Key takeaways
- How rule-based fraud systems break down: limited features, slow manual updates, and high false-positive costs that erode customer trust and inflate investigation budgets
- How Driverless AI automates feature engineering, turning 11 raw fields into over a thousand engineered features in minutes rather than the weeks a manual team would need
- How machine learning interpretability, including global and local explanations, Shapley values, and disparate impact analysis, keeps new models explainable to regulators and customers
- How models deploy as a MOJO for real-time scoring, in this demo running a prediction in 0.071 milliseconds
- How PayPal used Driverless AI to raise fraud model accuracy from 89 to 95 percent, cut model development time by a factor of six, and auto-generate documentation for regulatory approval
- How automatic oversampling handles the class imbalance common in fraud data, where fraudulent transactions are a small fraction of total volume
Overview
Fraud cost the UK economy billions in 2019, with 3.8 million reported instances and rising, yet many detection systems still rely on a handful of features and slow-to-update rules. In this session, John Spooner walks through how H2O.ai's Driverless AI automates feature engineering and model building to catch more fraud without slowing down decisions. He demos turning a raw transaction table into a scored model in minutes, then shows how PayPal used the same approach to lift model accuracy from 89 to 95 percent while cutting development time sixfold.