Improving the applicability of the ranked nodes method to build expert-driven Bayesian networks

Abstract

One challenge in constructing a Bayesian network (BN) is defining the node probability tables (NPTs), which can be learned from data or elicited from domain experts. In practice, for large-scale BN it is common not to have enough data for learning and elicitation from experts is unfeasible. Previous work proposed a solution to this problem: the Ranked Nodes Method (RNM). However, this solution needs to be applied by a RNM expert who, through the elicitation of expert judgement, identifies the necessary parameters for the RNM algorithm to generate the NPTs. Hence, this paper presents a novel approach to define NPT using the RNM with no ranked nodes-specific knowledge. The solution is named Simulated Bayesian Network Expert (SBNE). It consists of eliciting a subset of the NPT from the domain experts which is used as input to an algorithm that estimates the optimal parameters for the RNM to generate the NPTs. To validate our solution, we conducted an experiment with multiple domain experts and compared the results with other methods. Our solution outperformed the other methods (producing NPTs at least 12% more accurate) and is, therefore, a promising approach to apply RNM without relying on RNM experts.