密度敏感的谱聚类  

Density-Sensitive Spectral Clustering

作　　者：王玲[1] 薄列峰[1] 焦李成[1] 

机构地区：[1]西安电子科技大学智能信息处理研究所,陕西西安710071

出　　处：《电子学报》2007年第8期1577-1581,共5页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60372050;60133010);国家863高技术研究发展计划(No.2002AA135080);国防预研项目(No.A1420060172)

摘　　要：谱聚类是近来出现的一种性能极具竞争力的聚类方法,它的成功很大程度依赖于相似性度量的选择.本文通过分析这一性质并结合数据聚类特性,提出一种数据依赖的相似性度量——密度敏感的相似性度量.该相似性度量可以有效描述数据的实际聚类分布.将其引入谱聚类得到密度敏感的谱聚类算法.与原有的谱聚类算法相比,新算法不仅能够处理多尺度聚类问题,而且对参数选择相对不敏感.算法有效性分析以及实验验证了所提算法的有效性和可行性.Spectral clustering has become increasingly popular in recent years. Being a pairwise method, the success of spectral clustering depends heavily on the choice of similarity measure. Through analyzing the property of data dusters, a novel data-dependent similarity measure is proposed, namely density-sensitive similarity measure, which has the ability of describing the characters of data clustering compared with the traditional Euclidian metric based similarity measure. Based on the novel similarity measure, we have a density-sensitive spectra/clustering algorithm. Compared with the original spectral clustering,it has the advantages of effectively dealing with the multi-scale problems and relatively not sensitive to parameter. It obtains promising results not only on artificial datasets but also on USPS handwritten digit dataset.

关 键 词：聚类 谱聚类 距离测度 相似性度量 相似性矩阵 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]