基于结构系数的K-means初始聚类中心选择算法  

Algorithm for Selecting K-means Initial Clustering Centers Based on the Structure Coefficient

作　　者：李汉波 魏福义[1] 张嘉龙 刘志伟 黄杰[1] 方月宜 LI Hanbo;WEI Fuyi;ZHANG Jialong;LIU Zhiwei;HUANG Jie;FANG Yueyi(College of Mathematics and Information,South China Agricultural University,Guangzhou 510642)

机构地区：[1]华南农业大学数学与信息学院,广州510642

出　　处：《计算机与数字工程》2023年第5期993-996,1107,共5页Computer & Digital Engineering

基　　金：广东省大学生创新创业项目“基于图像变换与压缩算法的荔枝损伤研究”(编号:S202010564034);华南农业大学微达安产业学院项目资助。

摘　　要：传统的K-means算法选取初始聚类中心时的不确定性会导致聚类结果不稳定。论文提出了基于相异度的邻域及其结构系数的概念,从最小的结构系数开始,按照其递增顺序寻找初始聚类中心;随后采用依次缩小邻域的技巧逐步探索,直到找到K个初始聚类中心。该方法同时得到li(i=0,1,2,…,q)个初始聚类中心及其对应的数据分类结果。实验证明,对比于以往的算法,新算法具有更高的分类准确率以及更少的迭代次数。The uncertainty in selecting initial clustering centers usually leads to unstable clustering results of the traditional K-means method.This paper proposes a new concept of neighborhood and structure coefficients based on dissimilarity.This algorithm looks for the initial clustering centers in order of ascending structure coefficients,starting from the neighborhood with the smallest coefficient.The above procedure is repeated with the neighborhood reduced until the K initial clustering centers are found.By the proposed algorithm,it can finally obtain li(i=0,1,2,…,q)initial clustering centers and the corresponding data classification results.Experiments show that this algorithm achieves higher classification accuracy and needs less number of iterations than the existing algorithms.

关 键 词：K-MEANS聚类 相异度 初始聚类中心 结构系数 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]