It would be nice if the current operator was modified so that the proportion of classes can be specified. I have worked with several datasets where one class is underrepresented. As a result the model classifies everything as the class that is overrepresented. The modified operator I am proposing would be able to capture a number of cases by class. This way you could control how many cases were sampled from each class. This type of preprocesing in my experience leads to meaningful models.