It has been stated in many places that AUROC is an improper scoring rule.But I haven't seen anyone proving it. Does someone have a working example that shows that maximizing AUROC actually moves away probability estimates from actual probabilities.
Asked
Active
Viewed 32 times
2
-
1Gneiting & Vogel (2018, arXiv:1809.04808), in their footnote 4 note that AUC ignores calibration, referring to p. 346 in Wilks (2011), Statistical Methods in the Atmospheric Sciences. – Stephan Kolassa Sep 20 '19 at 20:27
-
1Take a look at Byrne (2016), as cited in this answer to the proposed duplicate. – Stephan Kolassa Sep 20 '19 at 20:32