I have a classification problem with highly unbalanced classes. In 4000 samples, 1% is 1s and 99% is 0s. Normally, I would use the balance technique only in the training set. However, I expect that the minority group could increase to 10% in the future. Can I undersample the majority group before the cross-validation to achieve 10% in 1s and 90% in 0s? If yes, is there any to paper use this approach?