Cloud-Native Deep Learning Framework for Scalable Insurance Fraud Detection and Predictive Analytics

Abstract

The continued escalation of insurance fraud represents a worrying trend for providers in the sector. To mitigate this risk, deep learning-based solutions have emerged to enhance predictive modelling. Nonetheless, deploying these approaches as part of a scalable cloud-native data pipeline remains challenging. This study presents an implementation within one of the core data analytical pillars of an InsurTech company, AiTR, currently in the pre-launch phase. Using a synthetic data generation approach, a set of data sources, including test data, has been made available in increasingly larger instances. The resulting pipeline enables end-to-end processing from a raw version of the input data to the prediction step for fraud likelihood scores through an ensemble of deep and extra tree models. It is integrated with an automated alerting mechanism, making it available for production testing in a live scenario. After a successful test run operating directly on the source transactional database in a pre-production environment, the pipeline has been deployed in a production instance, configured to refresh the prediction output in a dedicated storage solution on a weekly basis. A cron job has been implemented to trigger the Python script, and an additional job has been created to send alerts based on the prediction results. Following the set-up, the service will be in a pre-production phase during which the predictions will be reviewed and compared with actual contact results. Once signed off, the service will enter a live production phase where the predictions are operational and used to trigger contacts with customers.