Intermediate Decision Trees

Abstract

Intermediate decision trees are the subtrees of the full (unpruned) decision tree generated in a breadth-first order. An extensive empirical investigation evaluates the classification error of intermediate decision trees and compares their performance to full and pruned trees. Empirical results were generated using C4.5 with 66 databases from the UCI machine learning database repository. Results show that when attempting to minimize the error of the pruned tree produced by C4.5, the best intermediate tree performs significantly better in 46 of the 66 databases. These and other results question the effectiveness of decision tree pruning strategies and suggest further consideration of the full tree and its intermediates. Also, the results reveal specific properties satisfied by databases in which the intermediate full tree performs best. Such relationships improve guidelines for selecting appropriate inductive strategies based on domain properties. 1 Introduction Numerous decision tree ...