Mastercard has built a large tabular model — an LTM — trained on billions of card transactions to detect fraud and improve decision-making across its payments infrastructure.
Unlike large language models, which process unstructured text, the LTM examines relationships between fields in structured, multi-dimensional data tables. That makes it closer to classical machine learning than generative AI.
The training data includes payment events, merchant location, authorisation flows, fraud incidents, chargebacks, and loyalty activity. According to the announcement, personal identifiers were stripped before training began. The model works from behavioural patterns, not individual profiles.
Mastercard says removing personal data reduces privacy risks common in financial AI applications, and that the volume of anonymised behavioural data compensates for any loss of individual-level signal. The company describes the LTM as an ‘insights engine’ designed to sit inside existing products rather than replace them.
Where It’s Being Used
Cybersecurity is the first live deployment. Traditional fraud detection systems at the firm require human input upfront to define what counts as suspicious — unusual transaction frequency, purchases made in geographically distant locations within minutes, that kind of trigger. The LTM learns which patterns are anomalous without relying on those predefined rules.
Early results show the model outperforms conventional methods in specific cases. The company cites high-value, low-frequency purchases as one example: legacy models tend to flag these as anomalies, but the LTM can distinguish legitimate transactions from genuinely suspicious ones with greater accuracy.
The firm is not replacing existing systems. Hybrid deployments that combine established fraud detection procedures with the new model are planned, a reflection of the regulatory environment Mastercard operates under. The company acknowledges no single model performs well across all scenarios.
Infrastructure and Expansion
Technical infrastructure comes from Nvidia, which provides the computing platform, and Databricks, which handles data engineering and model development.
The intention is to expand training data from billions of transactions to hundreds of billions. Beyond fraud, the company says the model can be applied to loyalty programme monitoring, portfolio management, and internal analytics — any area with large volumes of structured data. Plans include API access and SDKs so internal teams can build new applications on top of the foundation model.
Running many task-specific models is expensive. Each requires separate training, validation, and ongoing monitoring. A single foundation model fine-tuned for different tasks could reduce that overhead significantly, though the company’s blog post acknowledges the corresponding risk: a failure in a widely deployed model carries system-wide consequences.
The announcement also emphasises privacy, model explainability, and auditability as core obligations attached to the LTM’s operation.
This article is a curated summary based on third-party sources. Source: Read the original article
