Table 6.

Average normalized mutual information for block recovery with different methods and the ground truth.

N, n	p, q	MGC	SGC	Our
N = 2, 000, n = 100	0.4, 0.01	0.47 (~7 h)	0.89 (~500 s)	0.99 ( ~ 26 s)
	0.5, 0.1	0.10 (~7 h)	0.39 (~500 s)	0.40 ( ~ 26 s)
	0.8, 0.3	0.09 (~7 h)	0.34 (~500 s)	0.37 ( ~ 26 s)
N = 4, 000, n = 200	0.4, 0.01	0.26 (~70 h)	0.90 (~3,700 s)	0.97 ( ~ 120 s)
	0.5, 0.1	0.09 (~70 h)	0.41 (~3,700 s)	0.44 ( ~ 120 s)
	0.8, 0.3	0.09 (~70 h)	0.40 (~3,700 s)	0.44 ( ~ 120 s)
N = 6, 000, n = 300	0.4, 0.01	N/A	0.87 (~12,000 s)	0.88 ( ~ 300 s)
	0.5, 0.1	N/A	0.43 (~12,000 s)	0.48 ( ~ 300 s)
	0.8, 0.3	N/A	0.40 (~12,000 s)	0.47 ( ~ 300 s)

The number in parenthesis shows the average running time per run. The last case for MGC is omitted (noted as N/A) due to the running time limit (we set it as 100 h). Bold value stands for the best performance for each dataset.