Mark Last

Most of decision tree induction algorithms use a greedy splitting criterion. One of the possible solutions to avoid this greediness is looking ahead to make better splits. Look-Ahead has not been used in most decision tree methods primarily because of its high computational complexity and its questionable contribution to predictive accuracy. In this paper we describe a new Look-Ahead approach to induction of decision tree models. We present a computationally efficient algorithm which evaluates quality of subtrees of variable-depth in order to determine the best split attribute out of a set of candidate attributes with a splitting criterion statistically indifferent from the best one.