基于集成分类的高维数据实体分辨  

High-dimensional data entity resolution based on ensemble classifying

作　　者：刘艺[1] 刁兴春[1] 曹建军 尚玉玲 Liu Yi;Diao Xingchun;Cao Jianjun;Shang Yuling(College of Command Information Systems,PLA University of Science&Technology,Nanjing 210007,China;Nanjing Telecommunication Technology Institute,Nanjing 210007,China)

机构地区：[1]解放军理工大学指挥信息系统学院,南京210007 [2]南京电讯技术研究所,南京210007

出　　处：《计算机应用研究》2018年第3期689-693,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61371196);中国博士后科学基金特别资助项目(201003797);解放军理工大学预研基金资助项目(20110604;41150301)

摘　　要：针对高维数据实体识别问题,为了有效利用高维特征的富信息,提高分辨性能,提出一种随机组合集成分类器。定义基分类器的分类性能指标,将分类正确性和特征子集的个数作为设计基分类器两个目标,使用聚合函数将其转换为单目标优化问题。采用蚁群优化求解基分类器模型,提出利用最大信息系数度量特征的相关性作为蚁群优化启发式信息,使用谷元距离度量选择特征多样性差异最大的基分类器组合集成分类器,集成分类器的决策函数采用投票表决输出。在标准数据集上进行验证与对比,结果表明了该方法的有效性。In order to effectively use rich information to improve performance of entity resolution in high-dimensional data,this paper proposed a random combinational ensemble classifiers’model.It defined the base classifier’s classification performance’s indicators,used the classification success rate and feature’s number as two objects for optimizing base classifier,and adopted an aggregation function to transform them into a single objective optimization problem.It applied ant colony optimization to design base classifier,and adopted maximal information coefficient to measure correlation between features as heuristic information.The ensemble classifiers were composed of base classifiers which had the best diversity evaluated by Tanimoto distance,and used voting way to decide the output of ensemble classifiers.This paper adopts some benchmark datasets to evaluate the method,and the results show the effectiveness of the method.

关 键 词：实体分辨 高维数据 集成分类器 蚁群优化 最大信息系数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]