A Cost-Minimization approach to Automobile Insurance Fraud Detection

Motor insurance fraud is the most common type of insurance fraud. It is estimated that while 10% to 20% of all claims are fraudulent [1], 21% to 36% of all motor insurance claims may contain elements of fraudulent activity. Detecting all fraudulent cases is difficult because of the limited human and financial resources businesses have at their disposal. Human resource cost can be reduced by deploying automated systems to detect potential fraud and then having humans do a closer inspection of these cases. We address this issue by applying Machine Learning to fraud detection but, instead of the traditional objective of maximizing detection accuracy, we instead minimize the total cost of detecting fraud. In this way we take into account the savings associated from detecting fraud balanced with the cost of human resources as well as the cost of inconveniencing customers by mistakenly accusing them of fraud. We employed a semi-supervised cost-sensitive learning framework where Isolation Forest detected anomalous cases and weighted logistic regression was used to minimize cost to the business.