Automobile Insurance Fraud Detection Using Data Mining: A Systematic Literature Review

Insurance is a pivotal element in modern society, but insurers face a persistent challenge from fraudulent behaviour performed by policyholders. This behaviour could be detrimental to both insurance companies and their honest customers, but the intricate nature of insurance fraud severely complicates its efficient, automated detection. This study surveys fifty recent publications on automobile insurance fraud detection, published between January 2019 and March 2023, and presents both the most commonly used data sets and methods for resampling and detection, as well as interesting, novel approaches. The study adopts the highly-cited Systematic Literature Review (SLR) methodology for software engineering research proposed by Kitchenham and Charters and collected studies from four online databases. The findings indicate limited public availability of automobile insurance fraud data sets. In terms of detection methods, the prevailing approach involves supervised machine learning methods that utilise structured, intrinsic features of claims or policies and that lack consideration of an example-dependent cost of misclassification. However, alternative techniques are also explored, including the use of graph-based methods, unstructured textual data, and cost-sensitive classifiers. The most common resampling approach was found to be oversampling. This SLR has identified commonly used methods in recent automobile insurance fraud detection research, and interesting directions for future research. It adds value over a related review by also including studies published from 2021 onward, and by detailing the used methodology. Limitations of this SLR include its restriction to a small number of considered publication years and limited validation of choices made during the process.