Some Asymptotic Properties of Model Selection Criteria in the Matent Block Model

International audienceCo-clustering designs in a same exercise a simultaneous clustering of the rows and the columns of a data array. The Latent Block Model (LBM) is a probabilis-tic model for co-clustering, based on a generalized mixture model. LBM parameter estimation is a difficult problem as the likelihood is numerically untractable. However , deterministic or stochastic strategies have been designed and the consistency and asymptotic normality have been recently solved when the number of blocks is known. We address model selection for LBM and propose here a class of penalized log-likelihood criteria that are consistent to select the true number of blocks for LBM