Michael Manapat, Head of Applied Machine Learning, Stripe
Three lessons from building a payment fraud prevention system
Stripe processes billions of dollars in payments a year and uses machine learning to detect and stop fraudulent transactions. In this talk, I’ll cover three lessons we learned as we built our systems for fraud detection:
* Fraud detection performance is as much about how you design your product, and how your product collects information relevant to fraud, as it is about machine learning.
* You need a plan (and budget) for counterfactual evaluation of machine learning model performance. Since your model is altering outcomes—by blocking payments, for example—you won’t always have the data you need to evaluate performance.
* Unstructured data can provide significant signal beyond what you might get with traditional feature engineering, and deep learning can help capture that signal.
Michael is the Head of Applied Machine Learning at Stripe. His group is responsible for fraud detection, financial products, and AI research. Prior to Stripe, he was a software engineer at Google and a research fellow at Harvard.