二阶K近邻和多簇合并的密度峰值聚类算法
Density peaks clustering with second-order K-nearest neighbors and multi-cluster merging
吕莉 1朱梅子 1康平 1韩龙哲1
作者信息
- 1. 南昌工程学院 信息工程学院,南昌 330099;南昌工程学院 南昌市智慧城市物联感知与协同计算重点实验室,南昌 330099
- 折叠
摘要
针对流形数据中密度峰值聚类(DPC)算法的局部密度易找到错误的类簇中心,且分配策略易导致远离类簇中心的剩余样本被错误分配的问题,本文提出二阶K近邻和多簇合并的密度峰值聚类(DPC-SKMM)算法.首先,利用最小二阶K近邻定义局部密度,凸显类簇中心与非类簇中心间的密度差异,从而找到正确的类簇中心;其次,利用K近邻找出样本局部代表点并依此确定核心点,用核心点指导微簇划分;最后,利用最小二阶K近邻及共享近邻定义的微簇间吸引度合并微簇,避免远离类簇中心的样本被错误分配,且微簇合并过程无须迭代.本文将DPC-SKMM算法与IDPC-FA、DPCSA、FNDPC、FKNN-DPC、DPC算法进行对比,实验结果表明,DPC-SKMM算法能有效聚类流形及UCI数据集.
Abstract
In the face of manifold data,the local density of density peaks clustering(DPC)algorithm is easy to find the wrong cluster center and the allocation strategy is easy to cause the residual samples far from the cluster center to be misallocation.In view of the above problems,this paper proposes density peaks clustering with second-order K-nearest neighbors and multi-cluster merging.Firstly,the minimum second-order K-nearest neighbor is used to define the local density,highlighting the density difference between the cluster center and the non-cluster center,so as to find the correct cluster center;Secondly,the K-nearest neighbor is used to find the local representative points of the sample and determine the core points,and the core points are used to guide the micro-cluster division;Finally,the inter-cluster attraction defined by the minimum second-order K-nearest neighbor and shared nearest neighbor is used to merge the micro-clusters,which avoids the misallocation of samples away from the cluster center,and the micro-cluster merging process does not require iteration.In this paper,DPC-SKMM algorithm is compared with IDPC-FA,DPCSA,FNDPC,FKNN-DPC,DPC algorithm.Experimental results show that DPC-SKMM algorithm can cluster manifolds and UCI data sets effectively.
关键词
密度峰值聚类/流形数据/二阶K近邻/K近邻/吸引度/多簇合并策略Key words
density peaks clustering/manifold data/second-order K-nearest neighbors/K-nearest neighbor/attractiveness/multi-cluster merging strategy引用本文复制引用
基金项目
国家自然科学基金(62066030)
国家自然科学基金(61962036)
出版年
2024
NETL
NSTL
万方数据