基于有效性函数的聚类算法

聚类作为一种无监督的学习方法，通常需要人为地提供聚类的簇数。在先验知识缺乏的情况下，通过人为指定聚类参数是不合实际的。近年来研究的聚类有效性函数（Cluster Validity Index） 用于估计簇的数目及聚类效果的优劣。本文提出了一种新的基于有效性指数的聚类算法，无需提供聚类的参数。算法每步合并两个簇，使有效性指数值增加最大或减小最少。本文运用引力模型度量相似度，对可能出现的异常点情况作均匀化的处理。实验表明，本文的算法能正确发现特定数据的簇个数，和其它聚类方法比较，聚类结果具有较低的错误率，并在效率上优于一般的基于有效性指数的聚类算法。

A Clustering Algorithm Based on Cluster Validity Indices

Being an unsupervised learning,clustering algorithm suffering from the limitation that the number of clusters is specified by a human user.However,it is not suitable for human to specify the cluster parameters while lacking of prior knowledge.The main purpose of cluster validity index studied recently is to estimate the number of clusters and the goodness or validity of the clusters.In this paper,we propose a new clustering algorithm based on cluster validity indices,which obviates the needs for cluster parameters.We choose at each step the merging of clusters that results in the greatest increase(or smallest decrease) in cluster validity index.And we take gravitation model as calculating similarity distance;dispose the abnormal points that could be appeared.Experimental results show that our methods can detect the number of clusters based on given dataset exactly,and the error rates of clustering results are lowerthan the other clustering algorithm.We outperform the other cluster algorithms based on cluster validity indices with higher efficiency.