Our fraud prevention models at Riskified are masters of identifying suspicious patterns in structured data. But what happens when the most crucial information is buried in unstructured text? Unfortunately, for our tree-based models, "the more that you say, the less I know."
When a major concert ticket merchant came on board, we found ourselves with a critical blind spot: our model couldn't effectively use the event titles to spot fraudulent orders. A free-text field with infinite possibilities is an impossible input for a tree-based model. We needed a solution that could translate the rich information in these titles into a single, usable signal. Eventually, we designed a simple neural network framework to translate any textual input into a risk score that could be seamlessly integrated into our existing tree-based models.
In this talk, I will share our journey from a specific, urgent problem to a generic, reusable solution. We'll see how our solution enabled our team to reuse the same framework to solve an entirely different problem: recognizing obfuscated addresses. You will leave with a practical understanding of how to proactively design for reusability, turning a single-purpose project into a powerful, adaptive tool for your entire organization.