Fig. 4.

Gene subset selection for cancer classification. Subsets of gene expression levels were chosen incrementally to maximize the information with the cancer class according to MIST₂, direct estimation of MI or mRMR and scored by the cross-validation error of an SVM classifier. For all data sets, 75% of the data were separated and used to select features and train the model; the classifier was then used to classify the remaining 25%. The mean classification error and standard error of the mean for 200 training/testing partitionings are reported. Genes were selected for data sets relating to (A) breast, (B) leukemia, (C) colon and (D) prostate cancer.