The impact of overconfidence bias on practical accuracy of Bayesian network models: an empirical study

In this paper, we examine the influence of overconfidence in parameter specification on the performance of a Bayesian network model in the context of Hepar II, a sizeable Bayesian network model for diagnosis of liver disorders. We enter noise in the parameters in such a way that the resulting distributions become biased toward extreme probabilities. We believe that this offers a systematic way of modeling expert overconfidence in probability estimates. It appears that the diagnostic accuracy of Hepar II is less sensitive to overconfidence in probabilities than it is to underconfidence and to random noise, especially when noise is very large.