«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

27

期数:

2017年09期

页码:

31-34

栏目:

智能、算法、系统工程

出版日期:

2017-09-10

文章信息/Info

Title:

 A Method for Determination of Optimal Value in K-means Clustering with Generalization

文章编号:

1673-629X（2017)09-0031-04

作者:

 张雄; 赵礼峰

 南京邮电大学 理学院

Author(s):

 ZHANG Xiong; ZHAO Li-feng

关键词:

 ; K-均值; 最佳聚类数; 泛化能力; 非监督学习

Keywords:

 ; K-means clustering; optimal number of clusters; generalization; unsupervised learning

分类号:

TP301

文献标志码:

A

摘要:

 
针对K-均值聚类算法需要事先确定聚类数,而人为设定聚类数存在极大主观性的缺点,提出了一种基于泛化能力的最佳聚类数确定方法.该方法认为:一个好的聚类结果,应该对未知的样本有着良好的泛化能力.其通过设计一种泛化能力指标(GA)来评价得到的聚类模型对未知样本的分类能力,泛化能力指标的值越大,则聚类模型的效果越好,以泛化能力最优的聚类模型所对应的K值作为最佳聚类数.为了测试所提出方法的稳定性和有效性,分别基于真实数据集Iris以及人造数据集对基于泛化能力的最佳聚类数确定方法进行了实验验证,均能准确找到数据集最佳聚类数.实验结果表明,该方法能够简单、高效地获得最佳聚类数,且对数据集的聚类效果良好.

Abstract:

 Aimed at the defect of K-means clustering algorithm determining the clustering number in advance which could be defined arti-ficially and is subjective in computations, a method of determining an optimal clustering value with generalization is proposed. It is thought that a good clustering result should have good generalization to the unknown samples. Therefore,a generalization index is de-signed to evaluate the classification of the unknown samples in the clustering model obtained. The more the value of generalization index, the better the effect of clustering model. The K value corresponded by clustering model with optimal generalization is selected as the opti-mal clustering value. In order to verify its stability and effectiveness, the experiments are carried out in optimal clustering determining methods based on generalization based on Iris and artificial data set,which indicate that it is simple and efficient to obtain the optimal clustering number,and has the good clustering effect.

常用功能

更新日期/Last Update: 2017-10-19