NVIDIA Kumo Relational: churn and fraud predictions straight from your database tables, with no training run

NVIDIA introduced Kumo Relational on 21 October, a pre-trained model that makes predictions directly on a relational database. It is the multi-table sibling of Kumo Tabular, which works on one flat table.

How it works

Most prediction models on business data need a flattened table of hand-built features, such as “average order value in the last 30 days”. Kumo Relational instead reads the tables, foreign keys and timestamps as they are. You state what to predict in its Predictive Query Language, and it gathers up to 10,000 past examples as context and predicts in one forward pass. There is no task-specific training, though you can fine-tune it.

For time-based tasks it replays the database as it looked before a cut-off time, which NVIDIA says prevents leakage from the future. The model was pre-trained only on synthetic relational databases.

What NVIDIA reports

All of these are NVIDIA’s and the paper authors’ results, on 41 benchmarks across 15 databases that were not in pre-training:

• On RelBench V1 classification, nearly 4 AUROC points better on average than gradient-boosted trees on expert-built features.
• On SAP SALT, 7 MRR points better than XGBoost out of the box, and 13 after fine-tuning.

NVIDIA also lists limits. It works best where relationships and time matter, such as user behavior. With very large training sets, a dedicated supervised model may still edge it out. And the useful signal has to sit within a few table hops, so NVIDIA suggests starting with a narrow query over a few tables.

Running it

The weights are on Hugging Face under the OpenMDW-1.1 license, ungated; the structured-data-models library is Apache 2.0 and downloads the weights on first use in a GPU notebook. You can also run it as an NVIDIA NIM service, and NVIDIA publishes Databricks examples. Data can stay in memory or come through database and warehouse connectors; NVIDIA says its graph engine handles more than 500 billion rows.

The research paper dates from April. The 21 October post is the public launch of the model and library.