Σωτήριος Τασουλής, Επίκουρος Καθηγητής του Τμήματος Πληροφορικής με Εφαρμογές στη Βιοϊατρική του Πανεπιστημίου Θεσσαλίας
«Fraud detection using statistical machine learning»
Fraud detection is a long-standing challenge for the insurance industry. Still, the problem remains far from being resolved while the consistently increasing services, the client personalization and the multilevel nature of particular domains impose further complexity. For example, fraud and abuse occur at many points in the healthcare system involving doctors, hospitals, nursing homes, diagnostic facilities and medical equipment suppliers. Traditionally, we deal with fraud detection by developing heuristics around fraud indicators such as rules that define if a case needs to be sent for investigation. In some cases, there is a checklist with scores for the various indicators of fraud which are designed manually. Unfortunately, relying heavily on manual intervention results in several constraints, such as the operation with a limited set of known parameters and the inability to understand context-specific relationships. In the meantime, recalibration needs to take place relatively often and still, discovery of new fraud variations is almost impossible. Leveraging the Machine Learning Capability for this domain seems like the only way to go. Exploring the relevance of the data elements without any prior knowledge is of particular value for discovering patterns in unusual behaviours. Additionally, based on identified frauds the development of a predictive model is more straightforward than ever nowadays with the availability of recent techniques that can guarantee robust results even when we have to deal with extremely unbalanced classes. Deep Learning techniques have been very promising for fraud detection where a relatively new research domain is expanding, namely “Deep Anomaly Detection”. It looks like a permanent solution to this problem could find its basis in Machine Learning while recent tools that allow faster integration and communication can persuade the industry to adopt Machine Learning in their workflow.