Corrigendum to "Likelihood ratio equivalence and imbalanced binary classification"

The authors regret a minor inconsistency detected after further experimental work. This note is to point out that the 2-step rebalancing method we presented in Section 8 does not use a neutral emphasis. This is because the emphasized samples in this method are those whose outputs are in some interval of values, and this does not guarantee invariance of the likelihood ratio. The normalization parameters estimates –formula (23a)– will be different, in general. Consequently, when working with classification costs, formula (24) of Section 6 must be applied. Nevertheless, if the selection of the threshold for the output of the final machine design is done to obtain some false alarm/false positive rate, as we do in Section 9, the results will still be valid.

Corrigendum to "Likelihood ratio equivalence and imbalanced binary classification" Articles