1
$\begingroup$

I have read many articles about ROC curve. Some specified a method to calculate the accuracy of a classifier using iso accuracy lines in a convex hull. Just like in this articles : http://mlwiki.org/index.php/ROC_Analysis

http://www0.cs.ucl.ac.uk/staff/ucacbbl/roc/

enter image description here

The question is : Most articles pointed out that, the intersection point between the accuracy line of a classifier (i.e C3) and the diagonal line can be used to determine the accuracy of that classifier, just like in the figure. Why? anyone can prove that for me :(

In addition, given the ratio neg/pos i understood that this ratio is the slope of the accuracy line , but how about the intercept, it is not clear why they can construct a line just with a slope

$\endgroup$
1
$\begingroup$

As I understand it, the tangent of a 45 degree line is the ‘optimal’ threshold when sensitivity and specificity are equally weighted (often the default). If you change the weighting of sens/spec, you change the slope of the tangent line.

This is not the ‘accuracy’ of the classifier - that’s usually tied to the AUC (accuracy of an ideal observer).

I don’t think the ‘intercept’ of this tangent line has any real interpretation.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.