Big Data is a hot topic today that stems back to the early days of high-performance computing and parallel computing, which I worked on during my time in theoretical physics at Duke and Los Alamos. These days, Big Data tools facilitate the ease in applying these concepts. Interestingly, much of the discussion around Big Data focuses on the size of data, but not as much on the fact that it’s changing the analytics paradigm. That paradigm shift centers around analytics “living in the stream”.
Streaming analytics is no stranger to FICO, and one of the best examples is with our fraud detection solution, FICO® Falcon® Fraud Manager. Falcon models rely on transaction profiles that summarize data in the stream as it passes by, in order to compute the pertinent fraud feature variables without relying on the persistence of data in production.
There is such power in this concept that we at FICO continue to drive innovations in streaming profiles for our fraud products. These advances include our global intelligent profiles, which identify abnormal behavior from cardholders as well as entities like ATMs and merchants, and self-calibrating profiles, which improve detection accuracy where service/channel usage and other customer behaviors are changing.
Another major impact of Big Data is that analytics must reduce reliance on persistent data, and allow analytic models to adjust on the fly in the stream. To meet the need of an increasingly dynamic stream, we have focused research efforts on self-learning techniques such as adaptive analytics and self-calibrating analytics. We strongly believe these are critical technologies to supplement traditionally trained neural network fraud models, which rely on persistent data. Self-learning technologies may even eventually replace neural networks in some regions.
Along these lines, we recently made major architectural and other improvements to our self-calibrating analytics technology. This patent-pending “multi-layered self-calibrating analytics” has a model architecture that resembles that of standard neural network models, but unlike neural networks, the models are fully self-calibrating in the stream. Tuning the multi-layered self-calibrating models also requires much less data examples than neural nets, and they can be coupled directly with adaptive analytics for even more dynamic fraud detection.
We see tremendous predictive lift when we benchmark the multi-layered self-calibrating models against our current self-calibrating technology. Compared to traditional neural network models, we often see better performance outside of model training period for certain portfolios dependent on how data patterns shift and change.