使用证据累积的文本聚类谱算法  

A document clustering spectral algorithm that uses evidence accumulation

作　　者：徐森[1,2] 卢志茂[1] 张春祥[3] 顾国昌 张琦[1] 

机构地区：[1]盐城工学院信息工程学院,江苏盐城224000 [2]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001 [3]哈尔滨理工大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2010年第8期1043-1047,共5页Journal of Harbin Engineering University

基　　金：国家自然科学基金资助项目(60603092;60903082;60975042);高等学校博士学科点专项科研基金资助项目(20070217043)

摘　　要：针对谱聚类算法相似度函数设置困难问题,提出了一种使用证据累积的文本聚类谱算法.该算法使用超球K均值算法对文本集进行多次聚类,并将每次得到的划分结果作为判断2个文本是否应该放在一个簇中的证据,由此构建文本的相似度矩阵和正则化拉普拉斯矩阵.在TREC和Reuters文本集上进行了实验,验证了本文算法的有效性,它比层次聚类算法和CLUTO提供的K均值算法更加优越.Spectral clustering＇s weakness is an inability to choose a similarity measure.To resolve this,a document clustering spectral algorithm using evidence accumulation was proposed.In this algorithm,spherical K-means was first performed over document sets multiple times.Each time the partitioning results were regarded as evidence when judging whether two documents should be put in the same cluster or not.On this basis,the similarity matrix and normalized Laplacian matrix of the documents were constructed.Experiments on the Text REtrieval Conference（TREC） and Reuters document sets demonstrated the effectiveness of the proposed algorithm.It outperformed hierarchical clustering algorithms as well as the K-means algorithm provided in the CLUTO general purpose clustering toolkit.

关 键 词：聚类分析 文本聚类 聚类谱 证据累积 超球K均值 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]