Table 4.

Ranks of methods based on individual and combined criteria. Smaller ranks correspond to better methods according to each criterion. As described in text, ranks were obtained using formal statistical comparison of the observed differences between methods; that is why they do not necessarily range between 1 and 27 (total number of tested methods).

Method		Rank
		N	PV	AUC	(PV, AUC)
TIE*	max-k = 3, α = 0.05	4	5	2	1
KIAMB	Number ofruns = 5000, α = 0.05, K = 0.7	4	2	4	5
	Number of runs = 5000, α = 0.05, K = 0.8	4	2	4	5
	Number of runs = 5000, α = 0.05, K = 0.9	4	2	4	5
EGS-NCMIGS	l = 7, δ = 0.015	6	1	4	3
	l = 7, K = 10	6	5	3	6
	l = 7, K = 50	6	9	3	11
	l = 5000, δ = 0.015	3	2	4	5
	l = 5000, K = 10	3	4	3	4
	l = 5000, K = 50	3	9	3	11
EGS-CMIM	l = 1, K = 10	6	5	3	6
	l = 7, K = 50	6	9	2	8
	l = 5000, K = 10	3	3	3	2
	l = 5000, K = 50	3	9	2	8
EGSG	Number of Markov boundaries = 30, t = 5	5	6	4	10
	Number of Markov boundaries = 30, t = 10	5	6	4	10
	Number of Markov boundaries = 30, t = 15	5	6	5	13
	Number of Markov boundaries = 5,000, t = 5	2	9	4	14
	Number of Markov boundaries = 5,000, t = 10	2	8	4	12
	Number of Markov boundaries = 5,000, t = 15	1	7	5	15
Resampling+RFE	without statistical comparison	2	10	2	9
	with statistical comparison (α = 0.05)	2	9	3	11
Resampling+UAF	without statistical comparison	3	11	1	7
	with statistical comparison (α = 0.05)	3	10	2	9
IR-HITON-PC	max-k = 3, α = 0.05	6	5	3	6
IR-SPLR	without statistical comparison	6	10	2	9
	with statistical comparison (α = 0.05)	5	9	3	11