基于Hubness现象的高维数据混合聚类算法  

A hybrid clustering algorithm for high dimensional data based on Hubness phenomenon

作　　者：王妍 马燕[1] 黄慧 李顺宝 张玉萍[1] WANG Yan;MA Yan;HUANG Hui;LI Shunbao;ZHANG Yuping(College of Information,Mechanical and Electrical Engineering,Shanghai Normal University,Shanghai 200234,China)

机构地区：[1]上海师范大学信息与机电工程学院

出　　处：《电视技术》2019年第6期17-23,共7页Video Engineering

基　　金：国家自然科学基金(61373004,61501297)资助项目

摘　　要：高维数据聚类是聚类分析中的难点。K-hubs聚类算法是在K-means方法基础上,结合高维数据空间的Hubness现象对数据进行聚类。针对K-hubs聚类算法需要随机确定初始聚类中心,不适用于非超球状簇等问题,本文提出了基于多阶段层次聚类和划分聚类的高维数据混合聚类算法。该算法将数据点按其Hub值分为Hub点,Midhub点和Antihub点三类,然后对Hub点和Midhub点分别采用层次聚类,接着进一步采用层次聚类合并簇,最后,对Antihub点利用划分聚类合并到最近的簇。在UCI数据集上的实验结果表明,与其它最新的聚类算法相比,本文提出的算法在高维数据集上得到了较好的聚类结果。High dimensional data clustering is a difficult task in clustering analysis.Based on K-means algorithm,K-hubs algorithm performs the clustering for the data combining the Hubness phenomenon in high dimensional data space.To tackle the issues such as random selection of initial clustering centers and not being adaptive to nonhyperspherical clusters,this paper proposes a hybrid clustering algorithm for high dimensional data based on multi-stage hierarchical clustering and partition clustering.The data points are classified into three categories including Hub points,Midhub points and Antihub points.Then,we conduct hierarchical clustering for Hub points and Midhub points.Next,the clusters are further merged with hierarchical clustering.Finally,Antihub points are merged into the nearest clusters with partition clustering.The experimental results on UCI data sets show that the proposed algorithm can achieve better clustering results on the high dimensional data set compared with the state of the art method.

关 键 词：高维数据 聚类 Hubness现象 层次聚类 K-MEANS算法 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]