Automobile Insurance Fraud Detection using the Evidential Reasoning Approach and Data-Driven Inferential Modelling

Automobile insurance fraud detection has become critically important for reducing the costs of insurance companies. The majority of insurance companies use expert knowledge to detect fraud. Experience-based knowledge are interpretable and re-usable but the simplistic way that this knowledge is used in practice, often leads to some degree of misjudgment. This paper aims to establish a unique Evidential Reasoning (ER) rule that combines independent evidence from both experience based indicators and probabilities of fraud obtained from historical data. Each piece of evidence is weighted and then combined conjunctively with the weights optimised using a maximum likelihood evidential reasoning (MAKER) framework for data-driven inferential modelling. Based on a real-world insurance claim dataset, our experimental results reveal that the proposed approach preserves the interpretability and usability of expert detection system, and anticipates the changes in fraud practices by tracking the trend of the weights of experience-based indicators. Furthermore, the experimental results show that the proposed approach outperforms a number of widely used machine learning models, such as logistic regression and random forests.