MASTERÂ Assignment
Evaluating Generalisability, Limitations, and Adaptations in Social Network Analytics-Based Insurance Fraud Detection
Type : Master M-CS
Period: Sep, 2023 - Feb, 2024
Student : Schrijver, G. (Gilian, Student M-CS)
Date Final project: February 20, 2024
Supervisors:
Abstract:
Despite insurance companies detecting a large worth of fraudulent insurance claims, detected insurance fraud is assumed to constitute only a small fraction of all insurance fraud. Meanwhile, a large number of claims flagged as `potentially fraudulent' by fraud detection systems are considered benign after manual review, indicating a high error rate. These combined observations reveal that, at least in theory, there is room for improving current fraud detection systems. In the current research, we extend upon a recently proposed social network analytics-based approach to automobile insurance claims fraud detection. This approach leverages the BiRank algorithm to calculate fraud scores in a graph of claims and stakeholders, which are then combined with other features to train a supervised machine learning classifier. We first establish that our real insurance data also suggests empirical evidence for the homophily assumption proposed by the original work's authors. We then reconstruct their proposed model and corroborate their finding that the inclusion of network-related features enhances fraud classification performance. As an extension, we assess the impact of incorporating time-weighted fraud influence and extending the graph with relations based on shared resources and reveal that our current approach yields limited additional value over the baseline model. Meanwhile, we identify limitations in the original work's methodology and experimental setup. The results of this study provide a deeper understanding of the value of using graph-based insurance fraud detection techniques in practice. These insights shall ultimately aid in saving insurers and their customers from the financial consequences of fraudulent claims.