密度敏感的谱聚类

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.

Density-Sensitive Spectral Clustering

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 clusters,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 spectral 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.