Agglomerative and divisive hierarchical Bayesian clustering

Abstract

Cluster analysis methods are designed to discover groups of subjects or objects in datasets by uncovering latent patterns in data. Two model-based Bayesian hierarchical clustering algorithms are presented—divisive and agglomerative—that return nested clustering configurations and provide guidance on the plausible number of clusters in a principled way. These algorithms outperform many existing clustering methods on benchmark data. The methods are applied to identify subpopulations among Parkinson's disease subjects using only baseline data, and differing patterns of progression in the few years following diagnosis are demonstrated in the identified clusters.