KumoRFM Foundation Model for Relational Data

This document is a technical report that presents Kumo Relational Foundation Model (KumoRFM), a novel pre-trained model designed specifically for making accurate predictions on relational databases without requiring data- or task-specific training. KumoRFM utilizes a table-agnostic encoding scheme and a Relational Graph Transformer to effectively handle complex multi-table relational data. The report outlines the capabilities of KumoRFM, including its ability to adapt to unseen database schemas, accommodate diverse column types, and perform a wide range of predictive tasks such as churn detection and fraud identification. It emphasizes the model's efficiency, delivering predictions in under one second, and its potential for scalable and explainable AI applications. The report also discusses the model's performance compared to conventional approaches, highlighting its advantages in terms of speed and accuracy. KumoRFM is evaluated on various predictive tasks using publicly available datasets, demonstrating its effectiveness in real-world applications.