The-Minh Luong (MAP5)

Fast estimation of posterior probabilities in change-point models through a constrained hidden Markov model approach

mardi 10 avril 2012, 14h30 - 15h30

Salle de réunion, espace Turing


In bioinformatics, a vast amount of methodology has been developed to identify an ideal set of change-points for detecting Copy Number Variation (CNV). We consider the important problem of assessing the uncertainty of a set of estimated change-point locations, which is typically estimated with quadratic complexity. We describe a constrained hidden Markov model that corresponds to the classical model of heterogeneous segments with contiguous observations. This model permits the use of a forward-backward algorithm of linear complexity for estimating quantities of interest, such as the posterior probability of a change-point at a given location. The methods are implemented in the R package postCP; we obtain confidence intervals around change-point estimates used to identify CNV in cancer. We also introduce an approach for fast model selection in change-point problems using results of postCP to estimate criteria such as the Integrated Completed Likelihood (ICL).

L’exposé sera en français. De plus, attention cette fois-ci, exceptionnellement, la séance commencera à {{14h30}}.