Recursive gene selection based on maximum margin criterion: a comparison with SVM-RFE

BMC Bioinformatics

Table 4 Performance comparison for multi-class datasets (continued).

Classifier+Selection criterion	Number of genes
	10	20	30	50	100
CNS
NMC+BW	31.1 ± 1.3	23.1 ± 1.2	20.1 ± 1.1	18.3 ± 1.0	15.9 ± 1.0
NMC+MMC-RFE(U)	27.2 ± 1.1	22.8 ± 0.9	21.9 ± 0.9	19.4 ± 0.8	16.8 ± 0.8
NMC+MMC-RFE(O)	24.4 ± 1.0	22.7 ± 0.8	22.1 ± 0.9	20.6 ± 0.9	18.9 ± 0.8
NMC+SVM-RFE(H)	45.6 ± 1.3	35.4 ± 1.0	33.3 ± 1.0	28.8 ± 0.9	24.9 ± 0.8
NMC+SVM-RFE(S)	45.4 ± 1.3	34.9 ± 1.0	32.5 ± 0.9	27.6 ± 0.8	24.6 ± 0.8
MMC+MMC-RFE(U)	27.6 ± 1.1	22.5 ± 0.9	21.3 ± 0.9	19.2 ± 0.8	16.9 ± 0.8
MMC+MMC-RFE(O)	24.4 ± 1.0	22.9 ± 0.8	22.2 ± 0.9	20.2 ± 0.9	19.4 ± 0.8
SVM+SVM-RFE(H)	54.0 ± 1.5	42.6 ± 1.4	36.8 ± 1.3	31.0 ± 0.9	25.2 ± 0.8
SVM+SVM-RFE(S)	47.3 ± 1.2	37.7 ± 1.1	32.6 ± 1.1	28.4 ± 1.0	26.6 ± 0.9
NCI60
NMC+BW	49.8 ± 1.2	44.0 ± 1.0	41.6 ± 1.0	39.1 ± 0.8	37.7 ± 0.7
NMC+MMC-RFE(U)	46.4 ± 0.8	38.9 ± 0.8	34.0 ± 0.9	29.8 ± 0.9	26.8 ± 0.7
NMC+MMC-RFE(O)	48.2 ± 0.9	39.6 ± 0.9	35.0 ± 0.9	31.6 ± 0.8	30.2 ± 0.9
NMC+SVM-RFE(H)	60.6 ± 1.0	51.4 ± 1.0	48.4 ± 1.0	43.4 ± 0.9	38.0 ± 0.8
NMC+SVM-RFE(S)	60.8 ± 1.0	52.2 ± 0.9	47.3 ± 1.0	41.3 ± 0.9	39.0 ± 0.9
MMC+MMC-RFE(U)	46.0 ± 0.9	37.3 ± 0.8	33.7 ± 0.8	29.0 ± 0.9	25.0 ± 0.7
MMC+MMC-RFE(O)	49.0 ± 1.0	38.6 ± 0.9	34.3 ± 0.9	30.4 ± 0.8	28.7 ± 0.9
SVM+SVM-RFE(H)	64.7 ± 1.2	54.3 ± 1.1	47.7 ± 1.0	42.0 ± 0.9	35.9 ± 0.9
SVM+SVM-RFE(S)	59.9 ± 1.1	50.3 ± 1.0	46.2 ± 1.0	42.8 ± 1.1	35.8 ± 0.9

The average error and standard error rates (%) for CNS and NCI60, when the number of genes is {10, 20, 30, 50, 100}. SVM-RFE(S) shows the best result with respect to the C parameter; NMC+SVM-RFE(S): C = 10, SVM+SVM-RFE(S): C = 0.1 for CNS; NMC+SVM-RFE(S): C = 100, SVM+SVM-RFE(S): C = 0.1 for NCI60.

ISSN: 1471-2105