Mastercard has developed a large tabular model (an LTM as opposed to an LLM) that’s trained on transaction data rather than text or images to help it address security and authenticity issues in digital payments.
The company has trained a foundation model on billions of card transactions, with the intention of expanding to hundreds of billions in time. The datasets include payment events and associated data such as merchant location, authorisation flows, fraud incidents, chargebacks, and loyalty activity. Mastercard says personal identifiers are removed before the training began, and that the model parses behavioural patterns rather than concern itself with individual identities.
By excluding personal data, the technology reduces privacy risks that may affect other forms of AI in financial services sector. The scale and richness of the data allow the model to infer patterns that are commercially valuable – the company said in a recent blog post – despite the lack of per-user information. Although anonymisation removes signals that could be argued as being useful in the area of risk assessment, Mastercard asserts that using sufficiently large volumes of behavioural data compensates for any loss of rich data.
What is an LTM (large tabular model)?
LTM architecture differs from that of large language models, which are trained on unstructured inputs and work by predicting the next token (typically but inaccurately described as a word) in a sequence. Mastercard’s LTM examines relationships between fields in multi-dimensional data tables, making a definition of the technology closer to that of pure machine learning rather than artificial intelligence.
The large tabular model learns from raw inputs exactly which relationships are predictable, so it can identify anomalous patterns not captured by predefined rules.
The company describes the LTM as an ‘insights engine’ that can be used in existing products, augmenting existing workflows. The operational risk of a model that interacts with customers (often an LLM) differs from that of one that’s part of internal decision-making.
Technical infrastructure for the LTM comes from Nvidia and Databricks, with the former providing the computing platform and Databricks handling data engineering and model development.
Where will we see an LTM in operation?
Cybersecurity at Mastercard is the first area to see active deployment of the tech. Like many institutions, Mastercard operates several fraud detection systems examining transaction data. These require human input at their outset – and ongoing attenuation – to define what constitutes as suspicious behaviour. These might include sudden increases in transaction frequency, or users making purchases in different parts of the world in a small space of time.
Early results indicate improved performance on conventional techniques in specific cases, the company says. It cites the example of high-value, low-frequency purchases which can be flagged as anomalies using traditional models, but the new model appears to be able to distinguish legitimate events more accurately than its counterparts.
The company plans to deploy hybrid systems that combine established procedures with the new model, a degree of caution that reflects the regulatory levels it operates under. It acknowledges that no single model is likely to perform well in all scenarios, so the LTM will take its place among the tools in this sphere.
It’s claimed the model can scan activity on loyalty programmes, be used in portfolio management, and for internal analytics, areas where there are large volumes of structured data. In current operations, companies often deploy many models adapted to each task, but this can involve multiples of training costs and validation and monitoring efforts. A single foundation model that can be fine-tuned for different tasks may simplify processes and keep costs down.
Risk and future plans
There’s a risk to the multi-function LTM approach, of course: A failure in a widely-deployed model could have system-wide consequences, which goes some way to explain Mastercard’s strategy of applying its technology alongside existing detection systems – at least, for the present.
Mastercard hopes to increase the scale of the data used on the model and its overall sophistication. It’s also planning on API access and SDKs to let internal teams build new applications.
The blog post emphasises the data responsibilities the LTM holds, mentioning privacy and transparency, model explainability, and auditability. Regulatory scrutiny of any system that influences credit decisions or fraud outcomes is to be expected in addition to any data practices involved in the LTM’s operation.
Highly structured data, as opposed to text or images, lies at the core of the LTM. Large tabular models may be the start of a new generation of AI systems in core banking and payments infrastructure. Evidence to date remains limited to vendor reports, so any performance claims should not necessarily be regarded as conclusive.
Robustness under adversarial conditions, long-term post-training costs, and regulatory acceptance are all issues on which tabular models may founder or thrive. These factors will determine the pace and extent of adoption, but it’s the area of the table where Mastercard is placing some of its bets at present.
(Image source: “Oversight” by United States Marine Corps Official Page is licensed under CC BY-NC 2.0.)
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post Mastercard keeps tabs on fraud with new foundation model appeared first on AI News.
