A New Timescale for Fraud Science: Insights from Our CSO

At Feedzai, we continue to break records in how quickly we begin delivering value to our customers. Speed of data science is important. The faster we can test and deploy new models, the sooner we can test multiple scenarios of rules and models and get to high detection rates. A faster system also makes for a more adaptive fraud solution, one that is future-proofed against new and evolving fraud, and effective in mission-critical scenarios.

In his new report, “A New Timescale for Fraud Science,” Pedro Bizarro, Feedzai’s co-founder and CSO, describes some of Feedzai’s research efforts to speed up the machine learning process. Five years ago, businesses in our market would take a full year to deploy a single model. However, fraudsters change tactics monthly or weekly, not yearly. In order to match the timescale of fraudsters, Feedzai has developed a system that could deploy models in days, rather than in many months.

To continue to speed up our data science work, we’re developing many tools and techniques that optimize every component of our six-stage data science process, which we call our Data Science Loop. For example, to speed up the “Evaluate & Compare” stage, we built a Scenario Optimizer™ feature. It lets data scientists evaluate and compare the different effects of making changes in a combined environment of rules and models.

We’ve also increased the speed of data science work by building an Integrated Data Science Environment, which allows for data scientists to seamlessly “travel” from one stage of work to the next.

Most recently, with a new internal tool called Plotzai™, we’ve managed to speed up our data science work to an entirely new timescale. As Pedro writes in his report, Plotzai has made it 500 times faster for our data scientists to perform data analysis, during both the “Analyze & Sample” stage and the “Evaluate & Compare” stage.

How does Plotzai work? We’ve optimized the data science technique called “drill up drill down” analysis by lacing a easy-to-use graphical interface over massive and complex data sets. Plotzai makes it easy for our data scientists to zoom up and back as they explore data and uncover insights that were previously invisible.

Pedro writes, “Many systems don’t even have the concept of model performance that can be broken down by any dimension. But breakdown ability is native to our system. By breakdown ability I mean what’s traditionally called “drill down drill up analysis.” This concept is not new. It’s maybe 40 years old. What is new is that we’ve been able to combine enterprise-scale, “drill down drill up” functionality within the machine learning cycle of things.”

Reflecting on Feedzai’s earliest vision of easy, large-scale fraud detection, Pedro Bizarro paints a picture of the next stage in this vision: the transformation of machine learning into “machine teaching.”

As AI technology grows more powerful, and as we begin to trust the machine with more and more decisions, we will find that machine learning is teaching us what’s going on in increasingly human terms. As a result, humans will be driven toward increasingly strategic and creative decision-making tasks.

Pedro writes: “Yes, we are very, very far from a conversational intelligent agent that explains to us what is going on in human terms. But before sending a rocket to the moon, the Chinese invented firecrackers that flew just a few meters. Working at Feedzai has taught me that the possibilities are endless.”

To learn more about this new timescale of fraud science, read the report: “A New Timescale for Fraud Science: Insights from Our CSO.