An analysis of bayesian network model-approximation techniques

Two approaches have been used to perform approximate inference in Bayesian networks for which exact inference is infeasible: employing an approximation algorithm, or approximating the structure. In this article we compare two structure-approximation techniques, edge-deletion and approximate structure learning based on sub-sampling, in terms of relative accuracy and computational efficiency. Our empirical results indicate that edge-deletion techniques dominate the subsampling/induction strategy, in both accuracy and performance of generating the approximate network. We show, for several large Bayesian networks, how edge-deletion can create approximate networks with order-of-magnitude inference speedups and relatively little loss of accuracy.