A Comparison of Constructing Different Types of New Feature For Decision Tree Learning

Abstract

This chapter studies the effects on decision tree learning of constructing four types of new feature (conjunctive, disjunctive, M-of-N, and X-of-N representations). To reduce effects of other factors such as tree learning methods, new feature search strategies, search starting points, evaluation functions, and stopping criteria, a single tree learning algorithm is developed. With different option settings, it can construct four different types of new feature, but all other factors are fixed. The study reveals that conjunctive and disjunctive representations have very similar performance in terms of prediction accuracy and theory complexity on a variety of concepts, even on DNF and CNF concepts that are usually thought to be suited only to one of the two kinds of representation. In addition, the study demonstrates that the stronger representation power of M-of-N than conjunction and disjunction and the stronger representation power of X-of-N than these three types of new feature can be reflected in the performance of decision tree learning in terms of higher prediction accuracy and lower theory complexity.