Check Fraud: Leveraging Both Unsupervised & Supervised Machine Learning Technology
- Machine learning is a vital tool in spotting and preventing fraud
- Two primary approaches deployed by Machine Learning: Unsupervised and Supervised Learning
- The best solution for most FIs is a properly configured combination of the two
Jeremy Chen, senior director of product management at DataVisor, explores the growing need for adoption of machine learning technologies in order for banks to stay one step ahead of fraudsters.
Mr. Chen focuses on machine learning, a subset of artificial intelligence that uses experience and repetition to allow systems to learn and improve. He explores the advantages and challenges of two primary machine learning approaches: Unsupervised Learning and Supervised Learning:
Unsupervised machine learning involves feeding data into the system without any preexisting labels or classifications. The system then identifies patterns and structures within the data, helping institutions detect suspicious activities that may indicate potential fraud. This approach is highly valuable in identifying emerging financial fraud patterns that might not be apparent through traditional methods.
On the other hand, supervised machine learning relies on labeled data, where historical fraud cases are used to train the model. The system learns from past examples and can classify new transactions as either fraudulent or legitimate. Supervised machine learning is particularly effective when dealing with known fraud patterns and is capable of providing accurate predictions.
Unsupervised Learning or Supervised Learning? No "One-Size-Fits-All"
So, how does a bank pick an approach? Well, banks need to evaluate every case.
For institutions dealing with a constant influx of transaction data with no historical knowledge of fraud patterns, unsupervised machine learning can be highly advantageous. It can detect unusual patterns in real time, flagging potential fraud cases that may otherwise have gone unnoticed.
Supervised learning becomes a highly effective approach when there is a clear understanding of fraud patterns that demand a more automated solution beyond self-adaptation methods. For complete, holistic protection, a combined or ensemble approach provides complete coverage against known and unknown threats.
Mr. Chen offers the following primary questions to ask when considering which fraud prevention strategy or provider to leverage:
- Does the fraud and risk platform operate in real time?
- Does the platform offer both SML and UML capabilities?
- What is the detection accuracy rate?
- Can the platform account for missing data and fix data quality issues?
- Does the platform provide human-readable explanations to support your investigators?
For check fraud detection, there is a case for both unsupervised and supervised learning. As we've noted in the past, banks need to take a multi-layered approach to check fraud -- leveraging a multitude of technologies to detect check fraud. Each system, however, will be trained differently.
Take, for instance, transactional or behavioral analytics. Because of the nature of transactions, it's important not to encode biases -- rather, let the system learn patterns on its own to find the anomalous behavior.
In contrast, image forensic AI requires a supervised machine learning system. While a "profile" of good checks are created, whenever a new fraudulent check is detected, a fraud analyst will provide feedback -- adding to the data that the system will look for in other checks. This includes looking for indicators on the check stock, signature, and writing styles.
Banks and credit unions must embrace the power of machine learning to meet the ever-evolving threat of fraud. By harnessing the capabilities of both unsupervised and supervised machine learning, financial institutions can enhance fraud detection, minimize losses, as well as foster trust and confidence among their customers.