面向密度分布不均数据的混合近邻密度峰值聚类算法  

Multiplex neighbor density peaks clustering for uneven density data sets

作　　者：吕莉[1] 朱梅子 康平 韩龙哲 LYU Liy;ZHU Mei-zi;KANG Ping;HAN Long-zhe(School of Information Engineering,Nanchang Institute of Technology,Nanchang Jiangxi 330099,China)

机构地区：[1]南昌工程学院信息工程学院,江西南昌330099

出　　处：《控制理论与应用》2024年第10期1821-1830,共10页Control Theory & Applications

基　　金：国家自然科学基金项目(62066030,61962036)资助.

摘　　要：密度峰值聚类算法(DPC)的局部密度忽略了密度分布不均数据的疏密差异,易导致类簇中心聚集在密集区域;其分配策略在分配剩余样本时,易将稀疏区域样本错误分配到密集区域,致使聚类效果不佳.为克服上述缺陷,本文提出了面向密度分布不均数据的混合近邻密度峰值聚类(MN-DPC)算法.首先,利用自然近邻信息定义样本的局部密度,平衡稀疏区域与密集区域样本之间的密度差异,从而正确找到稀疏区域的类簇中心;其次,利用样本之间的共享及自然近邻信息对样本相似度进行加权处理,加强了同一类簇样本间的相似度,有效的避免稀疏区域样本被错误分配.本文将MN-DPC算法与IDPC-FA,DPC-DBFN,DPCSA,FNDPC,FKNN-DPC,DPC算法进行对比.实验结果表明,MN-DPC算法能有效聚类密度分布不均及UCI数据集.The local density of density peaks clustering(DPC)algorithm ignores the density difference of the data with uneven density distribution,which easily leads to the cluster centers found in the dense area resulting in poor clustering effect.In order to overcome the above shortcomings,this paper proposes multiplex neighbor density peaks clustering for uneven density data sets(MN-DPC).Firstly,the natural nearest neighbor information is used to define the local density of samples to balance the density difference between samples in sparse and dense regions,so as to correctlyfind the class cluster centers in sparse regions;Secondly,the sample similarity is weighted by using the shared nearest neighbor and natural nearest neighbor information,which strengthens the similarity between samples of the same type of cluster and effectively avoids the misallocation of samples in sparse regions.This paper compares the MN-DPC algorithm with the IDPC-FA,DPC-DBFN,DPCSA,FNDPC,FKNN-DPC and DPC algorithms.The experimental results show that the MN-DPC algorithm can effectively cluster data sets with uneven density distribution and UCI data sets.

关 键 词：密度峰值聚类 局部密度 自然近邻 共享近邻 样本相似性 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]