参数估计决策树算法

DECISION TREE BY PARAMETER ESTIMATION

本文提出了一种新的决策树算法.引入了基于统计估计的方法,并针对学习问题做了修正,同时考虑了特征提取和特征选择的因素.基于UCI数据的实验结果以及实际应用的测试结果都表明,本文方法比C4.5的判决精度更高,计算速度更快.

Feature extraction, which decides the distribution of examples in feature space, and feature selection, which decides the suitable features for shaping best classifier, are both important for classification. But conventional methods for making decision tree such as C4. 5 think little about the former. Different from the entropy based inductive learning methods, a new method that considers both feature extraction and feature selection is proposed. Statistical estimation methods are introduced and modified for learning in the proposed method, selecting and deciding the kind of distribution shadowed on a certain dimension that is best for class discrimination. Evaluation results tested on UCI datasets and real applications show that the proposed method is much faster in making tree and has higher predictive accuracy than C4. 5 algorithm does.