Using the Noisy-OR Model Can Be Harmful... But It Often Is Not

摘要

The noisy-OR model and its generalizations are frequently used for alleviating the burden of probability elicitation upon building Bayesian networks with the help of domain experts. The results from empirical studies consistently suggest that, when compared with a fully expert-quantified network, using the noisy-OR model will just have a minor effect on the performance of a network. In this paper, we address this apparent robustness and investigate its origin. Our results show that ill-considered use of the noisy-OR model can substantially decrease a network's performance, yet also that the model has broader applicability than it was originally designed for.