Factorization of computations in Bayesian networks: Interpretation of factors

Abstract

Given a Bayesian network (BN) relative to a set I of discrete random variables, we are interested in computing the probability distribution PS, where the target S is a subset of I. The general idea is to express PS in the form of a product of factors whereby each factor is easily computed and can be interpreted in terms of conditional probabilities. In this paper, a condition statingwhen PS can be written as a product of conditional probability distributions is called a non-pathology condition. This paper also considers an interpretation of the factors involved in computing marginal probabilities in BNs and a representation of the probability target as a Bayesian network of level two. Establishing such a factorization and interpretations is indeed interesting and relevant in the case of large BNs.