Implement in custom_output the list of high-corr features to not loose information during training for the user interpretation