I have a dataset with a binary target variable. This variable is highly imbalanced i.e. the # of True case is ~1% and # of False cases is ~99% The other limitation I have is that I can only use Decision Tree because the target system where I will use this model can only take rules (IF-THEN-ELSE conditions) so I can't use more complex models like Random Forest etc. and the only system I am allowed to use to build a model is SAS Enterprise Miner.
- Oversample my rare class
- Adjust decision weights using "inverse prior weights" (SAS has this feature which was suggested here)
The post goes on to suggest
If you have an event which happens far less than 1% of the time, you may get better results by oversampling and then adjusting the probabilities later.
Is it really necessary to "undo the over sampling" when all I am interested in is to build a decision tree and then convert it into rules which are simple statements like if-then-else, that my target system can use? I am not going to use the actual model for scoring.
Any advise will be very appreciated.