决策粒K均值聚类算法

An algorithm of decision granular K-means clustering

针对传统K均值聚类效果不稳定,陷入局部最优解和簇边缘样本聚类不准确的问题,提出一种基于决策粒的K均值聚类算法.该算法利用卡方检验、卡方权值参数进行特征选择,再与样本密度和间隔差异结合,选取出优秀初始质心;然后,根据粒向量距离度量方法得到基础粒K均值聚类结果,构造决策粒模型,对决策粒向量进行聚类决策划分.最后,结合多个UCI数据集实验,将该算法与多种经典聚类算法在聚类结果、性能评估、迭代次数等方面进行比较.结果表明,提出的决策粒K均值聚类算法具有选择优秀初始质心和提高聚类性能的优点.

A K-means clustering algorithm based on decision granules is proposed,to solve the problem of unstable effect,falling into local optimal solutions and inaccurate clustering of edge samples between clusters of traditional K-means clustering.The algorithm uses chi-square test and chi-square weight parameters for feature selection,then combines with the sample density and interval difference to select the excellent initial centroids.Then,based on the granular vector distance measure,the basic granular K-means clustering results are obtained,and a decision granular model is constructed to cluster and partition the decision granular vector.Finally,combining experiments with multiple UCI datasets,this algorithm is compared with various classical clustering algorithms in terms of clustering results,performance evaluation,iteration times,and other aspects.The results show that the proposed decision granular Kmeans clustering algorithm has the advantages of selecting excellent initial centroids and improving clustering performance.