NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

Prediction on business data usually means building a separate model for each question. Someone flattens linked tables of users, orders and products into one row per customer by hand-writing features such as average order value over 30 days, then trains a model like XGBoost for that one task. NVIDIA's Kumo Relational is meant to remove both steps. It is pre-trained once, entirely on synthetic relational databases. For a new question, written in a declarative Predictive Query Language, it pulls up to 10,000 historical examples from your own database as context and predicts in a single forward pass with no task-specific training. For time-based tasks it rebuilds earlier states of the database, so each example only sees data from before the moment it is labelled. NVIDIA says this prevents leakage from the future.

The model keeps the database's structure instead of flattening it. A small network first relates values within each table, across columns and across rows. A larger one then passes information along the primary-key and foreign-key links between tables, which connects a user's profile to their transactions, and compares the entity being predicted with the labelled context examples. Because this second stage works on one embedding per row rather than on every cell, NVIDIA says it avoids quadratic cost as context grows.

The headline results are NVIDIA's own. Across 41 tasks on 15 databases from RelBench, SAP SALT and 4DBInfer, none seen during pre-training, NVIDIA reports that Kumo Relational beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines. On RelBench V1 classification it is nearly 4 AUROC points ahead of boosted trees on engineered features. On SAP SALT it beats AutoGluon ensembles by 8% in mean reciprocal rank, rising to 16% after fine-tuning. NVIDIA also states the limits. The model works best when the data has meaningful relationships and changes over time. With very large training sets, task-specific supervised models may still edge it out, since the benchmarks used at most 10,000 context examples, and fine-tuning is offered for that case. Useful signals also have to be reachable within a limited number of table hops.

The code ships under Apache 2.0 and the weights under the OpenMDW license. They are available in NVIDIA's open-source structured-data-models library, the same one behind its sister model Kumo Tabular, or as an NVIDIA NIM service.