Gene Selection for Cancer Classification Using DCA

Gene selection is a very important problem in microarray data analysis and has critical implications for the discovery of genes related to serious diseases. In this paper the problem of gene selection for cancer classification is considered. We develop a combined SVMs-feature selection approach based on the Smoothly Clipped Absolute Deviation penalty, minimizing directly the classifier performance. To solve our optimization problems, we apply the DCA (Difference of Convex functions Algorithms) which is a general framework for non-convex continuous optimization. This leads to a successive linear programming algorithm with finite convergence. Preliminary computational experiments on different real data demonstrate that our methods accomplish the desired goal: suppression of a large number of features with a small error of classification.