Factorized NML models
Petri Myllymäki, Teemu Roos, Tomi Silander, Petri Kontkanen and Henry Tirri
Festschrift in Honor of Jorma Rissanen on the Occasion of his 75th Birthday
Tampere International Center for Signal Processing
, Tampere, Finland
ISBN 978 952 15 1962 8
We consider probabilistic graphical models where a directed acyclic graph represents a factorization of a joint probability distribution: the joint probability of the variables is represented as a product of conditional probabilities, one for each variable conditioned on its immediate parents in the graph. For this type of models, computing the normalized maximum likelihood (NML) is computationally very demanding. We suggest a computationally feasible alternative to NML, the factorized NML, where the normalization is done locally for each conditional distribution, and not globally.