Human Centric Explainable AI: Operationalising Social Science Explanatory Techniques to Improve Fraud Prediction Model Interpretability

Wyrley-Birch, Justin (2026) Human Centric Explainable AI: Operationalising Social Science Explanatory Techniques to Improve Fraud Prediction Model Interpretability. Bachelor thesis, Data Science and Society (DSS).

PDF BA5602521JWyrley-Birch.pdf Restricted to Repository staff only Download (1MB)

Abstract

Fraud detection in insurance increasingly relies on machine learning models that flag suspicious claims for review. These models are accurate but difficult to understand, and a flagged claim carries real consequences for policyholders. Claim investigators are expected to scrutinise model outputs, yet a poorly designed explanation can leave them over-reliant on a prediction they do not understand. This thesis treats explanation as a social process rather than a technical output, drawing on Miller's (2019) account of social science informed explanations. It asks how these techniques can make SHAP-based explanations of an ensemble fraud model more interpretable for non-technical end users. The study was a qualitative, iterative design conducted with FRISS, refining two explanation artefacts across interviews with six client-facing participants. The two principles fared differently. The selective principle was well received throughout, and the local SHAP waterfall became the artefact all six relied on. The contrastive principle struggled in its first form, since comparing whole claims added cognitive load rather than insight. A later redesign reframed the comparison at the level of model behaviour relative to a fraud alert threshold, which appeared to help. Across the study, these explanations gave participants a clearer view of why the model had flagged a claim and a firmer basis for scrutinising it. Both principles appear to hold explanatory value, and the techniques used to operationalise them can inform how explanations are designed beyond this context.

Item Type:	Thesis (Bachelor)
Name supervisor:	Schauble, J.K.
Date Deposited:	10 Jun 2026 11:46
Last Modified:	10 Jun 2026 11:46
URI:	https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/818

Actions (login required)

View Item