Ralf Klinkenberg, RapidMiner
This presentation shows how to leverage machine learning to detect and prevent fraud and make fraud fighters more efficient and effective. The challenge of fighting fraud is that fraudsters often are intelligent, learn from mistakes, and continuously create new types of fraud. Hence, we need techniques that can capture known fraud patterns as well as new types of fraud. A first step is to embed domain expertise and known fraud patterns into entity features and fraud risk scores that can be automatically computed and used to systematically and automatically rank suspects. In a second and more sophisticated step, supervised machine learning, e.g. classification and association learners, can be used to learn detection models for known types of fraud. These models can then be deployed to automatically identify new instances/cases of known fraud patterns/types in the future. In order to also detect potential new fraud patterns and types, we leverage unsupervised machine learning, e.g. anomaly detection and outlier identification techniques. In order to not only identify fraud after it has already happened, but to prevent it, we also use machine learning and predictive analytics, e.g. for predicting expected treatments, medications, quantities, volumes, and costs and comparing them with the actual transactions and requested payments, so that the health insurance can decide to postpone or deny payments, if transactions or combinations of transactions seem highly suspicious, and perform investigations before processing payments. Overall, data mining and machine learning can help auditors and fraud busters to focus on high risk cases instead of wasting time with random checks, allow to integrate and consider various data sources, create meaningful features and scores, provide context and explanations, detect networks of fraudsters, and assist the auditors and fraud busters and make them more effective and efficient.