Selecting the best splits for classification trees with categorical variables