LoG² - Learning on Graphs and Geometry

​While foundation models have revolutionized unstructured data, structured enterprise data remains siloed in relational databases, typically requiring labor-intensive, task-specific model development.

​In this talk, we introduce the Kumo Relational Foundation Model (KumoRFM), a pre-trained architecture designed to generalize across arbitrary relational schemas and diverse predictive tasks without retraining.

​Central to this approach is the paradigm of Relational Deep Learning (RDL), which represents relational databases as temporal heterogeneous graphs where records are nodes and primary-foreign key links define the edges.

​KumoRFM employs a table-agnostic encoding scheme and a novel Relational Graph Transformer to reason across arbitrary multi-modal data and complex multi-table structures. This architecture enables in-context learning (ICL) by dynamically generating task-specific context labels from historical snapshots via temporal neighbor sampling, allowing the model to reason about the dependencies between subgraph-label pairs in a single forward pass.

​We demonstrate that KumoRFM delivers accurate predictions in under one second and outperforms conventional supervised approaches on the RelBench benchmark, paving the way for scalable and explainable AI on enterprise data.