Telescoping mixtures - Learning the number of components and data clusters in Bayesian mixture analysis

Description

Telescoping mixtures are an extension of sparse finite mixtures by assuming that additional to the unknown number of data clusters also the number of mixture components is unknown and has to be estimated. Telescoping mixtures explicitly distinguish between the number of data clusters K+ and components K in the mixture distribution, and purposely allow for more components than data clusters. By linking the prior on the number of components to the prior on the mixture weights, it is guaranteed that components remain empty as K increases, making the number of clusters in the data, defined through the partition implied by the allocation variables, random a priori. Telescoping mixtures can be seen as an alternative to infinite mixtures models. We present a simple algorithm for posterior MCMC sampling to jointly sample K, the number of components, and K+, the number of data clusters. The algorithm is compared to standard transdimensional algorithm such as the reversible jump Markov chain Monte Carlo and the Jain-Neal split-merge sampler.

Period	26 Aug 2019 → 29 Aug 2019
Event title	16th Conference of the International Federation of Classification Societies
Event type	Unknown
Degree of Recognition	International