计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (5): 71-78.DOI: 10.3778/j.issn.1002-8331.1910-0262

• 理论与研发 • 上一篇    下一篇

分布的自动阈值密度峰值聚类算法

彭启慧,宣士斌,高卿   

  1. 广西民族大学 信息科学与工程学院,南宁 530006
  • 出版日期:2021-03-01 发布日期:2021-03-02

Distribution Automatic Threshold Density Peak Clustering Algorithm

PENG Qihui, XUAN Shibin, GAO Qing   

  1. College of Information Science and Engineering, Guangxi University for Nationalities, Nanning 530006, China
  • Online:2021-03-01 Published:2021-03-02

摘要:

密度峰值聚类(DPC)是一种基于局部密度的聚类方法,在DPC中影响算法的效果的两个基本因素是局部密度定义和类中心选择。针对经典DPC在定义局部密度时没有考虑到邻域内样本点的分布情况,以及无法自动选择类中心等问题,提出一种基于分布的局部密度定义和基于最大类间差法的自动类中心选择策略。计算每个样本点截断距离圆圈内的数据点个数,同时考虑数据点的分布情况。当圈内具有相同的点个数时,如果圆圈内的数据点分布越均匀,该点的局部密度就越大,密度峰值的可能性越高。通过最大类间差法(Otsu)自动选择阈值找出类中心。实验结果表明,新算法不仅能够自动选择聚类中心,而且相比已有原算法能获得更高分类准确度。

关键词: 聚类, 密度峰值, 自动选择, 类中心点

Abstract:

Density Peak Clustering(DPC) is a clustering method based on local density. There are two basic factors in the DPC that affect the effect of the algorithm:local density definition and class center selection. For the classical DPC, the specific distribution of sample points in the neighborhood is not taken into consideration when defining the local density, and the cluster center cannot be automatically selected in the cluster. A local density definition based on distribution and an automatic class center selection strategy based on the maximum classes’ square error method are proposed. Firstly, the number of data points within the circle of each sample point is calculated, and the distribution of data points is considered. When there are the same number of points in the circle, if the distribution of data points in the circle is more uniform, the local density of the point is larger, and the probability of peak density is higher. The class center is then found by automatically selecting the threshold by the maximum classes’ square error method(Otsu). Experimental results show that the new algorithm can not only automatically select the clustering center, but also obtain higher classification accuracy than the existing original algorithm.

Key words: clustering, density peak, automatic selection, class midpoint