Designing Interpretable Explainability Measures for the Cross Model

Wyrley-Birch, Justin (2026) Designing Interpretable Explainability Measures for the Cross Model. Internship report thesis, Data Science and Society (DSS).

PDF Internship-report.pdf Restricted to Repository staff only Download (580kB)

Abstract

FRISS is an insurance technology company that helps P&C insurers detect fraud and assess risk across underwriting and claims. They offer an AI driven SaaS platform where investigators can directly screen claims and see an associated risk score comprised of different indicators and checks. This internship focused on applying explainability measures to the Cross Model, which is a fraud detection model used for new clients or ones which have too little labelled data to train a model on. It uses multiple XGBoost lite models trained on data from other clients to train a broadly applicable ensemble model, where the majority vote from the lite models is the returned prediction. The primary internship assignment was writing a framework for how to introduce effective explainability measures to the Cross Model. This was written by evaluating different types of explanations, requirements and cognitive techniques to improve interpretability of explanations. Ultimately producing example visualisations and suggestions for how these might be implemented. Alongside the primary focus of cross model explainability, the internship also involved contributing to analytical and client facing aspects of business. This included analysing an existing check and developing a standardised notebook for it which can be used for future business reviews. I presented the findings from this work directly to clients, clearly justifying the methodology behind decisions and recommendations, which were subsequently well received and implemented.

Item Type:	Thesis (Internship report)
Name supervisor:	Rodighiero, D.
Date Deposited:	11 Feb 2026 08:17
Last Modified:	11 Feb 2026 08:17
URI:	https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/780

Actions (login required)

View Item