A comprehensive comparison of random forests and support vector machines for microarray-based cancer classification

Table 1 Comparison of classification performance of SVMs and RFs without gene selection.

*Task & dataset*	*Classification performance metric*	*Classification performance*		*Nominally superior method*	*P-value*
		*SVM*	RF
Dx-Alon	AUC	0.867	0.867	-	1
Dx-Ramaswamy2	AUC	0.821	0.767	SVM	0.409
Dx-Shipp	AUC	0.992	0.973	SVM	0.500
Dx-Singh	AUC	0.964	0.944	SVM	0.377
Px-Beer	AUC	0.798	0.646	SVM	0.032
Px-Bhattacharjee	AUC	0.519	0.561	RF	0.546
Px-Iizuka	AUC	0.663	0.763	RF	0.061
Px-Pomeroy	AUC	0.692	0.600	SVM	0.235
Px-Rosenwald	AUC	0.689	0.629	SVM	0.140
Px-Veer	AUC	0.747	0.754	RF	0.867
Px-Yeoh	AUC	0.777	0.660	SVM	0.006
Dx-Alizadeh	RCI	1.000	1.000	-	1
Dx-Armstrong	RCI	0.944	0.894	SVM	0.658
Dx-Bhattacharjee	RCI	0.895	0.763	SVM	0.015
Dx-Golub	RCI	0.939	0.934	SVM	1
Dx-Khan	RCI	1.000	1.000	-	1
Dx-Nutt	RCI	0.775	0.733	SVM	0.498
Dx-Pomeroy	RCI	0.823	0.611	SVM	0.031
Dx-Ramaswamy	RCI	0.905	0.861	SVM	0.010
Dx-Staunton	RCI	0.770	0.819	RF	0.249
Dx-Su	RCI	0.958	0.910	SVM	0.004
Px-Veer2	RCI	0.451	0.304	SVM	0.004

The performance is estimated using area under ROC curve (AUC) for binary classification tasks and relative classifier information (RCI) for multicategory tasks. See subsection "Statistical comparison among classifiers" for the description of statistical test employed to compute reported p-values. P-values shown with boldface denote statistically significant differences between classification methods at the 0.05 α level.

ISSN: 1471-2105