Machine Learning to Assess Relatedness: The Advantage of Using Firm-Level Data

Table 2

Prediction performance of firm-based models. Our problem is characterized by a huge number of true negatives, and this precludes the use of metrics such as accuracy and ROC-AUC. In any case, the random forest outperforms all other models.

	RCA	Random forest	Product space	Taxonomy network
Best F1	0.050	0.166	0.110	0.126
AUC-PR	0.006	0.083	0.046	0.056
ROC-AUC	0.519	0.936	0.919	0.928
Precision@1000	0.156	0.399	0.207	0.265
mP@5	0.047	0.264	0.237	0.235
MCC	0.052	0.169	0.113	0.130
TP	2703	15584	11482	12706
FP	33387	100292	126310	116773
FN	69037	56156	60258	59034
TN	38952640	38885735	38859717	38869254