Constraints on priors and on estimations for learning Bayesian network parameters
Cassio Polpo de Campos, Qiang Ji · 2008
This paper describes a new approach to unify constraints on parameters with training data to perform parameter estimation in Bayesian networks of known structure. The method is general in the sense that any convex constraint is allowed, which includes many proposals in the literature. As prior distributions are also valuable for estimation accuracy, we present a new idea based on maximum entropy and the Imprecise Dirichlet Model to combine priors, constraints and data, and show that estimations can be found using convex programming. Synthetic and real data experiments indicate benefits of this framework.