变精度邻域等价粒的邻域决策树构造算法  

Neighborhood decision tree construction algorithm based on variable-precision neighborhood equivalent granules

作　　者：谢鑫 张贤勇 王旋晔[1,2] 唐鹏飞 XIE Xin;ZHANG Xianyong;WANG Xuanye;TANG Pengfei(School of Mathematical Sciences,Sichuan Normal University,Chengdu Sichuan 610068,China;Institute of Intelligent Information and Quantum Information,Sichuan Normal University,Chengdu Sichuan 610068,China)

机构地区：[1]四川师范大学数学科学学院,成都610068 [2]四川师范大学智能信息与量子信息研究所,成都610068

出　　处：《计算机应用》2022年第2期382-388,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61673258);四川省科技计划项目(2021YJ0085)。

摘　　要：针对现有决策树算法对连续性数据分类的信息丢失、效果不佳等缺点,提出一种邻域决策树(NDT)构造算法。首先,挖掘了邻域决策信息系统上的变精度邻域等价粒,并探讨了相关性质;然后基于变精度邻域等价粒构建邻域基尼指数度量,以度量邻域决策信息系统的不确定性;最后,用邻域基尼指数度量诱导出树节点的选取条件,并以变精度邻域等价粒为树分裂规则,从而构建NDT。在UCI数据集进行实验的结果表明,NDT算法的准确度比基于信息熵的决策树算法ID3、基于基尼指数的决策树算法CART、基于信息增益率的决策树(C4.5)算法和融合信息增益和基尼指数(IGGI)算法平均提高了20个百分点左右,验证了NDT算法的有效性。Aiming at the shortcomings such as information loss and poor effect of the existing decision tree algorithms for continuous data classification,a Neighborhood Decision Tree(NDT)construction algorithm was proposed.Firstly,the variable-precision neighborhood equivalent granules on the neighborhood decision information system were mined,and the related properties were discussed.Secondly,the neighborhood Gini index measure was constructed based on the variable-precision neighborhood equivalent granules to measure the uncertainty of the neighborhood decision information system.Finally,the neighborhood Gini index measure was used to induce the tree node selection conditions,and the variable-precision neighborhood equivalent granules were used as the tree splitting rules to construct NDT.Experimental results on UCI datasets show that the accuracy of NDT algorithm is generally improved by about 20 percentage points compared with those of Iterative Dichotomiser 3(ID3)algorithm,Classification And Regression Tree(CART)algorithm,C4.5 algorithm and combining Information Gain and Gini Index(IGGI)algorithm,indicating that the proposed NDT algorithm is effective.

关 键 词：不确定性度量 基尼指数 邻域决策信息系统 决策树 机器学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]