The Studies of Optimal Continuous Attribute Multiple Splitting in Decision Tree

• Main Authors: Ke Chih Ning, 甯格致
• Other Authors: T.Y. Tseng
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/98803818878451365043

碩士 === 華梵大學 === 資訊管理學系碩士班 === 96 === Traditionally, decision tree performs as not well as novel classifiers like BPN or SVM in aspect of accuracy. However, the human readable result generated by decision tree truly makes it capable to fulfill the requirement of some expert domain like medical、business…etc. But when the dataset has severe data confliction in it or there are large amount of information hidden in it, the decision tree will turns out a big size tree with high depth which makes it hard to understand by human. Although decision tree yields out average performance among usual applications, the inefficiency of binary cut which used to split the continuous data is still an academic topic, and there are several methods claims to facilitate the performance of continuous attribute splitting, but eventually the improvement lacks of generalization. In this research, we proposed a meta-heuristics method (e.g. simulated annealing) with a proper designed Objective Function and two pruning method (In-Build Pruning, Post Pruning) in the process the tree generation were joined. We choose 10 UCI datasets as our original data. All of the experiments were follow the procedure of 10-folds cross validation method. Our proposed method finally generated much more simplified decision trees with relatively small tree size & almost half of the tree depth compared with C4.5. In addition, the accuracy has no significant differences.