Meta Clustering

Abstract

Most clustering methods search for one optimal partitioning of the data. Often it is better to search for many different clusterings of the data and present the user with a means of efficiently navigating between them. We present two algorithms for generating many alternate clusterings: Sample-and-Merge and Component Reweighting. We then use clustering at a meta level to organize these different base-level clusterings. This {\em MetaClustering} partitions the base-level clusterings into groups of similar clusterings. We demonstrate MetaClustering on a synthetic data set, and on a real protein data set. The results show that the algorithms are effective at generating qualitatively different clusterings, and at organizing these clusterings so that similar ones are grouped together.