摘要
针对密度峰值聚类算法(Density Peak Clustering,DPC)在密度分布不均匀及同一个簇有多个高密度点的数据集中难以准确选取聚类中心的情况,提出一种改进节点凝聚度的密度峰值聚类算法.先将数据转化为一个加权的完全图.其次,引入改进后的节点凝聚度的思想构建节点重要度的评价函数,并计算网络中每个节点的局部重要度,聚类中心为局部重要度最高的节点并且与重要度大于该聚类中心重要度的点具有较大距离.然后,对节点重要度进行排序,比较选取节点重要度与距离乘积值异常大的点作为类簇中心.最后,利用所提出的算法和其他密度峰值聚类算法比较,在人工数据集和真实数据集上的实验仿真表明,该算法能够找到具有更高精度的聚类中心,从而可以实现更高的性能.
Aiming at density peak clustering algorithm(DPC),it is difficult to accurately select cluster centres in uneven density distribution datasets or multiple high-density points in the same cluster.It proposed a density peak clustering algorithm on improved node aggregation.Firstly,a dataset to be classified was converted into a weighted complete graph.Secondly,the enhanced aggregation method was constructed as an evaluation function of node importance and calculated the local significance of each node in the network.The cluster centre had a higher value of local importance than surrounding neighbour nodes.And the node,which is compared with other height-importance nodes,had more considerable distance.Then,we could sort node importance and select an extremely value that node importance producting node distance as a cluster centre in the same cluster.Finally,experimental simulations with the proposed algorithm and a few existing DPC based algorithms on both artificial and real datasets show ed that the proposed algorithm can find cluster centres with higher accuracy and thus can achieve improved performance.
作者
吴辰文
魏立鑫
刘晓光
WU Chen-wen;WEI Li-xin;LIU Xiao-guang(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Department of Computer Application,School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Department of Software Engineering,School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第7期1427-1432,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61762057,61662043)资助。
关键词
加权完全图
关键词
凝聚度
节点收缩
weighted complete graph
aggregation
node contraction
density peak clustering
