Skip to Main content Skip to Navigation
Conference papers

Model selection with untractable likelihood

Christine Keribin 1, 2 Vincent Brault 1 
2 SELECT - Model selection in statistical learning
Inria Saclay - Ile de France, LMO - Laboratoire de Mathématiques d'Orsay
Abstract : Penalised likelihood criteria such as AIC or BIC are popular methods used to deal with model selection and require to compute the maximised likelihood. Unfortunately, this maximised likelihood can be untractable, as it is the case for the latent block model (LBM). LBM is a mixture model that allows to perform the simultaneous clustering of rows and columns of large data matrices, also known as coclustering. Due to the complex dependency structure of the bservations conditionally to the row and column labels, approximations must be defined to perform the E step of the EM algorithm, leading to a lower bound of the maximised likelihood. For the same reason, the usual asymptotic approximation used to derive BIC is itself questionable. On the other hand, the integrated completed likelihood criterion (ICL) is exactly computed for LBM, but requires to investigate the influence of hyperparameters. This influence is investigated, links between the criteria are discussed and numerical experiments on both simulated and real data sets highlight the proposed model selection procedure.
Complete list of metadata
Contributor : Christine Keribin Connect in order to contact the contributor
Submitted on : Monday, January 6, 2014 - 2:50:14 PM
Last modification on : Sunday, June 26, 2022 - 12:00:24 PM


  • HAL Id : hal-00924197, version 1


Christine Keribin, Vincent Brault. Model selection with untractable likelihood. ERCIM - 6th International Conference of the ERCIM Working Group on Computing and Statistics, 2013, Dec 2013, London, United Kingdom. ⟨hal-00924197⟩



Record views