Iterative Aggregation of Bayesian Networks Incorporating Prior Knowledge
Jian Xu · OhioLink ETD Center (Ohio Library and Information Network) · 2004
Multi-source information integration has gained significant interest recently.We focus on integrating Bayesian networks (BNs) learned from data.The BN batch aggregation algorithm proposed by Maynard-Reid II and Chajewska (MC01) requires all sources' information be available at aggregation time and does not take the user's prior knowledge into account.We extend this algorithm to make the aggregation iterative, supporting "anytime" querying, and to allow the incorporation of the user's structural prior knowledge.We prove that the iterative extensions for joint distribution aggregation are independent of the order in which sources arrive.We show experimentally that iterative BN aggregation is order-dependent due to bias introduced by the algorithm's optimization nature and an "inertial" effect.However, we show that the results compare well with the batch algorithm in accuracy and efficiency.We also show experimentally that incorporating the user's structural prior knowledge can improve the accuracy and efficiency.