Abstract
A novel approach to knowledge extraction from neural network classifiers when applied to payment card fraud detection is proposed. Existing Fraud Management Systems (FMS) use neural network classifiers but do not have the ability to explain their learnt patterns of fraud. Rule extraction from such classifiers with a high level of abstraction and linguistic simplicity is proposed. Decompositional knowledge extraction methods are found to be too reliant on the architecture of the fraud classifer and current pedagogical rule extraction methods produce rules that are not sufficiently comprehensible. In this thesis the Sparse Oracle-based Adaptive Rule (SOAR) pedagogical extraction algorithm is proposed to extract generalising rules that explain patterns of fraud. SOAR uses sensitivity analysis to avoid the exhaustive searches of other pedagogical methods. By projecting into discretised space, polytopes are formed by SOAR covering the class convex hull of the classifier surface. A methodological and verifiable empirical evaluation on publicly available datasets in various domains is undertaken. These results show that SOAR extracts comprehensible rules that are sound from a deep learning neural network. When SOAR is applied to large datasets provided by payment card issuers it discovered new fraud types that were of key interest to payment risk/fraud analysts. SOAR provides an improved understanding of fraud vectors that will lead to a more secure payment process through informed payment fraud prevention steps and this work could therefore alter how fraud management is undertaken in the future.