Section: New Software and Platforms

BLOCKCLUSTER software

Participants : Gilles Celeux, Christine Keribin.

Mixture model, Block cluster analysis, Blockcluster is software devoted to model-based block clustering. It is developed in partnership with the MODAL team (Inria Lille). This year, some major bugs have been fixed, and the Bayesian point of view has been reinforced by including Gibbs sampling for binary and categorial data. This Gibbs sampler, coupled with the variational Bayes algorithm, provides solutions which are more stable and less dependent on the initial values of the algorithm. An exact expression of the ICL criterion has been provided. This non-asymptotic criterion appears to be more relevant than the BIC-like approximation of ICL.

Vincent Brault, Christine Keribin and Mahindra Mariadassou have shown the consistency and asymptotic normality of the maximum likelihood and variational estimators in stochastic or latent block models.