KumoRFM: A Relational Foundation Model for Revolutionizing Relational Database Predictions

KumoRFM is a groundbreaking Relational Foundation Model (RFM) capable of making accurate predictions on relational databases across a wide range of predictive tasks without requiring any data or task-specific training. It achieves this by transforming databases into temporal, heterogeneous graphs, employing a table-invariant encoding scheme and a Relational Graph Transformer to reason across multimodal data between tables. On the RelBench benchmark, KumoRFM outperforms traditional feature engineering and end-to-end supervised deep learning approaches by 2% to 8% on average, with further improvements of 10% to 30% after fine-tuning. Most importantly, KumoRFM is orders of magnitude faster than conventional supervised training approaches, offering a zero-code solution for real-time predictions.