Merging Partitions Using Similarities of Anchor Subsets

Abstract

This paper addresses the problem of merging pairs of partition matrices. Such partition matrices may be produced by collaborative clustering. We assume that each subset in one partition matrix matches one of the subsets in the other partition matrix. To align the arbitrarily ordered rows in the partition matrices we use the memberships of a set of anchor points and maximize their pairwise similarities. Here, we consider various set-theoretic similarity measures. Experiments with a simplified version of the well-known BIRCH benchmark data set illustrate the effectivity of the approach and show that all considered similarity measures are well suited for partition merging.