1
$\begingroup$

I've created ROC curves by calculating the TPR and FPR at various thresholds. The FPR range differs between models, so I'm wondering if AUC is still a valid way to compare the curves. A curve will have a low AUC value if it has a small FPR range, but this doesn't make it a worse model, does it?

Sidenote: These aren't machine learning models. I'm considering subsets of biological data as a model, and comparing results from each subset to those of the complete dataset to determine the accuracy of results.

ROC curvve image

$\endgroup$
0
$\begingroup$

You know that from a theoretical point of view the curve will always pass through the points (0,0) and (1,1) so you could extend all your curves to pass through those points and then compute AUC. Having said that the portions of the curves which you have currently look rather strange because the scales on the two axes are very different so if you tried to draw the 45 degree line it would look very flat so if you do not feel comfortable with including the two extreme points then at least plot on the same scales so it is clear how much you current curves deviate from the diagonal line.

$\endgroup$

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.