It’s a basic rule in data science: when precision and accuracy are required for fraud prevention, make sure you choose the algorithm appropriate to a bank’s data sets and needs. A complex algorithm does not mean better; a simple algorithm does not mean worse, writes Dr. Vivien Bonvin.
Not so long ago, a Tier 2 bank approached us about a proof-of-value (PoV) project. It had heard how we successfully use artificial intelligence and machine learning to identify fraudulent transactions precisely and accurately. It was keen for us to use one of our more complex algorithms in its project, believing that complex equated to more effective. Once we had examined the bank’s data set, however, we knew that using a complex algorithm would be a mistake. While it would be precise – it would not generate many false alerts – it would not be accurate as it wouldn’t detect all the fraud. Sometimes simple is better.
Fortunately, with our solution, it is easy to switch between different algorithms so we did what the bank wanted and set up the project with a complex algorithm but also prepared to run with the simple one.
A complex algorithm might use deep neural networks or graph analysis to draw on and analyze many data points. These data points act like markers scattered throughout an unknown landscape – for a bank they represent genuine and fraudulent transactions. When a new transaction is examined for fraud, it is compared against the historical data and given a risk score. But here’s the problem.
The bank didn’t have enough data on fraud for the complex algorithm to accurately fill in the unknown, background landscape. It had a lot of data about genuine transactions, but little on fraudulent ones – creating a significant labelled data imbalance. As a result, while the complex algorithm could identify frauds similar to those it had already come across, it was unable to accurately identify new fraud types. This meant that while the algorithm raised few false alerts (good in terms of it demanding little fraud investigation time), it still allowed fraud to take place (bad in terms of reputation and customer experience). It was precise, but not accurate.
This was exactly the outcome we had been expecting. Meanwhile, the simplicity of the alternative algorithm prevented it from learning the pattern of the fraudulent data too precisely. As a result, it learned to spot new frauds unseen before, picking up ones missed by the complex algorithm. It traded some precision for better accuracy, proving to the bank that in its case, simple was better.
You may also be interested in this article:
The key to choosing the right algorithm is knowing and understanding the banks’ data sets and business constraints. The algorithm’s job is to quantify the risk of fraud accurately and precisely for every transaction. If among a bank’s data there are only a few known frauds – which was the case with our Tier 2 bank – there is a strong labelled data imbalance. This means that the algorithm is unable to rely on the data to precisely paint an accurate picture of the unknown fraud landscape – because it is too accurate, it will miss frauds that lack precedents.
When a financial institution has a significant labelled data imbalance, a simple algorithm can be trained over a relatively short period of time. On the other hand, if more frauds lie among the institution’s data, the algorithm has more relevant information upon which it can draw and learn. It can accurately and precisely assess risk and by doing so can flag up suspicious transactions.
That being said, our 3D AI adaptive feedback pillar allows the system to build up a larger set of genuine and fraudulent transactions over time. This allows our solution to smoothly move at the appropriate time from simple AI to a more advanced one, better adapted to the new landscape.
Our parallel projects proved two things: they showed the bank that NetGuardians’ fraud-mitigation software is effective at spotting and stopping not just known frauds but new fraud types; and they showed we understood how to interpret its data set and choose the right algorithm so it got the most efficient results.
The benefits are the bank doesn’t waste resources funding large investigation teams that chase false alerts; and it doesn’t bother customers unnecessarily about genuine transactions. Our experience means we will ensure that our customers use the right algorithms. And the icing on the cake is that our software delivers a clear explanation as to what is suspicious about any suspended transaction, speeding investigation. In this respect, it already complies with existing and draft regulations concerning explainable AI and transparency.
So if your bank wants a truly efficient fraud-mitigation regime, one that can accurately spot fraud – both known and new types – with minimal false alerts, it should talk to us. Not only do we have the right algorithms, but we know which ones to use and when. It’s simple when you know what you’re doing.