Construction and evaluation of Bayesian networks with expert-defined latent variables

The structure and parameters of a Bayesian network can be determined by learning observed data or by eliciting expert knowledge during the design process. Structures most often contain latent variables when following an expert-driven approach (participatory modelling) to determine causal links between variables. Latent variables play an important role in the design as well as runtime process of a Bayesian network. For example, they are useful to provide intermediate variables and prevent the forming of very large conditional probability tables. Furthermore, experts understand abstract concepts that can be modelled in the network, but for which data is impossible to find. Given the typical placement of latent variables, usually intermediate, with several child and parent nodes, parameterising may not be straightforward. Also, a Bayesian network fully trained with observed data leaves little room for reasoning about stakeholder knowledge and perspectives. Fusion in Bayesian networks can take on three forms during runtime, namely context fusion, observation/measurement fusion and latent fusion. In this paper we encourage the inclusion of abstract latent variables in BN fusion systems by a) listing the considerations for evaluating the uncertainties of such variables b) illustrating a novel elicitation technique for parameterisation of large conditional probability tables and c) framing the uncertainty evaluation in a systems engineering validation and verification process. We present lessons learned from a case study in rhino poaching predictive modelling.