In our third session of fraud office hours, an attendee asked, "How does Unit21 use machine learning?" Watch this video and read below to see how Unit21's Head of Fraud Risk, Alex Faivusovich, responded.
How Unit21 Uses Machine Learning
"We introduced machine learning capabilities earlier this year, and our approach is rather unique.
While you have a lot of vendors and even in-house solutions today in the market, companies who decide to develop their own machine learning models, (which is absolutely fine), are focused on scoring at the transaction level. At Unit21, we decided that we wanted to provide something different.
Scoring On Alerts
The way we approach machine learning today is that we provide scoring on the alerts. So, when your team builds rules that generate alerts, those alerts are going to surface either true positive or false positive results. As such, the team will disposition those alerts in different ways.
Our model learns how your team works those alerts, and based on the way your team treats those alerts, (whether they're true positives, or false positives, and the way they disposition those alerts), the model would still learn how fraud actually looks on your platform. We also look at different attributes of the actual entity associated with the alerts.
Basically, we combine those two elements together. We see all the different attributes of the entity, and we consider how your operating fraud team has been working those types of alerts in the previous months. This is how we come up with the score. And we advise our customers use the score as a way to prioritize work. So, for example, if you have an alert and the associated score with the alert is 95, chances are the alert is fraudulent. And we will advise the team to work on those alerts first because they are probably a higher priority as they have a greater chance of actually being fraudulent.
How Risk & Compliance Teams Can Leverage This
With alert scores—rather than simple, traditional transaction scores—we are changing how teams can manage their alerts and prioritize their cases for better case management. We also provide data on how these scores are assessed—the data used in the machine learning model is available in a human-readable format. So your team isn’t just getting a black-box score that you have to blindly trust. Instead, we are empowering teams to interpret the data that the model is producing and allow you to make informed decisions.
High alert scores are often best to prioritize first, as they are likely to be true positives based on true positives your team has flagged in the past. This helps teams effectively manage their workflow, but it also offers another major benefit. Since the information on how scores are decided is available and human-readable, teams can analyze this information and use it to better refine their detection rules. And, since these alerts are based on what most commonly signals a true positive, this can allow your team—and the detection and prevention systems you use—drastically improve over time.
If you'd like to see a demo of how Unit21 uses machine learning in action, watch fraud office hours on demand!