Machine learning is nothing new, but during the pandemic, fraudulent activity hit an all-time high, and its popularity soared. Now, it is the primary tool used for mitigating fraud, and companies like ACI Worldwide are leading the charge in developing algorithms and models to serve each and every one of their customers.
To further discuss the benefits of machine learning and how it can better serve institutions looking to improve their fraud prevention technologies, PaymentsJournal sat down with Patricia Rojas, Senior Manager Data Scientist at ACI Worldwide, and Tim Sloane, VP of Payments Innovation at Mercator Advisory Group.
Machine learning is essential for fraud prevention
It is now clear that machine learning is a valuable tool for fraud prevention, and most experts would agree that it has become essential for mitigating cybercrime. On a high level, detecting fraud is about learning the difference between normal spending behaviors and unusual, fraudulent purchases. With machine learning, the technology can analyze all available data and educate itself on the difference between an honest transaction and a fraudulent one.
“These type[s] of models, when they’re properly trained and get the feel for one specific merchant or one specific sector, they can help increase the fraud detection accuracy in your overall strategy by as much as 40 to 50%,” claimed Rojas. She warns, however, that merchants and PSPs need to understand the specifics when implementing machine learning algorithms, because there are many different techniques and levels of sophistication. It is also important to note that these algorithms are limited by the amount and quality of data within the institution.
There are many different applications of machine learning, and its evolution shows no signs of slowing down. With fraud also occurring in a fast-paced environment, a company like ACI is necessary to correctly apply machine learning to fraud prevention.
Machine learning trumps other fraud prevention tools
Identifying fraudulent behavior can be a complex and time-consuming task, especially for institutions with an abundance of data. In such cases, machine learning models are ideal because of their efficiency and ability to analyze massive amounts of data to identify trends. Not only are they more precise, but they are also exponentially quicker.
“This is very important because different behaviors change very quickly,” said Rojas. “You need to be able to stay on top of that and to adapt your strategy to be able to capture those new fraudulent behaviors.” Overall, machine learning is a tool that can help its users improve their fraud prevention strategy and minimize the ‘false positive’ transactions. It can even assist in reducing friction for customers at checkout.
Tim Sloane breaks down the process to offer a better understanding: “You have data at the merchant location. You have [data] about the account individual, their behavior. You have data coming from the network. You have data at the acquirer. And you have data that, if you’re lucky, you can get from the issuer to be able to tie it all together. [Machine learning can] pull those signals together and learn more than you possibly could any other way.”
All machine learning is not created equal
There are a multitude of machine learning models, as well as many different algorithms that can be used, case-by-case. While tree-based algorithms tend to work best for fraud detection, different use cases might require a different approach. It is crucial to first use the right model, and then to optimize that model for a specific merchant or sector. When models are trained with specificity, they are more effective because they take into account the nuances of customer behavior, fraud trends, and spending patterns.
“At ACI, one of the things we do to improve the performance of our model is to leverage the power of the consortiums by building strong models for our merchants,” explained Rojas. “We do this by identifying similar merchants and then combining all that information to train our models.” This gives ACI a larger set of data to provide information for the model they are building, which then enhances the ability to correctly identify fraudulent behaviors and make more accurate predictions for future transactions. The performance result is significantly increased.
ACI is also developing new incremental learning models. This type of models differs from static models mainly in how they are built and maintained over time. With a static machine learning model, a historical set of data is used to build the model and, over time, that model becomes less efficient as fraudulent behavior evolves and model will need to be retrained to learn the new fraudulent behaviors to be able to make an accurate prediction. With the new learning model, the technology is able to think for itself and adapt to new behaviors without having to relearn everything it already knows which not only makes the training phase more efficient but also a more accurate prediction using more recent and relevant data to prevent future fraudulent transactions.
“These types of models will perform better in production for longer, and it’s reduced the number of retraining[s] that we need to do…it’s a smooth process for the customers,” concluded Rojas.
Mitigating the limitations of machine learning
“Sometimes a merchant has a special offer going out,” explained Sloane. “And that special offer is going to generate new types of traffic that needs to be coordinated with the machine learning tools and the people who are operating them to make sure that that special offer is done in a safe fashion and doesn’t throw off the models.”
Seasonality can significantly impact the performance of models. High sales peak seasons and the launch of a new product can both impact the reading of normal and abnormal behavior.
Everybody has different goals, and merchants are no exception. While one merchant may be looking to reduce false positives, another might want to maximize the fraud detection rate. ACI engages with merchants at a very early stage to understand their goals and offer a multi-layered technology to optimize the overall fraud strategy in a way that best caters to the needs of the merchants. It takes into account seasonality, peak sales seasons, new product launches and other special circumstances to ensure the merchant is protected against fraud and revenue is not impacted.
Part of ACI multi-layered technology is the Rule Intelligence process, which is a machine learning model that generates human readable rules in an automated way that is tailored to merchant-specific needs. The rules generated by this process are a small set of high performing rules, which reduces the false positives, reduces the time needed to create a fraud strategy, and can be refreshed to adapt to changes in behaviors.