Gene Selection for Cancer Classification from Microarray Data Using Data Overlap Measure

Abstract

Cancer detection is one of the major applications of clinical microarray data. High dimensionality is one of the important challenges in microarrays. Most of genes in microarrays have no importance or contribution on the class prediction and on the other side a lot of resources and memory are needed to processing this amount of genes. Thus the reduction in number of dimensions seems to be staple to predict cancer. In this paper a gene selection method using data complexity measures on microarray gene expression cancer data is presented. Two overlap measures as data complexity measure namely fisher discriminant ratio and attribute efficiency are applied to ranking the genes and afterward the high rank genes are considered as important ones to contribute in cancer diagnosis. Five well-known binary microarray cancer data are considered for evaluation and also the applied classifiers are Decision Tree (DT), naïve bayes (NB) and K-Nearest Neighbor (KNN). Two approaches that were considered are fisher-based and (attribute +fisher)-based gene selection. The results indicate that the model created by genes selected by fisher-based method can detect the cancerous samples with high accuracy.