Two-stage gene selection for support vector machine classification of microarray data

This paper proposes a new stable gene selection method for support vector machines (SVM) classification of microarray data, aiming to improve the classification accuracy. A two-stage algorithm is used to select genes, leading to the construction of a compact multivariate linear regression model, which contains only genes less than the number of experiments as well as a weight vector for each gene index. An SVM then learns the microarray data based on this linear regression model. The experimental results, from two well-known microarray datasets, show that SVMs with two-stage gene selection maintains a consistently high accuracy with a small number of genes. It is also shown that the proposed method outperforms the two other typical gene selection methods - baseline method and significance analysis of microarrays in terms of accuracy.