Extracting local representations from large Bayesian networks

As Bayesian network models continue to grow larger and more complex, the efficiency of probabilistic inference and knowledge representations always cannot meet all demands. In fact, large applications always involve several fields and particular applications may only pay attention to a local domain in these large Bayesian networks within a long period of time. Motivated by this situation, this paper shifts attention to local representations, and concentrates on finding a method of extraction without loss of any information and free of dataset by capturing local graphical characterizations of large Bayesian networks. From the graphical perspective, this paper first captures two local graphical characterizations to determine the skeleton and V-structure of local models respectively. Moreover, several transformation rules are elicited from these local characterizations, and local representations can be obtained though applying these rules to the global network. The local representations can not only offer compact and intuitive representations for local domains concerned, but also allow evidence propagate in a smaller model other than the entire model.