Abstract:

The clusters of a high dimensional dataset are often hidden in the subspaces of the corresponding low dimensional datasets. In order to successfully find the subspaces, Agrawal proposes the conception of projective clustering, converting the data into subspaces with mapping and using another method to find clusters. EPCH is the latest projective clustering algorithm. This paper incorporates Mean-Shift into EPCH to divide a high dimensional dataset into the corresponding subspaces. Experiments demonstrate that the approach is comparable to EPCH in the sense of obtaining the reasonable clusters, however, it doesn’t require any parameter and can reduce the number of subspaces.

Key words: subspace partition, histograms, Mean-Shift, projective clustering

摘要： 高维数据的聚类都隐含在低维的子空间内。为找出有效的子空间，Agrawal等人提出了投影聚类概念，通过映射变换转换到子空间里，然后借助其他方法找到聚类。该文基于目前最新的投影聚类算法EPCH，提出了PCMF算法，借助Mean-Shift划分子空间聚类。与EPCH算法相比，PCMF在划分子空间中数据时，无须输入参数（EPCH中是最大聚类个数），能够有效降低划分出的子空间数量，获得与EPCH相媲美的实验结果。

关键词: 子空间划分, 直方图, Mean-Shift, 投影聚类

CLC Number: 

• TP301.6