Training sample set generation from imbalanced data in view of user goals