For several years now, banking institutions have been relying on machine learning methods for card fraud detection.
In such cases, algorithms undergo training to help them learn how to identify the behavioral patterns of fraudulent transactions so that they can block or thwart subsequent suspicious activities.
The drawback of using these machine learning techniques in detecting card fraud is the resulting false positives or rather the number of genuine transactions that they incorrectly qualify as fraudulent.
For instance, when customers use their cards far away from their normal location or pay an extraordinary sum of money, banks usually block their cards since the system mistakenly identifies such unusual actions as “suspicious”.
The reason behind this is that such methods, even though successful in other areas, are susceptible to several drawbacks when used in detecting card fraud.
These machine learning techniques derive more false positives as their ability to identify fraudulent transactions increases.
Enhancing user experience in such situations without interfering with the existing security levels marks one of the permanent obstacles affecting BBVA.
Motivated by this goal, the Corporate Security and Engineering Risk team of BBVA started partnering with the Computer Science and Artificial Intelligence Lab of MIT back in 2015 in a move to search for new techniques of minimizing the number of false positives.
The results of this partnership were published in the form of a joint study that MIT presented at the recently held European Conference for Machine Learning.
The joint study was also included in one of the articles posted on MIT’s news portal.
Among the issues spotted was the shortage of mathematical variables or otherwise known as ‘’features”, which could be derived from the data used in training machine learning systems.
Another obstacle identified was the lack of enough useful examples for training models or an equivalent of what researchers refer to as “looking for needles in a haystack.”
Algorithms ought to be able to spot irregularities that occur rarely.
The BBVA and MIT research team leveraged revolutionary automated engineering approaches in trying to make this situation better.
They utilized the Deep Feature Synthesis technique, a model created by two MIT researchers including James Max Kanter and Kalyan Veeramachaneni.
The model allows detailed features to be derived automatically from datasets.
This groundbreaking method enabled the research team to derive more than 200 new extra features from each card transaction.
These features are useful when it comes to describing card transaction behavior in detail as well as improving the outcomes generated from the detection engines.
Some of the new features utilized in refining true positive outcomes show, for instance, the average amount of time spent in various stores or whether the card user was physically present during the transaction.
Up to now, tests are continuing to produce promising outcomes.
The number of false positives has also declined considerably while maintaining the same fraud detection level in comparison to other systems and products on the market today.
Furthermore, applying the Deep Feature Synthesis technique to a dataset made up of 1.8 million BBVA card transactions helped in reducing the number of false positives by 54 percent over that of conventional models.
“Taking as a starting point this model that was designed and prepared by the MIT team, as internal research in the bank, we have synthesized new features using additional card and business data that were not available at the time of the original research,” said Carlos Capmany, who manages the project for BBVA’s CS&ER area.
“As a result, we have managed to consistently reproduce the already impressive results of false positive reduction obtained by the MIT research team, and this with a minimum increase in the overall computational cost of the system,” added Capmany.
The Advantages of the Model
Applying the outcomes of this study to the bank’s fraud detection systems has shown their ability to boost customer satisfaction, particularly in card transactions without interfering with the existing security standards.
What’s more, BBVA’s CS&ER area is applying this method in other areas of research, specifically on machine learning (ML) applied to cybersecurity with similar obstacles such as disparate input data and scarce variables for training the machine learning system.
“This automated feature synthesis technique and the overall knowledge MIT contributed to this project have shown us a new way of refocusing research in other challenges in which we initially have a reduced set of features. For example, we are obtaining equally promising results in the detection of anomalous behavior in internal network traffic or in market operations, just to mention two,” added Sergio Iglesias, another member of the CS&ER team, who was involved in the project.