Reconstruction of a decision tree with learning examples generated from an original tree and its characteristics
Tohru Asami, Hachisu Unoki, Kazuo Hashimoto, Seiichi Yamamoto · Systems and Computers in Japan · 1994
Abstract This paper describes the a posteriori method of decision tree learning after the tree is applied to a real domain, such as medical diagnoses. Without collecting a new set of diagnosis examples, the presented algorithm reconstructs a decision tree preserving the error rate of diagnosis from an original tree and a frequency of diagnoses, which is counted at reaching the corresponding terminal node of that tree when applied to a real domain. The new tree has a shorter path length to diagnose and a logically same meaning with the original tree because of generating a set of pseudoexamples whose unobserved attribute values uniformly distribute in the value range. To reduce the computational cost, a method to avoid a generation of a pseudoexample set also is presented. The context dependencies between attributes are considered by introducing an attribute concatenation. The experiments show that an average path length will be reduced by 6 to 10 percent after reconstruction of a randomly generated decision tree with nonoptimized diagnosis frequencies.