Probabilistic graphical models parameter learning with transferred prior and constraints

Learning accurate Bayesian networks (BNs) is a key challenge in real-world applications, especially when training data are hard to acquire. Two approaches have been used to address this challenge: 1) introducing expert judgements and 2) transferring knowledge from related domains. This is the first paper to present a generic framework that combines both approaches to improve BN parameter learning. This framework is built upon an extended multinomial parameter learning model, that itself is an auxiliary BN. It serves to integrate both knowledge transfer and expert constraints. Experimental results demonstrate improved accuracy of the new method on a variety of benchmark BNs, showing its potential to benefit many real-world problems