Data Set Homeomorphism Transformation Based Meta-clustering

Clustering analysis is an important data mining technique with a variety of applications. In this paper, the data set is treated in a dynamic way and a Data Set Homeomorphism Transformation Based Meta-Clustering algorithm (DSHTBMC) is proposed. DSHTBMC decomposes the task of clustering into multiple stages. It firstly constructs a series of homeomorphous data sets ranging from high regularity to low, and then iteratively clusters each homeomorphism data set based on the clustering result of the preceding homeomorphism data set. Since data sets of high regularities are easier to be clustered, and the clustering result of each homeomorphism data set can be used to induce high quality clusters in the following-up homeomorphism data set, in this way, the hardness of the problem is decreased. Two strategies (i.e., Displacement and Noising) for data set homeomorphism transformation are proposed, with classical hierarchical divisive method---Bisecting k-means as DSHTBMC's subordinate clustering algorithm, two new clustering algorithms---HD-DSHTBMC-D and HD-DSHTBMC-N are obtained. Experimental results indicate that the new clustering algorithms are remarkably better than Bisecting k-means algorithm in terms of clustering quality.