基于密度RPCL的K-means算法

目的探索同时确定K-means算法的最佳聚类数K和最佳初始聚类中心的方法,使K-means算法的聚类结果尽可能地收敛于全局最优解或近似全局最优解。方法以次胜者受罚竞争学习(Rival Penalized Competitive Learning,RPCL)作为K-means的预处理步骤,以其学习结果作为K-means的聚类数和初始聚类中心并依据数据集样本自然分布定义样本密度,将此密度引入RPCL的节点权值调整,以此密度RPCL的输出作为K-means的最佳聚类数K和最佳初始聚类中心。采用UCI机器学习数据库数据集以及随机生成的带有噪音点的人工模拟数据集进行实验测试,并用不同的聚类结果评价指标对聚类结果作了分析。结果提出的密度RPCL为K-means提供了最佳的类簇数和最佳的初始聚类中心。结论基于密度RPCL的K-means算法具有很好的聚类效果,对噪音数据有很强的抗干扰性能。

A density RPCL based K-means algorithm

Aim To solve the two open problems of K-means algorithm that needs the cluster number in advance and its clustering results depend on initial seeds.Methods The advantages of RPCL(Rival Penalized Competitive Learning) are introduced to uncover the clusters of a dataset and the best initial seeds for K-means as well.The deficiency of the available RPCL algorithms is analyzed and a density RPCL is proposed here,where the density for a sample is defined according to the pattern distribution of samples in a dataset and the density is introduced to adjust the weights of nodes in RPCL.The density RPCL is used as the preprocessing procedure to determine the proper number of clusters and the best initial seeds for K-means algorithm.The density RPCL based K-means clustering algorithm is tested on some well-known data sets from UCI machine learning repository and on some synthetic data sets with noise samples,and the experimental results are analyzed upon many different criteria.Results The density RPCL algorithm uncovers the proper clusters of a dataset and the best initial seeds for K-means.Conclusion The density RPCL based K-means algorithm achieves better clustering result and is insensitive to noisy data.