Josh Johnston: I get asked sometimes, "Why is fraud an artificial intelligence problem? How is it hard enough? Why is it that computers aren't better at this?"
There's really, I think two reasons. One reason that fraud is hard, is that there's not an easy answer. When you're looking for landmines, digging a hole is pretty cheap. If you're not sure whether there's a landmine there or not, you say, "Yeah, I think there's a landmine there," and somebody goes and digs a hole.
If they dig an empty hole that's okay. That's a lot better than missing it. With fraud, we don't have an easy answer like that. If we take a transaction and we say, this is fraud, somebody declines it. If that's a false positive, that costs a lot of money. You lose that customer, possibly forever, all the marketing costs that you've sunk to acquire that customer. You lose the specific sale that you're working on, bad reputation in the marketplace.
The same thing if we let fraud go through. You get lost merchandise. You have to pay the person back. You get the chargeback fee. You might get on a chargeback program. There's not an easy way to answer this. That means that we really do have to be right about fraud. We can't err on one side or the other.
The other thing that makes fraud really hard, is that there's this tremendous class imbalance. Almost every transaction is a good transaction. Fraud is a very small part of the business.
If you have three cups of flour, and you have one teaspoon of salt, this one teaspoon of salt that you mix into three cups of flour is about 7/10th of a percent. 70 basis points, which is about the fraud rate that most eCommerce merchants would see. If you imagine that I mix these two together, and then later I'm trying to sort them back out, I need to have a big pile of flour, that's my good transactions and little pile of salt, that's my chargebacks. That's really hard to do. There's not a lot of problems that are like that.