Mastercard takes wraps off new GenAI model

Unlike the popular large language models trained on huge amounts of unstructured data, Mastercard, working with Nvidia and Databricks, has built a large tabular model, or LTM, which is trained on structured data, such as large-scale tables or datasets.

The LTM is trained on billions of anonymised transactions, with Mastercard planning to ramp this up to include hundreds of billions of payments transactions, as well as additional types of datasets, including merchant location data, fraud data, authorisation data, chargeback data and loyalty programme data.

The model is already paying off in the realm of cybersecurity, outperforming standard industry machine learning techniques, says Mastercard distinguished engineer Steve Flinter.

“For instance, very expensive but very infrequent purchases — such as when someone buys a wedding ring — tend to trigger current models today and cause a lot of false positives. In our experiments, our foundation model can better identify these legitimate transactions, with the model able to learn from relatively weak signals in the data,” writes Flinter in a blog.

Meanwhile, Flinter explains that Mastercard currently needs to build, train and maintain thousands of AI models to run its network, each for different markets, use cases or customers. The new LTM could become flexible enough to help cut down significantly on having to maintain so many different models.

Now, the firm is developing APIs and toolkits to give teams across Mastercard access to the new foundation model so they can build new applications on top of it.

Writes Flinter: “We plan to use this new foundation model — not to build a chatbot — but as an insights engine that will make many of our tools and services even better, from cyber defenses to loyalty programmes to small business tools.”