Table 1.

The best performing gene selection procedure with the .632+ bootstrap error identified by FiGS for each of the six microarray datasets.

Dataset	Feature selection method	k	Gene expression pattern	Feature discretization	Feature vector addition	Classifier	Error
Leukemia	Wilcoxon rank sum test	10	Down-regulated	Not apply	Not apply	SVM	0.02
Leukemia	Wilcoxon rank sum test	10	Down-regulated	Not apply	Not apply	RF	0.02
Leukemia	Wilcoxon rank sum test	10	Down-regulated	Not apply	Apply	SVM	0.02
Leukemia	Wilcoxon rank sum test	10	Down-regulated	Not apply	Apply	RF	0.02
Leukemia	Wilcoxon rank sum test	10	Down-regulated	Apply	Not apply	SVM	0.02
Leukemia	Information gain method	10	Down-regulated	Not apply	Not apply	SVM	0.02
Leukemia	Information gain method	10	Down-regulated	Not apply	Not apply	RF	0.02
Leukemia	Information gain method	10	Down-regulated	Not apply	Apply	SVM	0.02
Leukemia	Information gain method	10	Down-regulated	Not apply	Apply	RF	0.02
Leukemia	Information gain method	10	Down-regulated	Apply	Not apply	SVM	0.02
Leukemia	Information gain method	10	Down-regulated	Apply	Not apply	RF	0.02
Colon	Information gain method	30	Up-regulated	Not apply	Not apply	RF	0.11
Prostate	Information gain method	25	Total	Not apply	Not apply	RF	0.05
Adenocarcinoma	Wilcoxon rank sum test	10	Up-regulated	Not apply	Not apply	RF	0.10
Breast	Wilcoxon rank sum test	15	Down-regulated	Not apply	Apply	SVM	0.31
Breast	Information gain method	15	Down-regulated	Not apply	Apply	SVM	0.31
DLBCL	Wilcoxon rank sum test	20	Total	Not apply	Not apply	RF	0.08

k is the number of selected genes; and error is the .632+ bootstrap error achieved by the best performing gene selection procedure tested on 100 bootstrap samples. In the case of the leukemia and breast datasets, the multiple gene selection procedures are the best.