The implication of data diversity for a classifier-free ensemble selection in random subspaces