An abstraction scheme is developed to simplify Bayesian belief network structures for future inference sessions. The concepts of abstract networks and abstract junction trees are proposed. Based on the inference time efficiency, good abstractions are characterized. Furthermore, an approach for automatic discovery of good abstractions from the past inference sessions is presented. The learned abstract network is guaranteed to have a better average inference time efficiency if the characteristic of the future sessions remains moreorless the same. A preliminary experiment is conducted to demonstrate the feasibility of this abstraction scheme.