Recombining partitions from multivariate data: a clustering method on Bayes factors

Abstract

We introduce SAGRA (Split And Group Recombining Algorithm), a cluster analysis methodology which split the data set into small homogeneous groups and later recombine those groups using Bayes factors. We compare the performance of SAGRA with other three cluster analysis algorithms: SAR, M-clust and K-means, using five quality measures: Purity, number of groups, Rand index, adjusted Rand index, and F1, over four different data configurations. Results indicate that the SAGRA algorithm obtain consistently similar or better indexes than the other algorithms over all measures and data configurations