基于交互信息的连续值属性决策树学习算法  

Using Mutual Information For Selecting Continuous-Valued Attribute in Decision Tree Learning

作　　者：李华[1] 胡晓绢[2] 

机构地区：[1]石家庄铁道学院数理系,河北石家庄050043 [2]石家庄铁道学院电气与电子工程学院,河北石家庄050043

出　　处：《南华大学学报（自然科学版）》2008年第2期84-87,共4页Journal of University of South China：Science and Technology

摘　　要：Fayyad连续值属性决策树学习算法使用信息熵的下降速度作为选取扩展属性标准的启发式,本文针对其易选取重复的条件属性等不足之处,引入属性间的交互信息,提出了一种改进算法——基于交互信息的连续值属性决策树学习算法,它的核心是使用信息熵和交互信息的下降速度作为选取扩展属性标准的启发式.实验结果表明,与Fayyad决策树学习算法相比,该算法降低了决策树中同一扩展属性的重复选取率,实现了信息熵的真正减少,提高了训练精度和测试精度,能构造出更优的决策树.In this paper, we proposed a learning algorithm using the information entropy minimization heuristic and mutual information entropy heuristic to select expanded attributes. For a data set, of which the values of condition attributes are continuous, most of the current decision trees learning algorithms often select the previously selected attributes for branching. The repeated selection limits the accuracy of training and testing and the structure of decision trees may become complex. So in the selection of attributes, the previously selected attributes and the other attributes, which have high correlation to the previously selected attributes, should not be selected again. Here, we use mutual information to a- void selecting the previously selected attributes in the generation of decision trees and our test results show that this method can obtain good performance.

关 键 词：经验概念学习 决策树 最小信息熵 交互信息 离散化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]