Evidence accumulation clustering based on the K-means algorithm

Abstract

The idea of evidence accumulation for the combination of multiple clusterings was recently proposed [7]. Taking the K-means as the basic algorithm for the decomposition of data into a large number, k, of compact clusters, evidence on pattern association is accumulated, by a voting mechanism, over multiple clusterings obtained by random initializations of the K-means algorithm. This produces a mapping of the clusterings into a new similarity measure between patterns. The final data partition is obtained by applying the single-link method over this similarity matrix. In this paper we further explore and extend this idea, by proposing: (a) the combination of multiple K-means clusterings using variable k; (b) using cluster lifetime as the criterion for extracting the final clusters; and (c) the adaptation of this approach to string patterns. This leads to a more robust clustering technique, with fewer design parameters than the previous approach and potential applications in a wider range of problems.