Approaches for Reducing Expert Burden in Bayesian Network Parameterization.

Bayesian networks (BNs) are popular models that represent complex relationships among variables. In the discrete case, these relationships can be quantified by conditional probability tables (CPTs). CPTs can be derived from data, but if data are not sufficient, experts can be involved to assess the probabilities in the CPTs through Structured Expert Judgment (SEJ). This is often a burdensome task due to the large number of probabilities that need to be assessed and the structured protocols that need to be followed. To lighten the elicitation burden, several methods have previously been developed to construct CPTs using a limited number of input parameters, such as InterBeta, the Ranked Nodes Method (RNM), and Functional Interpolation. In this study, the burden/accuracy trade-off of InterBeta is researched by applying the method to reconstruct previously elicited CPTs and simulated CPTs, first by comparing these CPTs to ones constructed using RNM and Functional Interpolation. After that, InterBeta extensions are proposed and tested, including an extra mean function (shifted geometric mean), the elicitation of additional middle rows, and the newly proposed extension ExtraBeta. InterBeta with parent weights is found to be the best-performing method, and the ExtraBeta extension is found to be promising and is proposed for further exploration.