In automobile insurance industry, fraudulent incidents and claims bring tremendous loss to the insurance company and have gained increased attention. Collaborating with an automobile insurance company in mainland China, we extract the data of great volume and variety from their business system. By mining these data, a fraud detection model is built. Both supervised and semi-supervised learning algorithms are applied in our system. A comparison is made and the results show that semi-supervised approach outperforms the others. In the test set, 81% frauds fall into the cases with the top 10% highest fraud score predicted by our model. It demonstrates the capability of our system to guide the insurance reviewers to carry out the follow up actions, i.e., effciently allocating limited resou...[ Read more ]

In automobile insurance industry, fraudulent incidents and claims bring tremendous loss to the insurance company and have gained increased attention. Collaborating with an automobile insurance company in mainland China, we extract the data of great volume and variety from their business system. By mining these data, a fraud detection model is built. Both supervised and semi-supervised learning algorithms are applied in our system. A comparison is made and the results show that semi-supervised approach outperforms the others. In the test set, 81% frauds fall into the cases with the top 10% highest fraud score predicted by our model. It demonstrates the capability of our system to guide the insurance reviewers to carry out the follow up actions, i.e., effciently allocating limited resources to investigate the cases with higher fraud probability to save money from frauds while quickly settling the cases with lower fraud probability to improve customer satisfaction. The system has been tested and deployed in the company and it turns out to perform much better than the counterpart fraud detection system built by FICO.