Probabilistic Inference Using Function Factorization and Divergence Minimization
Terence Chan, Raymond W. Yeung · Birkhäuser Boston eBooks · 2011
This chapter addresses modeling issues in statistical inference problems. We will focus specifically on factorization model which is a generalization of Markov random fields and Bayesian networks. For any positive function (say an estimated probability distribution), we present a mechanical approach which approximates the function with one in a factorization model that is as simple as possible, subject to an upper bound on approximation error. We also rewrite a probabilistic inference problem into a divergence minimization (DM) problem where iterative algorithms are proposed to solve the DM problem. We prove that the well-known EM algorithm is a special case of our proposed iterative algorithm.