Verifying monotonicity of Bayesian networks with domain experts

Abstract

In many realistic problem domains, the main variable of interest behaves monotonically in the observable variables, in the sense that higher values for the variable of interest become more likely with higher-ordered observations. This type of knowledge appears to naturally emerge from experts during knowledge elicitation, without explicit prompting from the knowledge engineer. The experts’ concept of monotonicity, however, may not correspond to the mathematical concept of monotonicity in Bayesian networks. We present a method that provides both for verifying whether or not a network exhibits the properties of monotonicity suggested by the experts and for studying the violated properties with the experts. We illustrate the application of our method for a real Bayesian network in veterinary science.