Decision Theoretic Clustering

Using Bayesian mixture models for cluster analysis is highly sensitive to model misspecification, typically resulting in more clusters than are actually present in the data. In “Bayesian Clustering via Fusing of Localized Densities”, we propose to extricate components from clusters, instead characterizing clusters with multiple components via Fusing of Localized Densities (FOLD). Our method has has a fully decision theoretic justification, can be easily implemented as an add-on to a Bayesian procedure, and leads to appealing properties in the large sample limit. FOLD can be implemented in practice using the foldcluster package, which is available on Github.