最近最远得分的聚类性能评价指标  

A clustering evaluation index based on the nearest and furthest score

作　　者：冯柳伟 常冬霞[1,2] 邓勇[3] 赵耀[1,2] 

机构地区：[1]北京交通大学信息科学研究所,北京100044 [2]北京交通大学计算机与信息科学学院,北京100044 [3]中国科学院软件研究所,北京100190

出　　处：《智能系统学报》2017年第1期67-74,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金"重点"项目(61532005)

摘　　要：聚类算法是数据分析中广泛使用的方法之一,而类别数往往是决定聚类算法性能的关键。目前,大部分聚类算法需要预先给定类别数,在很多情况下,很难根据数据集的先验知识获得有效的类别数。因此,为了获得数据集的类别数,本文基于最近邻一致性和最远邻相异性的准则,提出了一种最近最远得分评价指标,并在此基础上提出了一种自动确定类别数的聚类算法。实验结果证明了所提评价指标在确定类别数时的有效性和可行性。The clustering algorithm is one of the widely-used methods in data analysis. However ’ the number of clusters is essential to determine the performance of the clustering algorithm. At present ’ the number of clusters usually need to be specified in advance. In most cases ’ it is difficult to obtain the valid cluster number according to a priori knowledge of the dataset. To obtain the number of clusters automatically ’ a Nearest and Furthest Score （NFS） index was proposed based on the principles of the nearest neighbor consistency and the furthest neighbor difference. Moreover,an Automatic Clustering NFS （ACNFS） algorithm was also proposed’ which can determine the number of clusters automatically. The experimental results prove the index is reasonable and practicable to determine the cluster number.

关 键 词：最近邻一致性 最远邻相异性 K-MEANS聚类算法 评分机制 评价指标 层次聚类 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]