Explaining anomalies : an approach to anomaly-driven revision of a theory

This thesis focuses on the explanation of anomalies as an approach to anomaly-driven revision of a theory. An anomaly is identified when a theory (or model of a domain) does not accurately reflect a domain observation, indicating that the theory (or model) requires refinement. In some cases an explanation can be generated for an anomalous observation using existing domain knowledge and hence a revision to the existing theory can be provided. Ontologies have been used in both stages of an investigation presented in this thesis; in the first stage, a domain ontology and expert-acquired strategies have been used as part of a knowledge-based system, EIRA (Explaining, Inferencing, and Reasoning about Anomalies), to generate explanations for an anomaly; in the second stage, domain ontologies have been used to suggest refinements to an incorrect or incomplete domain ontology. In the first stage of the investigation, extensive interviews were held with domain experts; the analysis of which led to the identification of both examples of anomalies encountered in the domain and the strategies used by the domain experts to provide (appropriate) explanations for the anomalies. EIRA is able to replicate these explanations; when EIRA is presented with an anomaly, potential explanations are generated by the application of expert-acquired strategies to the domain knowledge, patient data, and information about the clinical situation. To evaluate this approach, EIRA has been applied in the Intensive Care Unit (ICU) domain and ICU clinicians have evaluated the explanations produced by EIRA. The strategies used by EIRA have been abstracted further to form generic strategies for anomaly resolution. In the second stage, EIRA has been extended to investigate the use of domain ontologies to suggest refinements to an incomplete or incorrect ontology. These additional refinements are generated by reasoning about analogous concepts from the domain ontology. The findings described in this thesis support the belief that ontologies can be used to generate explanations to refine a theory, further, that the extensive domain knowledge contained in an ontology allows for sophisticated refinements of a knowledge base. Previous approaches to theory revision have largely focused on the refinement of an instantiated rule base, in which limited domain knowledge is incorporated in the rules and hence the refinements are essentially captured in a particular knowledge base. In these earlier approaches, refinements to remove the anomaly were generally suggested after applying machine learning techniques on data from the domain; however, this process requires large datasets, the refinements generated are not always acceptable to domain experts, and providing explanations (using an ontology) to account for anomalies have not been investigated. I believe that the findings reported in this thesis are significant and make a number of contributions including a novel approach to anomaly-driven revision of a theory.EThOS - Electronic Theses Online ServiceGBUnited Kingdo