"[SOLVED] ROC: which is the discriminating threshold for classifiers?"

calang

The documentation of the Compare ROC operator states that the ROC plots the classifier curve "as its discrimination threshold is varied".

I think I should then be able to choose a setting for the classifier that is closest to the upper-left corner, by selecting the "discrimination threshold" for that classifier corresponding to a point that made its curve come closest to that corner.

However, I'm not clear about which among the many input parameters is the discrimination threshold for each of the classifier operators, like tree induction and logistic regression.  Naive Bayes is most puzzling as it only has a single binary input parameter.

Can someone tell me which the discrimination threshold parameter is/are for each of those operators (tree, logistic, bayes) - and the other classifiers, if possible? And am I thinking correctly about this?

Thanks in advance.
cl

MariusHelf

Hi,

wessel has explained how ROC plots are created: http://rapid-i.com/rapidforum/index.php/topic,5698.msg20162.html#msg20162

Does that help?

Best regards,
Marius