When is it important to perform a non-random split between the training and testing datasets?