Table 2.

Stability of decision tree models. Up: the models built for up-regulated genes. Down: the models built for down-regulated genes. Noise: the number of noisy instances added into the training set. TP: the number of true positive genes predicted by models built on the original data. TP': the number of true positive genes predicted by models built on the noisy data. Loss: the number of positive instances correctly classified in the original data but mis-classified in the noisy data. Rescue: the number of positive instances correctly classified in the noisy data but mis-classified originally. FP: the number of newly added noise genes classified as positive. Each value is an average across 223 up-regulated or 223 down-regulated gene sets. The standard errors for loss, rescue and FP are all less than 0.2.

		Predefined motifs					Auto motifs
	Noise	TP	TP'	Loss	Rescue	FP	TP	TP'	Loss	Rescue	FP
Up	0	16.6	16.6	0.0	0.0	0.0	26.4	26.4	0.0	0.0	0.0
	5	16.6	17.1	2.5	2.4	0.6	26.4	27.7	7.8	7.6	1.5
	10	16.6	17.4	3.6	3.2	1.1	26.4	28.9	8.1	7.7	2.9
	15	16.6	17.7	4.0	3.5	1.6	26.4	30.3	8.2	7.7	4.4
	20	16.6	18.8	4.2	4.1	2.4	26.4	31.7	8.3	7.8	5.8
	25	16.6	19.3	4.5	4.2	3.0	26.4	32.6	8.5	7.8	6.8
	50	16.6	21.2	5.8	4.6	5.7	26.4	37.8	9.1	7.2	13.3
	0	19.1	19.1	0.0	0.0	0.0	27.9	27.9	0.0	0.0	0.0
Down	5	19.1	19.6	2.0	2.0	0.5	27.9	29.1	7.3	7.0	1.5
	10	19.1	20.5	2.7	2.9	1.1	27.9	29.6	8.0	6.8	2.9
	15	19.1	20.7	3.2	3.2	1.6	27.9	30.6	8.1	6.6	4.1
	20	19.1	20.9	3.8	3.5	2.1	27.9	30.9	8.8	6.5	5.3
	25	19.1	21.2	4.1	3.6	2.6	27.9	32.8	8.5	6.6	6.9
	50	19.1	22.6	5.3	3.6	5.2	27.9	37.8	9.7	6.2	13.4