Bridging the Gap

This thesis investigates ways in which domain knowledge can mediate between opaque ML models and specialist users in high-risk domains, such as law enforcement and counter-terrorism. These users are domain experts in their field, but not necessarily in AI itself. The central challenge lies in reconciling the technical complexity of models with domain experts' mode of reasoning on the basis of knowledge about their domain. This conceptual gap is exacerbated by a trade-off between performance and interpretability. Opaque models often achieve superior performance scores, yet resist meaningful interpretation. Fully interpretable approaches may sacrifice the very performance that makes automation efforts worthwhile. In high-risk domains, this trade-off is particularly concerning, given that the consequences of wrongful decisions can be severe. Also, such applications of ML often involve legal requirements pertaining to transparency and interpretability. As a result, it is difficult to produce a high-performing model that `speaks the same language' as the domain experts who rely on it. This thesis demonstrates how domain knowledge about terrorist modes of operation can be structured systematically using a story-scheme ontology. The prototype developed for this purpose allows for rule-based classification that maintains a certain degree of interpretability. It also reveals a fundamental limitation: the knowledge acquisition bottleneck, inherent to this type of work, remains significant, and representing real-world complexity within rigid structures continues to be a challenge. This serves to motivate the use of explanation or justification techniques in the rest of this thesis. A model-agnostic approach known as `AF-CBA' is considered. AF-CBA generates domain-specific justifications for binary classification predictions through case-based argumentation, constructing dispute trees that can be presented to users. Two modifications of AF-CBA improve its applicability. The first is the introduction of authoritativeness, an expression of the degree to which a precedent is supported or contradicted, which helps mitigate the effects of precedential inconsistency. The second is the optional reliance on archival data rather than a model's training set. The effectiveness of AF-CBA in providing domain-specific justifications hinges on the availability and interpretability of domain knowledge concerning preference relations between data features. This is addressed through a secondary argumentation framework that supports reasoning about such preferences, combined with a semi-automatic approach to knowledge discovery that produces potential preference relations. The resulting body of domain knowledge remains open for further refinement by domain experts, which would create a feedback loop. These modifications raise an obvious question: would AF-CBA be usable as an interpretable classification model in its own right? Experiments suggest that it can indeed function as such, comparable to decision trees both in traditional performance metrics and, arguably, in interpretability. Situations in which AF-CBA proves to be an unsatisfactory classification model instead contribute to the motivation for its original role as a justification approach for higher-performing opaque models.