This involves creating a new training dataset with an approximately equal number of occurrences for each class.