Relational Foundation Models for Enterprise Data with Jure Leskovec - #768

2026-05-21 · Source: The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Jure Leskovec, co-founder of Kumo and Stanford professor, discusses his work in AI for science and relational deep learning. His AI Virtual Cell project develops multiscale foundation models for biomedical data, learning representations from proteins to patients using single-cell RNA-seq, ESM, and AlphaFold without hand-encoding biology. He also details Kumo's Relational Foundation Model (RFM2), a pre-trained system that reframes enterprise databases as graphs. RFM2 performs in-context learning over subgraphs to make accurate predictions on new databases and tasks without model training, achieving up to 12% relative accuracy gains over state-of-the-art supervised models on benchmarks like RELBench and SALT. Real-world deployments include Reddit, DoorDash, and Coinbase, demonstrating significant business impact in areas like ad click-through rates and fraud detection.

Key takeaway

For AI Engineers building predictive systems on complex enterprise data, you should evaluate Relational Foundation Models like Kumo's RFM2. This approach eliminates extensive manual feature engineering and can deliver double-digit accuracy improvements on multi-table problems, even with noisy or cold-start data. Consider integrating RFM2 via its API into agentic workflows to enable real-time, context-aware decision-making without continuous model retraining.

Key insights

Relational Foundation Models enable accurate, no-training predictions on multi-table enterprise data by learning directly from graph-structured relationships.

Principles

• Treat enterprise databases as graphs for deep learning.
• Neural networks can learn directly from raw relational data.
• In-context learning allows pre-trained models to predict without fine-tuning.

Method

RFM2 extracts labeled in-context subgraphs from a database based on a specified task, then performs a single forward pass through a pre-trained neural network to make predictions.

In practice

• Apply RFM2 for fraud detection and recommender systems.
• Integrate RFM2 via agent-friendly APIs for autonomous decision-making.

Topics

• Relational Deep Learning
• Foundation Models
• Enterprise Data
• Graph Neural Networks
• In-Context Learning
• AI for Science
• Fraud Detection

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• SE Radio 724: Jure Leskovec on Relational Graph and Foundational Models — Relational deep learning transforms enterprise tabular data into graphs for superior, automated predictive AI.
• Why Foundation Models Haven’t Replaced Classical Machine Learning — Classical ML remains vital for proprietary, structured data, requiring context-aware agentic systems to navigate fragmented enterprise ecosystems.
• No Need to Train Your RDB Foundation Model — A new RDB encoder enables existing single-table ICL foundation models to operate on multi-table RDBs without retraining.
• Relational Foundation Models for Enterprise Data [Jure Leskovec] - 768 — Relational Foundation Models leverage graph neural networks for in-context learning on raw multi-tabular data, enabling accurate predictions without explicit training.
• How do One Big Table and AI fit together — Context engineering is crucial for AI to deliver consistent, low-latency, and cost-effective answers.
• Context Payload Optimization for ICL-Based Tabular Foundation Models — Optimizing context payloads is crucial for balancing accuracy, latency, and cost in ICL tabular foundation models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence).