By Jimmy Hennessy, Director of Data Science, ACI Worldwide
In recent years, merchants have faced continuous change. The rapid acceleration of digital consumer purchasing, accelerated by COVID-19, has forced rapid digital transformation for merchants. This has bought many benefits, but has also created many new opportunities for online fraud. One significant growth area with more people ordering products to their homes has been the incidence of fraud using the brands of regularly used delivery and courier companies. When detecting and analysing online fraud trends, the rapidly evolving nature of fraud ‘vectors’ has shown that existing technologies, namely traditional machine learning, aren’t able to keep up.
Between 2021 and 2025, merchants are predicted to lose $206 billion to online payment fraud. With retailers expected to spend approximately $9.6 billion annually on fraud detection and prevention. However, technology is developing to combat increasingly advanced fraudsters, promising to free up resources and lower operational costs. Incremental learning is the new machine learning capability, uniquely suited to combat this problem and adapt on the ever-evolving fraud frontline.
Traditional machine learning techniques have expired
Although traditional machine learning methods form an important part of a retailer’s multi-layer defence strategy, it can struggle to cope with rapidly changing situations. The more dynamic the market, the harder it is for a human analyst to interpret the data without help.
After a few months of building a machine learning model, they are then deployed with additional months of training on massive data sets that represent that moment in time. However, as new behaviours and trends emerge these models can quickly deteriorate. There is an opportunity to retrain the models but that can take weeks and with continuous change traditional machine learning will inevitably have to play endless catch up.
Combatting fraud in real-time
Today, learning only from historical data, has limited value. Whilst a model that is fit for purpose to tackle fraud can be augmented and refreshed live, incremental learning has been shown to outperform traditional machine learning models in accuracy by 10%.
It operates in the present, avoids needing to constantly be updated and uses real-time data to improve performance over time. It can reduce fraud losses by as much as 75% and help improve fraud detection by up to 85%.
By making ongoing small adjustments on a regular basis, incremental learning can adapt to new behaviours in real-time. Not only does this maintain consistent performance, but it also improves it, allowing incremental learning to remain capably reactive to new fraud intelligence with minimal intervention.
The quiet but giant leap
The beauty of incremental learning is that it is effective and gets the job done. With the added benefit of not needing complex rebuilding, retraining or model revisions. Much like humans, new knowledge and behaviours can be added to existing knowledge and maintain a consistent high level of coverage against fraud risks without disruption. This allows merchants to allocate their time and internal resources in other places that need it more.
Incremental learning provides fewer disruptions and errors whilst allowing merchants to benefit from greater protection and reduce fraud. This is the bold and crucial step in the right direction that is needed and it embodies what automated tech should be – silent and effortless. With the constant and increasing pace of evolution in fraud trends, incremental learning is the clear next major weapon to deploy in the fight against fraudsters.