Table 4.

We Examine How Perturbing Edge Weights Affect the Performance of Our LocalCut Algorithm

	Modules		BPMs
Dataset	Accepted	Enriched	Accepted	Enriched for same function	Enriched for same or related function	Enriched for different functions	One mod enriched	No mods enriched
LocalCut	112	103 (92%)	58	39 (67%)	43 (74%)	6 (10%)	9 (16%)	0 (0%)
LocalCut – Variant 1	50	46 (92%)	26	17 (65%)	19 (73%)	2 (8%)	5 (19%)	0 (0%)
LocalCut – Variant 2	133	61 (46%)	68	4 (6%)	6 (9%)	9 (13%)	33 (49%)	20 (29%)
LocalCut – Variant 3	54	37 (69%)	30	3 (10%)	7 (23%)	6 (20%)	17 (57%)	0 (0%)
LocalCut – Variant 4	21	14 (67%)	12	1 (8%)	2 (17%)	3 (25%)	7 (58%)	0 (0%)
LocalCut – Variant 5	98	82 (84%)	49	21 (43%)	30 (61%)	5 (10%)	12 (24%)	2 (4%)
LocalCut – Control	0	0 (0%)	0	0 (0%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)

The results validate our supposition that the nuances of scalar data are more informative than binary weights, and that positive-weight interaction edges matter as well as negative-weight edges.