I have tried a 2 different versions of a gbm in a multinomial classification problem. The second model results in better confusion matrix but in worse Log Loss value (at the test sample). How is that possible.
Further are the results of the two models.
I thought that it could be because the Class A is much more oversampled and the small decrease of that class could lead to such a deterioration of hte logloss?
Any ideas? Thank you