Practical Fraud Prevention

Chapter 5. Machine Learning and Fraud Modeling

You keep using that word. I do not think it means what you think it means.

Inigo Montoya, The Princess Bride¹

Before machine learning became a buzzword, there were rules. Most fraud analysts still remember using rules exclusively in their fraud automation, and rules continue to be important. But nowadays, they’re often used in combination with machine learning. The dominance of machine learning can make it feel like it has been in place for years, but it’s actually a relatively new development.

Up until the early 2010s, virtually no fraud team was using machine learning, though many organizations boasted about their early experiments in the field. One of the true pioneers—Fraud Sciences, acquired by PayPal in 2008—made it a habit to hold routine brainstorms to promote a steady stream of “creative juices” as its teams were building the first fraud classifiers. Each and every manually investigated fraud transaction would be inspected by a fraud analyst, who needed to suggest ideas for a model feature (or at the very least, select the most prominent “reasons for analyst decisions” so that those reasons would be turned into features in the model; a list of guesstimated options would then be presented on the in-house fraud investigation UI). Such practices boosted the art and science of fraud detection and helped it scale up rapidly.

Even five years ago, machine learning still felt new, and not every company was willing to take the plunge ...