attempts sample classification based on the analysis of gene expression values. In this paper a linear regression based feature selection algorithm for two class microarray datasets has been developed which divides the training dataset into two subtypes based on the class information. Using one of the classes as the base condition, a linear regression based model is developed. Using this regression model the divergence of each gene across the two classes are calculated and thus genes with higher divergence values are selected as important features from the second subtype of the training data. The classification performance of the proposed approach is evaluated with SVM, Random Forest and AdaBoost classifiers. Results show that the proposed approach provides better accuracy values compared to other existing approaches i.e. ReliefF, CFS, decision tree based attribute selector and attribute selection using correlation analysis.
