We presented exact solutions for Bayesian networks associated with global neighborhoods. Such neighborhoods specify perturbations in the joint probabilistic model, and are useful when one seeks to assess the global influence of parameter variations and independence assumptions. Inference bounds must be carefully analyzed, since global perturbations over multivariate structures may lead to very large intervals of probability.
Bayesian networks have not been analyzed with respect to robustness to structure, i.e., how the inferences degrade as the structure of the network loses accuracy. Such analysis must be part of inference procedures, but so far the classes that admit exact inference have been restrictive or inefficient. This paper offers the first analysis of global neighborhoods for Bayesian networks, demonstrating that some classes of models are tractable for inferences.
Thu Jan 23 15:54:13 EST 1997