Abstract
The Bayes coding algorithm for context tree source is a successful example ofBayesian tree estimation in text compression in information theory. Thisalgorithm provides an efficient parametric representation of the posterior treedistribution and exact updating of its parameters. We apply this algorithm to aclustering task in machine learning. More specifically, we apply it to Bayesianestimation of the tree-structured stick-breaking process (TS-SBP) mixturemodels. For TS-SBP mixture models, only Markov chain Monte Carlo methods havebeen proposed so far, but any variational Bayesian methods have not beenproposed yet. In this paper, we propose a variational Bayesian method that hasa subroutine similar to the Bayes coding algorithm for context tree sources. Weconfirm its behavior by a numerical experiment on a toy example.