I’ve got an interesting question at one interview.
Assume that we already trained and deployed some fraud detection model for some online service, and it has helped us to decrease amount of fraudulent transactions by 50% (for example, we temporarily ban suspicious cards).
This is great, but how can we deal with the fact that we now don’t have any explicit feedback for transactions that our model marked fraudulent?
How can we retrain our model keeping this in mind?
I see pseudo-labeling such transactions as one possible solution. And, probably, it would be good to use soft labels for them to deal with the cases when our model was not very confident with its decision.
But are there any other options? And what are real-world solutions to this?
I did’t get much during the research of the internet. Most of articles mention concept drift in terms of user behaviour change, but I haven’t seen my particular question to be discussed anywhere.
I haven’t had real-world fraud detection experience, so it would be really interesting to hear your opinion.
