Graphical models for eliciting structural information

The structure of a Bayesian Network is a priori plausible if the directed acyclic graph has one or more plausible structural features. Expert beliefs about the structure of a Bayesian Network may be substantial but limited both to a subset of nodes or to a set of network features indirectly related to network edges. Complex elicitation tasks involving dozens of reference features may be cognitively too difficult for the expert, unless limited subsets of features may be considered at one time. In this paper chain graph models on descriptors of structural features are proposed as a tool to elicit the degree of belief associated to the structure of a Bayesian Network. An algorithm and a parameterization are developed to support the elicitation.