I find myself in a position of calculating numerous PR / ROC curves and their associated area under the PR curves (AUPR) / area under the ROC curve (AUROC).

Its is quite easy to perform those calculations with standards implementations (I am using sklearn metrics) or even by 'hand'. However, with millions of instances, exact calculations are very time consuming for precision that do not seem necessary for exploratory analyses.

I have been looking for fast approximations for quite a while now. There are some people that propose to optimise the exact calculations, there are some people that propose some approximations for AUROC.

But do we have good approximations for fast PR / ROC curves calculations ?

  • $\begingroup$ What are the factors that impact fast PR / ROC calculations? Could there be floating error or too simplified algorithms that could lead to errors? It could be interesting to calculate the approximation accuracy in a particularly difficult data set (i.e. selecting lot of false positives or false negatives) in order to see the PR / ROC curves at the boundaries and highlight easily the approximation quality. $\endgroup$ Nov 8, 2021 at 15:59


Your Answer

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

Browse other questions tagged or ask your own question.