Finding minimum gene subsets with heuristic breadth-first search algorithm for robust tumor classification

BMC Bioinformatics

Table 3 Prediction accuracies of the ensemble SVM(Biased) and SVM(Unbiased) classifiers

Dataset	Ensemble modes	#Individual	Acc.%	Acc.%
		classifiers	(Biased)	(Unbiased)
Leukemia	Top 300 gene subsets	147	92.31	84.62
	10-Fold >98*	47	96.15	88.46
	10-Fold = 100 and Full-fold > =99	5	88.46	86.54
DLBCL	Top 300 gene subsets	61	95.24	85.71
	10-Fold = 100	143**	95.24	85.71
	10-Fold = 100 and Full-fold = 100	29**	95.24	85.71
Prostate	Top 300 gene subsets	300	97.06	88.24
	Full-fold > 98	290	97.06	88.24
	Full-fold > 99	139	97.06	88.24
SRBCT	Top 300 gene subsets	300	90	80
	Full-fold > 98	114	95	85
	Full-fold > 98 and 10-Fold = 100	8	100	90
ALL	Top 300 gene subsets	300	96	96
ALL	10-Fold = 100	59	97	96
	10-Fold = 100 and Full-fold > =99	42	95	95
Colon	Top 300 gene subsets	300	90	70
	10-Fold = 100	62	85	65
	10-Fold = 100 and Full-fold > =98	59	85	65

* The corresponding prediction accuracies (Biased and Unbiased) are obtained on the Leukemia52 test set, respectively. The item 10-Fold > 98 means that the gene subsets with 10-fold CV accuracy greater than 98% are selected from the 300 top-ranked gene subsets in which only 47 gene subsets are shared between the Leukemia72 training set and Leukemia52 test set. Thus the final ensemble classifier consists of the 47 individual classifiers respectively constructed from these 47 gene subsets; the corresponding prediction accuracies (Biased and Unbiased) are obtained by the ensemble classifiers constructed by SVM(Biased) and SVM(Unbiased) on the Leukemia52 test set, respectively.
** The individual classifiers are constructed from the gene subsets that are selected from all nodes in last layer, not limited to the 300 top-ranked nodes in last layer because more than 300 gene subsets can obtain 100% 10-fold CV accuracy on DLBCL.

ISSN: 1471-2105