Machine learning (ML) is effective at identifying both known and novel fraud patterns in a variety of domains, from financial to insurance to retail. Real-time fraud detection, however, brings many challenges related to predictable real-time data processing and shifting left as much feature computation as possible.

In this webinar, we will look at different ways to implement the queen of features for real-time fraud - rolling aggregations. Rolling aggregations can capture trends and patterns in a compressed representation, enabling solutions in low data regimes, and meet the lowest latency requirements for feature freshness, enabling the most reactive real-time AI systems.

We will look at how legacy approaches, such as sliding-windows and tiled-aggregations, are being replaced by incremental views. In particular, we will look at how the open-source Feldera streaming engine can be used with Hopsworks to compute rolling aggregations at scale, helping solve the most challenging real-time fraud problems.