利用一维投影分析的无参数多密度聚类算法  

Parameter Free Multi-density Clustering Using One-dimensional Projection Analysis

作　　者：钱美旋[1] 叶东毅[1] 

机构地区：[1]福州大学数学与计算机学院,福州350108

出　　处：《小型微型计算机系统》2013年第8期1866-1871,共6页Journal of Chinese Computer Systems

基　　金：福建省自然科学基金项目(2010J01329)资助;福建省高校产学研重大专项(2010H6012)资助

摘　　要：针对DBSCAN(Density Based Spatial Clustering of Applications with Noise)算法对参数敏感且无法适用于多密度数据集聚类的缺点,提出一种改进的基于一维投影分析的无参数多密度聚类算法PFMDBSCAN(Parameter Free Multi-Density Clus-tering Using One-dimensional Projection Analysis).算法首先对数据集进行一维投影,并对投影后的数据进行高斯核密度估计,据此采用极值策略得到多个局部密度估计值,将每个局部密度估计值转换为参数后依次调用DBSCAN进行聚类,最终得到完整的聚类结果.该算法达到了聚类无参数化且能适用于多密度的目标.实验表明,本文提出的无参数算法对单密度和多密度数据集都有较好的聚类效果,能适用于任意形状、任意密度的数据集,且具有较强的抗噪性.与近期文献中提出的无参数多密度聚类算法APSCAN相比,不仅聚类效果更好,且计算复杂性更低.In view of the two shortcomings of DBSCAN, sensitive to the parameter and not applicable to cluster multi-density dataset, a parameter free multi-density clustering algorithm based on one-dimensional projection analysis （ PP＇MDBSCAN } is proposed. Our algorithm first makes one-dimensional projection, and then calculates the kernel density estimation of the projective data. After that we apply the extremum strategy to get multiple local densities. By transforming these densities into parameters, DBSCAN is executed in turn and a complete clustering result of dataset is obtained. As a result, the goal of parameter free and being applicable to multiple densities clustering is met. The experiments indicate that the proposed parameter free algorithm performs well on single and multiple densities datasets. It can be applied to discover clusters of arbitrary shapes and densities and be robust to noise data. Compare to the parameter free Multi-Density Clustering algorithm APSCAN proposed in recent paper, PFMDBSCAN has higher clustering accuracy and lower computational complexity.

关 键 词：投影分析 高斯核密度估计 无参数多密度聚类 DBSCAN 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]