ABSTRACT

A natural goodness of split criterion is to take that split at any node which most reduces the resubstitution estimate of tree misclassification cost. In the two-class problem, a class of splitting criteria is introduced which remedies the deficiencies, and the simplest of these is adopted for use. This chapter deals with the introduction of variable misclassification costs into the splitting criterion in two ways: through a generalization of the Gini criterion and by alteration of the priors. The overall misclassification rate of the tree constructed is not sensitive to the choice of a splitting rule, as long as it is within a reasonable class of rules. The impurity is never increased by splitting. It stays the same only under rare circumstances. In parentheses in the brackets are the classes whose populations are already so small in the parent node that their effect in the split is negligible.