I am working on a project with high binary class imbalance.
I've tried other methods such as class weight and using an ensemble, but I want to try out a third: upsampling.
Is it ideal to upsample the test data?
Thanks.
$\begingroup$
$\endgroup$
3
-
$\begingroup$ your test data should be as realistic as possible and there should be no information „leak“ from the train set. I would not upsample the test set. $\endgroup$– PeterAug 6, 2019 at 20:26
-
$\begingroup$ Ok. Thanks Peter. $\endgroup$– GozieAug 6, 2019 at 20:32
-
2$\begingroup$ I think you should specify some more things before the question can be answered. What kind of data do you have? How do you pretend to upsample? The question "Is it ideal to upsample the test data?" can't be abstractly answered. $\endgroup$– 89f3a1cAug 6, 2019 at 22:56
Add a comment
|