基于几何的K-means初始聚类中心优化算法研究  

Research on Optimization Algorithm for Initial Clustering Centers of K-means Based on Geometry

作　　者：周晓东 董海清 张昆鹏 侯俊丞 孙树峰 ZHOU Xiaodong;DONG Haiqing;ZHANG Kunpeng;HOU Juncheng;SUN Shufeng(Shanghai Police College,Shanghai 200137,China;Ctrip Network Technology(Shanghai)Co.,Ltd.,Shanghai 200335,China;Postal Savings Bank of China,Beijing 100808,China;Shanghai Research Institute of Criminal Science and Technology,Shanghai 200083,China)

机构地区：[1]上海公安学院,上海200137 [2]携程旅游网络技术(上海)有限公司,上海200335 [3]中国邮政储蓄银行股份有限公司,北京100808 [4]上海市刑事技术科学研究院,上海200083

出　　处：《仪表技术》2025年第2期66-69,73,共5页Instrumentation Technology

基　　金：上海市科学技术管理委员会专项课题(22DZ1200500);上海市智慧警务协同创新中心专项课题(23xtcx04);上海市智慧警务协同创新中心项目(XC202303006)。

摘　　要：传统的K-means算法对初始聚类中心较为敏感,聚类结果随初始输入不同而波动显著,且易陷入局部最优解。为消除该敏感性,提出了一种改进初始聚类中心选取的新方法。运用主成分分析将高维数据降至平面二维,随后计算每个数据对象的欧氏距离与向量角度参数,建立距离角度混合评价模型,选取k个分散性最高的数据点作为初始聚类中心。实验结果表明,该算法对处理高维数据具有一定的优越性,尤其对非簇状数据集能产生较优的聚类结果,并且消除了初始输入的敏感性。The traditional K-means algorithm is sensitive to the initial clustering centers.The clustering results fluctuate significantly with different initial inputs and are prone to fall into local optimal solutions.To eliminate that sensi-tivity,a new method is proposed to improve the selection of the initial clustering center.The PCA is used to make high-dimensional data down to a two-dimensional in this method.Then Euclidean distance and vector angle are calculated for each data object parameter,and Distance-Angle hybrid evaluation model is set up to select the k points with high disper-sion as the initial clustering centers.Experiments show that the proposed algorithm has certain advantages while dealing with high-dimensional data,especially for the non-cluster data sets,and it eliminates sensitivity of the initial inputs.

关 键 词：聚类中心 K-MEANS算法 欧氏距离 角度参数 

分 类 号：P751.1[交通运输工程—港口、海岸及近海工程]