Any innovation that helps merchants outfox fraudsters can't come soon enough.
An innovative, industry-first approach to machine learning, incremental learning represents a step change in fraud prevention. It identifies patterns earlier and more accurately, automating decisions and actions to keep merchants better protected than ever.
Why do we need a new approach to fighting fraud?
Gone are the days of linear, clearly defined sales journeys. Today's buying behaviors are more varied, diverse and fragmented. This makes it difficult to identify definitive fraud patterns until they are large and consistent enough to detect - by which time, the damage to revenue and reputation may already be done. This is made even more challenging by the following:
What, where, when and how a shopper, gamer or traveler buys one day, may be completely different from the next. More choice means there's more friendly fraud and sensitive information in myriad new pipelines - from one-click marketplaces and social platforms to digital wallets and mobile and gaming apps.
Online and in-store, there are an increasing number of mobile, digital and hybrid options - from BOPIS (buy online, pick-up in-store) and self-serve, to contactless, QR code and mobile payments. This is exacerbated by flexible opening times, extended seasonal hours and recent lockdowns, which contribute to making long-term transaction and trading patterns more volatile and harder to read.
Synthetic fraud trends
Fraud tactics are becoming more sophisticated. According to the recent ACI eCommerce Fraud Index, synthetic fraud and bots are on the rise, creating even faster attack vectors. As soon as they spot a vulnerability, connected fraudsters can ramp up attacks on multiple brands and across sectors before the trend has been flagged and anti-fraud mechanisms have been triggered.
In dynamic omni-channel environments, even advanced machine learning may struggle to keep up
To identify patterns and create models, machine learning tools need a long-term bank of historic information to work from, often more than 12 months of transactional data. In the online world, they also have to contend with speech and image recognition and learn to identify new keywords and images.
And here's the issue. To remain accurate, fraud models need to be retrained over time as they are exposed to new items. But fraud can take weeks to be reported and information used to train them is often inaccurate, incomplete or incorrectly labelled. This makes it harder to classify between fraud versus genuine transactions, leading to skewed models with reduced efficiency.
As a result, anti-fraud rules and risk thresholds may be set too low, resulting in increased fraud rates. Alternatively, they may be set too high, creating friction for the user, stifling conversion and leading to false declines.
Incremental learning helps stay ahead of the threat curve and deliver a great experience
What makes incremental learning models different, is that they are able to "think for themselves" making small adjustments, on an ongoing basis, to remain relevant even as fraudsters and genuine consumers change their behaviors.
In addition, traditional machine learning models need to be retrained, recreated and then redeployed every time fraud patterns change. This is because every new version forgets everything previously leaned. It's a bit like having to relearn the alphabet every time you learn a new word.
However, with incremental learning nothing is lost. New trends can be added automatically, while models retain all their previous knowledge. It is also fully compatible for use within multi-layer anti-fraud environments. After each model refresh, the score threshold used to alert fraudulent or suspicious transactions is also retained and remains similar and stable.
When it comes to tackling friendly fraud, false declines and chargebacks, there's much to get excited about
Here are six advantages that incremental learning offers over traditional machine learning methods:
Wes noted that after 3 months, incremental learning models outperformed traditional models by more than 5 percent
It can be used across all sectors and channels, as well as purchasing behaviors and non-numeric data
Constant learning from recent data responds to "new normal" as they arise
Accuracy is not impacted by over-reliance on historic 12-month data trends
Identify and fight new fraud threats before they have time to mature
Reduce response timeframes to significantly reduce retailers' losses
What does it mean for the battle against fraud?
Ultimately, incremental learning helps merchants overcome current challenges around the development and training of machine learning models, including having to work with small amounts of data, constantly evolving distribution patterns and long-lead modelling times.
Instead, it makes it possible to be more accurate with less information, so that a multi-layered fraud system can work smarter and faster than ever before. More importantly, it helps fight payments fraud before a small pattern becomes a widescale trend, which could threaten both revenue and reputation.
The patented incremental learning approach to fraud preventionis exclusive to ACI Worldwide