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

NVIDIA Kumo Relational is a groundbreaking pre-trained relational foundation model that significantly improves predictive outcomes on enterprise databases without the need for training, feature engineering, or flattening tables. This innovative model outperforms traditional supervised machine learning methods on 41 tasks across 15 databases, even surpassing expert-engineered gradient-boosted tree pipelines.

On the challenging SAP SALT enterprise benchmark, Kumo Relational achieves an 8% improvement in mean reciprocal rank (MRR) over AutoGluon ensembles.

The model is capable of handling databases with 500 billion or more rows and benefits both human users and agents by leveraging NVIDIA's open-source, GPU-native structured-data-models library or as an NVIDIA NIM. Similar to the transition from specialized models for language and vision tasks, the advent of structured data foundation models like Kumo Relational is reshaping the landscape for business decisions powered by relational databases.

Kumo Relational eliminates the time-consuming process of feature engineering and task-specific training, which typically take weeks to complete. Instead, it offers superior accuracy out of the box, often outperforming established pipelines such as XGBoost. Notably, on RelBench V1 classification tasks, Kumo Relational surpasses gradient-boosted trees by nearly 4 AUROC points on average. Furthermore, on SAP SALT, it outperforms XGBoost by 7 MRR points without fine-tuning and by an impressive 13 MRR points with fine-tuning.

The simplicity and agility provided by Kumo Relational are transformative, as one pre-trained model can replace multiple specialized pipelines, allowing ML teams to tackle a broader range of use cases with ease. Additionally, the model is designed to be called by agents, with its Predictive Query Language (PQL) offering a straightforward interface for asking predictive questions.

Kumo Relational connects to a database, specifies what to predict, and delivers the results in a single forward pass, without requiring any training.