ROC-AUC is high but PR-AUC value is very low

21

That’s great news! :slight_smile:
How large is your batch size currently and how does the current confusion matrix look?
Could you try to apply a more aggressive data augmentation for the positive samples (or both classes) and check how the PR AUC changes?

22

:slight_smile: Thank you!!
My batch size is 1024. Currently the confusion matrix looks like this,
[[1000 0]
[ 10 14]]

[[998 2]
[ 17 7]]

[[1018 0]
[ 6 0]]

[[1024]]

Since I do 10-fold cross validation, there were a lot but just to give you a glimpse, I have shown a few here. I do not know about data augmentation. I will read about it and try and let you know.

Thanks a lot for your suggestions :slight_smile: