In The Kitchen With Fraud
By: Josh Johnston, Data Scientist at Kount
When I want to spend all day watching something cook, I usually smoke a pork shoulder or nice rack of ribs. I don’t have the patience to watch bread rise. I do make a ridiculous falafel with tzatziki, though, which demands a quality flatbread.
My flatbread recipe uses 3 cups of flour. There are a few billion grains of flour in those three cups. That’s how many transactions there are in the database we use to train our classifiers at Kount. My recipe also uses 1 tsp of salt. The mixture is 0.7% salt by volume. That’s about the chargeback rate many of our customers experience before they come to Kount. If we pretend the grains are the same size, we can imagine the grains of flour are good transactions and the grains of salt are bad ones.
After the flour and salt are mixed together, stopping chargebacks is like sorting the grains back into a pile of flour and a pile of salt.
There’s about 150 grains of flour for every grain of salt, so our sorting assembly line needs to almost always put the grain in the flour pile. We can’t just say “this looks more like salt than flour, so I’ll put it in the salt pile”. If we do, our flour and salt piles will probably be close to the same size, and we know that’s wrong.
In fact, we need to be able to say “I’m 99.3% sure this grain is salt”. If we’re that sure, we’ll be right more than half the time. This skew in prior probability is one of the ways fraud is particularly hard. You can read more about that here.
If you actually want to separate salt from flour, I hope you won’t look one grain at a time. I take a scientist’s approach to my cooking, so I know that salt dissolves easily in water while the longer proteins and carbohydrates in flour form dough. You can separate salt from flour by adding a small amount of water, mixing it well, and draining it off. Then, boil off the water to recover the salt and sift or grind the flour to return it to a fine consistency*.
Now I’m going to stretch this analogy as thin as I roll my flatbread. Looking at a good transaction and a bad transaction side by side is a lot like comparing two little white grains of flour or salt. It is hard to tell the difference by looking at the visible characteristics. The better way is to find a mechanism that disproportionately affects one over the other, like water dissolving salt. This is how Kount’s fraud protection works.
If your business is running fraud control like an assembly line, we help identify the patterns you can’t see and focus your attention on the highest risk transactions. Kount’s device fingerprinting and machine learning algorithms provide reviewers with much more information than just the payment and customer information that’s easy to collect.
Rather than looking at each transaction by itself, Kount’s Persona Linking finds related transactions across our platform. Fraudsters rarely strike once, but finding the pattern requires knowing where to look, collecting all possible data, and linking it in real time.
These are just a few examples of how artificial intelligence helps fight fraud. At Kount, we recommend AI to complement human intelligence. People excel at guiding machines based on specific business objectives, while AI can quickly look at many hidden features of each transaction. Using the two techniques together is imperative. Read the eBook, Two Heads Are Better Than One: Artificial Intelligence + Human Insight, to learn more.
* Using water also separates simple sugar from flour. A fun science demonstration separates sugar from salt by adding alcohol, then pouring through a sieve to remove the salt. (Sugar is alcohol-soluble while salt is not.)