Incremental Learning: The Real-Time Hero Taking on Fraud

In the fast-moving world of merchant fraud, we can no longer rely on the past to predict the future. The speed and scale of change is such that traditional machine learning (ML) methods, which analyze historical fraud trends, can’t keep up.  Luckily, there’s a new tech in town – incremental learning. Amongst other benefits, it halts the model degradation associated with many current backward-looking algorithms and makes life much more difficult for today’s fraudsters.

Incremental learning (IL) aims to remove the burden of monitoring fraud by constantly analyzing and filtering data to identify potential threat signals. It doesn’t just rely on historic data. 

IL operates in the present, using real-time data to learn and adapt. Not only does this help it to improve fraud detection – by up to 85 percent – it can also reduce fraud losses by as much as 75 percent.  With merchants predicted to lose $206 billion to online payment fraud between 2021 and 2025, this level of saving could have a significant impact on their financial performance. 

Retailers are also expected to spend approximately $9.6 billion annually on fraud detection and prevention. By improving performance, IL promises to lower operational costs and free up resources for other enterprise or customer-facing priorities. 

Why traditional ML techniques are no longer enough

Though it forms an important part of a retailer’s multi-layer defense strategy, traditional ML methods can struggle to cope with continuous and accelerated change – for example, what we’ve witnessed during the COVID-19 pandemic.   

Typically, a model is combined with rules-based engines to help spot transaction and fraud patterns.

They are built and then deployed after months of training on massive data sets. But these models can degrade quickly, as new trends and behaviors emerge.  While these models can be retrained, it can take weeks.  Near constant change means that traditional implementations of ML will have to play catch up once new behaviors are seen.   

Incremental learning has been shown to outperform traditional ML models in accuracy by 10 percent

The secret of IL’s success lies in its ability to make small incremental changes and implement them in near real time.  It combines learning from past behaviors with live data to make changes and update its own models. Like a human, IL learns from experiences as it goes along.

By making small adjustments on an ongoing basis, it can adapt as soon as new behaviors are observed. Not only does this maintain performance, but it also actually improves it, allowing IL to keep pace with new fraud intelligence with minimal intervention. 

And – going back to the issue of feedback – it also delivers more explainable decisions to help ensure more ethical use of AI.

A big step forward in invisible support

For the most part, incremental learning simply gets on and does the job. There are no additional skills required, no complex rebuilding, retraining and rerelease of models. Models are augmented and refreshed “on the fly” based on the streaming data. They maintain a consistently high level of coverage against fraud risks without effort or disruption, which helps merchants make best use of their internal resources and removes the need for additional skillsets. 

Why is this so important for merchants? There’s a growing skills gap, not just in payments but across the IT sector as post-pandemic businesses accelerate digitalization. Around three-quarters of IT decision-makers worldwide claim to be facing critical skills gaps across tech departments.

In fact, you could say that IL does what automated tech does best – it’s effortless and invisible. Merchants can enjoy less fraud, as well as fewer disruptions, chargebacks and false positives. Automation offers greater protection for less effort – a smart reason to build incremental learning into an omni-channel fraud prevention strategy.