De l'identification de structure de réseaux bayésiens à la reconnaissance de formes à partir d'informations complètes ou incomplètes.
Olivier François · HAL (Le Centre pour la Communication Scientifique Directe) · 2006
We have performed an empirical study of various deterministic Bayesian networks structure learning algorithms. The first test step has allowed us to emphasise which learning technics need a specific initialisation et we have proposed a way to do this. In the second stage of this doctoral study, we have adapted some learning technics to incomplete datasets. Then, we have proposed an efficient algorithm to learn a tree-augmented naive Bayes classifier in a general way from an incomplete dataset. We have also introduced an original formalism to model incomplete dataset generation processes with MCAR or MAR assumptions. Finally, various synthetic datasets and real datasets have been used to empirically compare structure learning methods from incomplete datasets.