Classification rules discovery from selected trees for thinning with the C4.5 machine learning system

Abstract

To use heuristic information efficiently, it is necessary to develop a knowledge-based system and to digitize acquired knowledge. The purpose of this article is to discover classification rules from selected trees for thinning with a machine learning system, C4.5, and to compare these to results from a neural network model. An algorithm used for information entropy, the gain ratio criterion, was used in order to induce decision trees and production rules. The number of samples used was 503 and two kinds of thinning types were used: binary case and four cases. The rate of accurate classification when trees were classified into four thinning types was about 63%–78%. In the case of a binary decision, whose output was “thinned” or “unthinned,” about 88%–97% of its answers were correct. The C4.5 machine learning system can be constructed from 10 to 20 rules, in comparison with the large data sample in this study. Where only estimation accuracy for unseen cases was compared, estimation accuracy of the neural network was superior to that of C4.5 for each thinning type. However, C4.5 has an advantage in that the rules are shown as linguistic information such as if-then type rules, which people can easily comprehend and use.

Key words:

• Decision trees
• Gain ratio criterion
• Production rules
• Machine learning
• Neural network