进化多目标距离矩阵聚类研究  

Research on Multi-objective Evolutionary Matrix Distance Clustering

作　　者：刘丛[1] 邬春学[1] 

机构地区：[1]上海市现代光学系统重点实验室,教育部光学仪器与系统工程研究中心,上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2016年第6期1298-1302,共5页Journal of Chinese Computer Systems

基　　金：上海重点科技攻关项目(14511107902)资助;上海智能家居大规模物联共性技术工程中心项目(GCZX14014)资助;上海市一流学科建设项目(XTKX2012)资助;沪江基金研究基地专项项目(C14001)资助

摘　　要：聚类分析在科学研究和现实生活中都有广泛的应用.然而,当前的聚类算法仍然面临一些挑战,自动确定最佳聚类数目和复杂分布数据聚类是最主要的两种,自动确定复杂分布数据的聚类数目并对其正确聚类是两者的结合.提出一种基于进化多目标的距离矩阵聚类算法(Multi-objective Distance Matrix Evolutionary Clustering,MODMEC).首先使用一种实数-标签的编码方式表示染色体,该染色体可两次解码成聚类候选解.其次使用聚类代表点代替聚类中心点设计聚类算法,通过类内紧凑度和类间离散性自动确定最佳聚类数目.最后使用进化多目标框架并行优化.将MODMEC在不同分布的五种人工测试集和两种UCI测试集上与四种常用的聚类算法做了比较.实验结果表明,M ODM EC在检测最佳聚类数目和聚类精度上都获得了良好的效果.Data clustering is widely used in both science and real-world applications. However, current clustering methods also f^tce some challenges, such as how to determine the fittest clustering numbers automatically and how to split the data set with complex disWibufion. How to determine the fittest clustering number of data set with complex distribution automatically is a direction for future research. In this paper, a novel method, named multi-objective distance matrix evolutionary clustering, was proposed to solve this problem. First, a real-label coding scheme is used to establish chromosome, which can be changed into candidate clustering result by two decoding steps. Then,cluster representative point is applied to design clustering algorithm instead of cluster center, and it was deter- mine the fittest cluster number automatically by two functions, which are compactness and separation. Finally, the two functions are simultaneous optimization by a multi-objective evolutionary algorithm. The approach is compared with four widely used clustering methods on five artificial and two UCI data sets. Experimental results show that MODMEC not only successfully detects the correct cluster numbers but also achieves satisfactory results for most of test data sets.

关 键 词：聚类 最佳聚类数目 进化多目标算法 进化多目标距离矩阵聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]