Efficient Principled Learning of Thin Junction Trees
Poster M45 Anton Chechetka and Carlos Guestrin Carnegie Mellon University Motivation: from data, learn the structure of a provably good probabilistic graphical model that allows tractable exact inference Learning the best model is NP-complete However, our algorithm provides  global quality guarantees (small KL from the true distribution)  in polynomial time  PAC learnability for strongly dependent JTs  promising empirical results

Learning

Data
Junction t re e structure