Quality of feature selection based on microarray gene expression data

Abstract

This paper is devoted to the problem of feature selection for class prediction based on results of DNA microarray experiments. A method is presented to objectively compare quality of feature sets obtained using different gene-ranking methods. The quality of feature sets is expressed in terms of predictive performance of classification models built using these features. A comparative study is performed involving means comparison, fold difference and rank-test (Wilcoxon statistic) methods. The study shows that best performance can be obtained using the rank-test approach. It is also shown that the means comparison method can be significantly improved by also taking into account fold-change information. Performance of such mixed methods of feature selection can surpass performance of rank-test methods. © 2008 Springer-Verlag Berlin Heidelberg.