Eliciting Bayesian networks via online surveys: a new approach to knowledge elicitation

Bayesian Networks (BNs) are popular computer models used to perform reasoning under uncertainty. They are popular as they are one of the few computer models able to be constructed by analysing historical data or expert elicitation. The main contribution of this thesis is a new workflow for expert elicitation of BNs, aiming to reduce the associated knowledge elicitation bottleneck. This bottleneck is caused by logistical issues associated with interviewing multiple experts, resolving conflicting opinions among experts, and the combinatorial explosion that occurs when eliciting probabilities. The workflow proposed in this thesis brings together research from the fields of BN construction, knowledge acquisition, the survey methodology, and crowd sourcing. The end result is a workflow for conducting online surveys (SEBN) which has been implemented in an open source online survey tool called BN Elicitator (BNE). This workflow allows a greater number of experts to contribute to the construction of BNs compared to traditional elicitation approaches. It also reduces the workload of the knowledge engineer facilitating the BN elicitation and minimizes the time required of each expert contributing knowledge to the BN. Two evaluation surveys were conducted using the BNE software to measure how successfully the newly proposed technique was able to elicit BNs. The results of these evaluations showed a small improvement over the currently employed methods of eliciting BNs from experts (primarily face to face interviews) but also many undesirable outcomes. The improvements obtained allowed the BNE software to facilitate the elicitation of a BN was faster than would have otherwise been required for face to face interviews. However, the BN resulting from the evaluation did not compare favourably to an existing published BN. Issues with the methods used to collate survey responses into a BN were identified and discussed, such as the choice of crowd sourcing algorithms. Such issues highlight the need for further research in this area to improve the accuracy of BNs elicited using online surveys.