一种基于K-Means局部最优性的高效聚类算法

K-Means聚类算法只能保证收敛到局部最优,从而导致聚类结果对初始代表点的选择非常敏感.许多研究工作都着力于降低这种敏感性.然而,K-Means的局部最优和结果敏感性却构成了K-MeanSCAN聚类算法的基础.K-MeanSCAN算法对数据集进行多次采样和K-Means预聚类以产生多组不同的聚类结果,来自不同聚类结果的子簇之间必然会存在交集.算法的核心思想是,利用这些交集构造出关于子簇的加权连通图,并根据连通性合并子簇.理论和实验证明,K-MeanScan算法可以在很大程度上提高聚类结果的质量和算法的效率.

An Efficient Clustering Algorithm Based on Local Optimality of K-Means

K-Means is the most popular clustering algorithm with the convergence to one of numerous local minima,which results in much sensitivity to initial representatives.Many researches are made to overcome the sensitivity of K-Means algorithm.However,this paper proposes a novel clustering algorithm called K-MeanSCAN by means of the local optimality and sensitivity of K-Means.The core idea is to build the connectivity between sub-clusters based on the multiple clustering results of K-Means,where these clustering results are distinct because of local optimality and sensitivity of K-Means.Then a weighted connected graph of the sub-clusters is constructed using the connectivity,and the sub-clusters are merged by the graph search algorithm.Theoretic analysis and experimental demonstrations show that K-MeanSCAN outperforms existing algorithms in clustering quality and efficiency.