A Bayesian approach for simultaneous segmentation and classification of count data

A Bayesian approach is proposed that provides a concise description of a series of counts under the form of homogeneous consecutive data segments that are classified based on their marginal distribution. Due to the flexibility of the corresponding model, carrying out the actual inference turns out to be a complex task for which an original combination of several Markov chain Monte Carlo (MCMC) simulation tools is developed. The proposed MCMC sampler makes use of reversible jump moves to achieve communication between models with different numbers of both segments and classes. A large section of the paper is devoted to the discussion of the results obtained on a medium-duration section (a few minutes) of a publicly available teletraffic trace taken from the Internet traffic archive