Since considering features separately struggles so far, we need to discard the original data structure. We don't have time fall into useless results which have been confirmed before.