The paper presents a unified approach to learning Bayesian networks from both user knowledge and statistical data. It introduces two key properties, *event equivalence* and *parameter modularity*, which simplify the encoding of prior knowledge. The authors develop scoring metrics based on these properties, allowing users to express their knowledge as a single *prior Bayesian network*. The paper focuses on directed belief networks, where the metric scores directed networks, and provides a detailed derivation of the metric, including assumptions about parameter independence and Dirichlet distributions. The metric is evaluated using the Alarm network in the domain of ICU ventilator management, demonstrating its effectiveness in learning accurate network structures from limited data. The results show that the metric performs well, even with noisy prior knowledge, and that the optimal equivalent sample size can be calibrated to balance the use of prior knowledge and statistical data.The paper presents a unified approach to learning Bayesian networks from both user knowledge and statistical data. It introduces two key properties, *event equivalence* and *parameter modularity*, which simplify the encoding of prior knowledge. The authors develop scoring metrics based on these properties, allowing users to express their knowledge as a single *prior Bayesian network*. The paper focuses on directed belief networks, where the metric scores directed networks, and provides a detailed derivation of the metric, including assumptions about parameter independence and Dirichlet distributions. The metric is evaluated using the Alarm network in the domain of ICU ventilator management, demonstrating its effectiveness in learning accurate network structures from limited data. The results show that the metric performs well, even with noisy prior knowledge, and that the optimal equivalent sample size can be calibrated to balance the use of prior knowledge and statistical data.