AUC greater than 1

TCB

Hi. I have been tracking performance of EvoSVM by outputting mean values of several measures after 10-fold cross-validations, including AUC. For several different parameter combinations, RM returns values for AUC greater than 1, such as, 1.06. This is especially worrying because I need to present these data in a few days. This only happens with EvoSVM and not with, say, WAODE The other performance measures (recall, precision, F) all seem sensible.

I would most definitely appreciate any timely explanations.

Thanks.

land

Hi,
I guess, the EvoSVM is just that good...:)
Ok, just joking. If the EvoSVM is the only learner causing this problems, the confidence calculation might be wrong. Did you check if the confidences are reasonable (between 0 and 1 and adding to 1)?  Perhabs you could use the Plattscaling Operator after prediction to get the confidences right.

Greetings,
  Sebastian