NVIDIA Kumo Relational: A Game Changer for Predictive Modeling in Relational Databases

NVIDIA has introduced Kumo Relational, a pre-trained relational foundation model designed to make predictions directly on enterprise databases. This model can gather up to 10,000 historical examples as context to predict outcomes such as churn, demand, or fraud in a single forward pass, eliminating the need for task-specific training and feature engineering.

Kumo Relational is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks. It surpasses supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines across 41 tasks on 15 databases. Notably, on the SAP SALT enterprise benchmark, it exceeds AutoGluon ensembles by 8% in mean reciprocal rank (MRR). The model is capable of scaling to databases with over 500 billion rows and can be utilized by both humans and agents through NVIDIA’s open-source, GPU-native structured-data-models library or as an NVIDIA NIM.

The Shift to Structured Data Foundation Models

Foundation models have transformed the landscape of natural language processing (NLP) and computer vision. Previously, production NLP involved dedicated models for each task, requiring extensive pipelines for tokenization and feature engineering. Large language models (LLMs) have made these processes largely obsolete due to their superior accuracy, simplicity, and flexibility.

A similar transformation is occurring for structured data. Relational databases are crucial for business decision-making, yet predictive modeling on them typically demands extensive feature engineering and task-specific training. Structured data foundation models like Kumo Relational address these challenges by providing better accuracy, simplicity, and the ability for agents to predict outcomes.

How Kumo Relational Works

Kumo Relational operates directly on relational databases, preserving their tables, foreign keys, and timestamps. Unlike most foundation models that work on flat tables, Kumo Relational respects the inherent structure of relational data. It is pre-trained on synthetic relational data that mimics real-world datasets and adapts to new tasks through in-context learning.

To make predictions, Kumo Relational must understand the values within each table, relate records across tables, and connect context examples with known targets. It utilizes a Predictive Query Language (PQL) to define tasks declaratively, allowing users to specify what to predict without detailing how to do it. The model can sample historical database states to prevent data leakage during temporal tasks.

Performance

Kumo Relational has been evaluated on 41 benchmarks across 15 relational databases from RelBench V1 and V2, SAP SALT, and 4DBInfer. It outperforms the strongest supervised models in both classification and regression tasks, including relational deep learning models and expert-engineered tabular ensembles. On the RelArena-α leaderboard, Kumo Relational ranks highly for accuracy and efficiency, demonstrating its capability to deliver results faster than competing models.

In addition to its performance, Kumo Relational shows resilience under challenging conditions, maintaining effectiveness with minimal training data and when faced with missing foreign-key links or noisy columns.

Limitations

While Kumo Relational excels in scenarios with meaningful relationships and temporal dynamics, it may not always outperform task-specific supervised models in cases with very large training sets. The model is designed to be fine-tuned for specific datasets and tasks when maximum accuracy is required.

Getting Started with Kumo Relational

NVIDIA Kumo Relational is available for use, allowing developers to connect databases, define predictive queries, and obtain predictions without the need for extensive training. The model can be integrated into existing systems through NVIDIA’s structured-data-models library, which supports various database connectors and can handle large datasets efficiently.

For more information, developers can refer to the research paper and documentation provided by NVIDIA.

Source: Hugging Face Blog