Abstract: K-means clustering algorithm clusters datasets according to the certain clustering number k. However,k cannot be confirmed beforehand. A new clustering validity index was designed from the standpoint of sample geometry. Based on the index, a new method for determining the optimal clustering number in K-means clustering algorithm was proposed. Theoretical research and experimental results demonstrate the validity and good performance of the above-mentioned algorithm.

Key words: K-means clustering, number of clusters, clustering validity index, cluster analysis

摘要： K-means聚类算法是以确定的类数k为前提对数据集进行聚类的，通常聚类数事先无法确定。从样本几何结构的角度设计了一种新的聚类有效性指标，在此基础上提出了一种新的确定K-means算法最佳聚类数的方法。理论研究和实验结果验证了以上算法方案的有效性和良好性能。

关键词: K-means聚类, 聚类数, 聚类有效性指标, 聚类分析