关键词: K-means算法, 初始中心, 聚类数, 基于密度

Abstract: The traditional K-means algorithm is sensitive to the selection of the initial clustering center, moreover the clustering number-k can not be confirmed beforehand. Then it is not conductive to the stability of the clustering. Considering this defection, a new improved algorithm—NDK-means（Nonuniqueness high-Density K-means） is proposed which is based on the global clustering center. In this algorithm, the highest-density points are not unique. To determine the radius, the standard deviation is applied to the new algorithm, then the initial clustering centers can be selected from a set of the high-density areas of points. When the highest-density points are not unique, the proposed new algorithm selects a set of initial clustering centers which are the furthest distance from the global clustering center. Besides, the new algorithm can determine the optimal number of cluster by selecting the BWP validity index. Finally, the experimental and analysis results show the new proposed method outperforms than the traditional K-means algorithm in terms of the validity and stability.

Key words: K-means algorithm, initial clustering center, number of clusters, density-based