Parameter Learning for Bayesian Networks with Strict Qualitative Influences

We propose a new method for learning the parameters of a Bayesian network with qualitative influences. The proposed method aims to remove unwanted (context-specific) independencies that are created by the order-constrained maximum likelihood (OCML) estimator. This is achieved by averaging the OCML estimator with the fitted probabilities of a first-order logistic regression model. We show experimentally that the new learning algorithm does not perform worse than OCML, and resolves a large part of the independencies.