There are, however, challenges around the development and training of machine learning models that merchants need to understand:
- Fraud represents a small fraction of all daily transactions
- The distribution or pattern of fraudulent and good transactions evolves over time, due to seasonality, changes in customer behavior and new fraud attack strategies
- Without timely checks, the true nature of most transactions is usually only revealed a few days – or even weeks – after they actually take place
What this means is that the training of machine learning models requires a very large amount of historical transaction data. And, as good and bad behaviors change, the performance of most machine learning models will degrade. They will need to be retrained on large amounts of data.
Enter incremental learning…
We recently announced the launch of incremental learning capabilities to address this challenge. This exciting new approach to modeling is designed to ensure that machine learning models perform to the highest quality for longer than current traditional models.
Incremental learning models are able to “think for themselves” and make small adjustments on an ongoing basis, ensuring they remain relevant even as fraudsters and genuine consumers change their behaviors. This represents a significant advancement over traditional machine learning models that need to be retrained as fraud patterns change.
Learning the new tunes
To understand how incremental learning works, think of it like playing the piano. As we get better at playing the piano, we learn new tunes. We build our repertoire over time – and we don’t need to re-learn all the old tunes when we learn a new one.
This is how an incremental learning model works – and this is the point: a traditional machine learning model would have to re-learn all the tunes it previously knew to be able to learn the new one.
Incremental learning algorithms allow a model to adapt to new behaviors without needing to re-learn everything it already knows. New data is fed to the algorithm every 24 hours and it spots new behaviors as they happen. If the change in behavior is enough to cause a degradation in performance, the model will adapt to what it has recently learned. This ensures that model performance quality lasts for longer periods of time, without degradation. In tests, our incremental learning models maintained their performance over 13 months, while others started to degrade after just three months.
What this means for merchants
Incremental learning reduces the need to retrain over large data sets, and the need to continuously re-deploy to production, because the changing data is incorporated on an ongoing basis. That reduces the demand on scarce merchant resources.
Most importantly though, incremental learning performs better than traditional methods. By using the most up-to-date fraud intelligence and customer behavior data, and continually adjusting in response to that data, models can protect merchants more effectively, over longer periods of time.
Given the continuous evolution of customer preferences, shopping trends and fraudster behavior, assessing transactions against the latest data has never been more important. Fraudsters will continue to adapt their tactics and techniques in an effort to evade detection – incremental learning will be instrumental in keeping up with fraudsters and stopping them in their tracks.
Watch our on-demand webinar: Find a New Ally in the Fight Against Fraud with Machine Learning