Maximum likelihood combination of multiple clusterings

A promising direction for more robust clustering is to derive multiple candidate clusterings over a common set of objects and then combine them into a consolidated one, which is expected to be better than any candidate. Given a candidate clustering set, we show that with a particular pairwise potential used in Markov random fields, the maximum likelihood estimation is the one closest to the set in terms of a metric distance between clusterings. To minimize such a distance, we present two combining methods based on the new similarity determined by the whole candidate set. We evaluate them on both artificial and real datasets, with candidate clusterings either from full space or subspace. Experiments show that they not only lead to a closer distance to the candidate set, but also achieve a smaller or comparable distance to the true clustering.