Extending factor graphs so as to unify directed and undirected graphical models

摘要: The two most popular types of graphical model are Bayesian networks (BNs) and Markov random fields (MRFs). These offer complementary properties in construction, expressing conditional independencies, arbitrary factorizations joint distributions, formulating messagepassing inference algorithms. We show how the notation semantics factor graphs (a relatively new type model) can be extended so as to combine strengths BNs MRFs. Every BN or MRF easily converted a graph that expresses same factorization distribution, used for probabilistic through application single, simple message-passing algorithm. describe modified "Bayes-ball" algorithm establishing independence graphs, we form strict superset In particular, give an example commonly-used fragment, whose independencies cannot represented MRF, but graph. For readers who use chain further extension enables them represent graphs.