一种可重叠子空间K-Means聚类算法

An Overlapping Subspace K-Means Clustering Algorithm

现有聚类算法面向高维稀疏数据时多数未考虑类簇可重叠和离群点的存在,导致聚类效果不理想。为此,提出一种可重叠子空间K-Means聚类算法。设计类簇子空间计算策略,在聚类过程中动态更新每个类簇的属性子空间,并定义合理的约束函数指导聚类过程,从而实现类簇的可重叠性与离群点的控制。在此基础上定义合理的目标函数对传统K-Means算法进行修正,利用熵权约束分别计算每个类簇中各维度的权重,使用权重值标识不同类簇中维度的相对重要性,并加入控制重叠程度和离群值数量的参数。在人工数据集和真实数据集上的实验结果表明,该算法在NMI、F1指标上均优于EWKM、NEO-K-Means、OKM等子空间聚类算法,具有更好的聚类结果。

Most of existing clustering algorithms for high-dimensional sparse data do not consider overlapping class clusters and outliers,resulting in unsatisfactory clustering results.Therefore,this paper proposes an overlapping subspace K-Means clustering algorithm.The computing strategy for class cluster subspace is given.The attribute subspace of each class cluster is dynamically updated in the clustering process,and a reasonable constraint function is defined to guide the clustering process,so as to realize the overlap of clusters and the control of outliers.On this basis,a reasonable objective function is defined to modify the traditional K-Means algorithm,and the weight of each dimension in each class cluster is calculated by using the entropy weight constraint.The value of weight is used to identify the relative importance of the dimensions in different class clusters.And some parameters are added to control the degree of overlap and the number of outliers.Experimental results on artificial data set and real data set show that the proposed algorithm outperforms EWKM,NEO-K-Means,OKM and other subspace clustering algorithms in terms of NMI and F1 indicators with better clustering results.