Back to Video

Yigit Yildirim: The model is very, very precise because we tried very hard to decrease the bias and variance that might be inherent in many machine learning models or so in some more expert driven systems at the same time. So we use a method called human-in-the-loop machine learning. Basically, what it does is it calls for human experts to get involved whenever the model feels uncertain or it’s not able to get the information that it’s looking for.

So, in those cases, we ask our fraud investigators to come in and really provide us insights and more training data to be able to be more precise. At the same time, when fraud investigators are not ready, we rely on our consortium labels to be able to decrease their bias in their investigations. So it goes both ways.

It’s cyclical and it’s iterative. So whenever the model is not certain about making decisions, we use humans. When humans are not certain and they’re not agreeing with each other, we use very sophisticated machine learning techniques to be able to decrease the bias and also inherently variance on it.