Nature of the ROC-AUC calculation #3
Description
Dear Christopher,
thank you very much for this great package. I have a question about the ROC-AUC calculation. I noticed that the value in your method does not correspond to the mean value of caret models that are based on resampling, for example, as using bootstrapping.
I have looked at your code. If I understood it correctly, you are averaging the probabilities per rowIndex from the Caret object. I had the same implementation in my code.
However, this does not correspond 1 to 1 to the ROC-AUC value that Caret returns to you. The ROC-AUC value is the average of all ROC-AUC values and not the average of all probabilities.
How do you see this? Is there any literature that justifies this kind of approach?
Best regards
BP