David Lock, Senior AI Product Manager

Normal for Norfolk is an expression English doctors once used to denote the outliers to the rest of the population. As a proud resident of Norfolk for nearly 20-years, I feel I don’t qualify for that tag, but that’s not for me to say.

This line of thinking does, though, bring me on to how we classify fraud in a machine learning environment.

The conventional wisdom is to use global models; be they global models based on large amounts of consortium data or global models based on your own data. They identify trends based on global experience and will determine things like whether an internet transaction with a low value followed by a high value one is potentially fraudulent.

Now every fraud manager has their own pedigree of villain to battle, much like any other superhero. Unfortunately, this can soon turn into profiling in potentially an unethical way. So, I’m going to create a semi-fictional environment, where the names and places have been changed to protect the innocent – and hopefully the chances of offending anyone is very very low.

So, to make this point, I used the example of the island of Tristan da Cunha in the middle of the Atlantic Ocean. With a population of around 250 people and no bank issuing cards on the island, I’d say I’m reasonably safe on not insulting anyone – in the unlikely event you are from Tristan da Cunha in this fictional example, I apologise.

So in this example, Tristan da Cunha cards are used in petrol stations in the UK to convert stolen cards numbers stolen from the highly insecure banks in Tristan da Cunha at regular intervals - this is the fraud pattern. Not hard to identify and simple to build a rule to track the pattern, or train machine learning to pick it up.

However, the scenario fraud managers face is that there is a petrol station next to the Tristan Embassy in London. The ambassador and his staff legitimately use it to fill up their large gas-guzzling SUVs on a regular basis. Obviously, these are not stolen cards, indeed they are VIPs.

Now, this is not an outlier, yet it is likely to be picked up by any global fraud detection model. Normally, the reliance is on expensive humans to detect these ‘outliers’ and detect that they are false positives.

So, what is actually needed is two models. The first to track the global model, and the second to determine if this is normal for a particular customer i.e. as it is for the ambassador and his wife, by profiling the transaction against their account.

From a fraud perspective, can we treat this as ‘normal for Norfolk’, or a false positive. At the very least, reduce its priority when it comes to investigations. This is cheaper and easier than relying on the global or consortium model.

This is just another way Renovite is using AI and cloud-native technology to improve transaction fraud management.

Again, I stress the scenario is completely fictional and not a reflection on the good people of Tristan da Cunha. Indeed, as I did my research for this I found they were recently hit by a storm which damaged the island’s infrastructure, so if you’d like to donate to the appeal the website, please click here.