Recovering the Global Structure from Multiple Local Bayesian Networks

Bayesian networks are powerful tools for common knowledge representation and reasoning of partial beliefs under uncertainty. In the last decade, Bayesian networks have been successfully applied to a variety of problem domains and many Bayesian networks have been established. Confronted with many real-world applications, each Bayesian network established may be a local model of the whole knowledge domain. It is desirable to combine all local models of the whole domain into a global and more general representation. It is not realistic to expect the domain experts to construct the global model manually due to the too broad domain. As well, it is not feasible to relearn the model since the dataset may have been discarded or the whole domain may be distributed. Thus, constructing the global model by combining local models is doomed to be an alternative solution. This paper concentrates on finding a method of combination without loss of any information and free of datasets by capturing the graphical characterizations of global models. From the graphical perspective, this paper first captures two graphical characterizations to determine the skeleton and V-structure of global models. Moreover, a simple algorithm is elicited from these graphical characterizations to recover the global underlying model from multiple local models. A preliminary experiment demonstrates empirically that our algorithm is feasible.