Advancing incorporation of expert knowledge into Bayesian networks

Bayesian networks (BNs) are used in many areas to support risk management and decision-making under uncertainty. A BN represents probabilistic relationships of variables and allows to explore their interaction through various types of analyses. In applications, a lack of suitable data often necessitates that a BN is constructed at least partly based on the knowledge of a domain expert. Then, in order to manage limited time and the cognitive workload on the expert, it is vital to have efficient means to support the construction process. This Dissertation elaborates and improves so-called ranked nodes method (RNM) that is used to quantify expert views on the probabilistic relationships of variables, i.e., nodes, of a BN. RNM is designed for nodes with discrete ordinal scales. With such nodes, the relationship of a descendant node and its direct ancestors is defined in a conditional probability table (CPT) that may consist of dozens or hundreds of conditional probabilities. RNM allows the generation of the CPT based on a small number of parameters elicited from the expert. However, the effective use of RNM can be difficult due to a lack of exact guidelines concerning the parameter elicitation and other user-controlled features. Furthermore, there remains ambiguity regarding the underlying theoretical principle of RNM. In addition, a scarcity of knowledge exists on the general ability of CPTs generated with RNM to portray probabilistic relationships appearing in application areas of BNs. The Dissertation advances RNM with regard to the above shortcomings. The underlying theoretical principle of RNM is clarified and experimental verification is provided on the general practical applicability of the method. The Dissertation also presents novel approaches for the elicitation of RNM parameters. These include separate designs for nodes whose ordinal scales consist of subjective labeled states and for nodes formed by discretizing continuous scales. Two novel approaches are also presented for the discretization of continuous scales of nodes. The first one produces static discretizations that stay intact when a BN is used. The other one involves discretizations updating dynamically during the use of the BN. The theoretical and experimental insight that the Dissertation provides on RNM clears the way for its further development and helps to justify its deployment in applications. In turn, the novel elicitation and discretization approaches offer thorough and well-structured means for easier as well as more flexible and versatile utilization of RNM in applications. Consequently, the Dissertation also facilitates and promotes the effective and diverse use of BNs in various domains.