Predicting Hidden Links in Supply Networks

Table 2

Comparison of predictive accuracy.

(a) Confusion Matrices for each case study, AUC and Accuracy

Algorithm	Logistic Regression		Naïve Bayes
	Link	No-link	Link	No link
JLR
Predicted true	501	45878	573	117164
Predicted false	248	273582	176	202296
Class recall	66.89%	85.64%	76.50%	63.32%
AUC	0.81		0.80
Accuracy	0.76		0.70
Saab
Predicted true	590	114585	405	18877
Predicted false	159	197070	344	292778
Class recall	78.77%	63.23%	54.07%	93.94%
AUC	0.81		0.80
Accuracy	0.71		0.74
Volvo
Predicted true	1040	134402	834	42997
Predicted false	547	865865	753	957270
Class recall	65.53%	86.56%	52.55%	95.70%
AUC	0.81		0.79
Accuracy	0.76		0.74

(b) PR-AUC and PR-CAUC

	Volvo	JLR	Saab
Logistic Regression
PR-AUC	0.140	0.085	0.073
PR-CAUC	0.113	0.024	0.012
Naïve Bayes
PR-AUC	0.143	0.033	0.044
PR-CAUC	0.063	0.002	0.0004

(c) PR-AUC and PR-CAUC with increased training sizes

	Volvo	JLR	Saab
PR-AUC	0.1	0.073	0.073
PR-CAUC	0.9	0.019	0.009
PR-AUC	0.14	0.085	0.073
PR-CAUC	0.113	0.024	0.012
PR-AUC	0.16	0.11	0.11
PR-CAUC	0.12	0.04	0.03