Tabular foundation models deliver top accuracy on evolving data streams
Rethinking Tabular Foundation Models On Data Streams
Machine LearningArtificial Intelligence
Summary
Data scientists often need to make predictions from data that continuously changes and arrives over time, called data streams. This paper studies special AI models known as tabular foundation models (TFMs) that learn from examples in their memory without changing their inner workings. The authors found these models predict better than traditional methods and quickly adapt to changes in the data, even when they get information late. However, the best accuracy comes with higher computing costs, suggesting improvements are needed to make these models run more efficiently in real-time scenarios.
What this means in practice
- •For real-time analytics teams: Improve prediction accuracy on continuously evolving tabular data by leveraging tabular foundation models that adapt via example context updates.
- •For data platform engineers: Optimize memory usage and model update scheduling for machine learning systems handling streaming tabular data with proven effective retention strategies.
Authors
Nilesh Verma, Daniel Nowak-Assis, Afonso Lourenço, Albert Bifet, Bernhard Pfahringer, Maroua Bahri, Nick Jin Sean Lim
Abstract
Tabular foundation models (TFMs) outperform established machine learning models on tabular benchmarks through in-context learning. Building on this success, interest is growing in applying them to data streams, where data arrive continuously and evolve over time. On a stream, a TFM adapts by updating its context rather than its parameters, so its accuracy and cost depend on which examples it keeps and how often it rebuilds its context. We therefore present a systematic study of TFMs on data streams, covering memory management, computational cost, and stream-specific challenges such as concept drift and delayed labels. We find that TFMs achieve the highest predictive performance and that simply retaining the most recent examples is as effective as existing memory management techniques. They also recover faster than streaming learners after drift and keep the highest accuracy under label delay. This accuracy, however, comes at a high serving cost, since a nearly unchanged context is re-encoded at every prediction. These results point to architectural efficiency as the way forward for in-context stream learning.