SE Radio 724: Jure Leskovec on Relational Graph and Foundational Models

2026-06-10 · Source: Software Engineering Radio - the podcast for professional software developers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, extended

Summary

Jure Leskovec, Professor of Computer Science at Stanford University and Chief Scientist at Kumo.ai, highlights relational deep learning and relational foundation models as a transformative solution for enterprise predictive modeling. He notes that traditional AI, while advanced in NLP and computer vision, has largely neglected operational data in relational databases, relying on outdated, expensive, and manual feature engineering methods. The proposed approach represents databases as graphs of linkages, utilizing specialized attention mechanisms to learn directly from raw tables. This method, differing from traditional Graph Neural Networks, offers "superhuman accuracy" and "productivity gains," evidenced by 30-50% improvements at Pinterest. These models are also less data-hungry and computationally intensive than large language models, enabling on-the-fly predictions, accurate uncertainty estimates, and natural language explanations for critical business decisions.

Key takeaway

For Data Scientists and Machine Learning Engineers tasked with building predictive models over complex enterprise data, consider adopting relational deep learning and foundation models. You can bypass time-consuming manual feature engineering and achieve significantly higher accuracy by training directly on raw relational data. Explore platforms like Kumo.ai or open-source frameworks such as PyTorch Geometric to implement graph-based attention, enhancing decision-making and operational efficiency.

Key insights

Relational deep learning transforms enterprise tabular data into graphs for superior, automated predictive AI.

Principles

• Tabular data is a "missing modality" in current AI.
• Attention mechanisms excel at combining raw relational data.
• Graph-based attention avoids over-smoothing issues of GNNs.

Method

Represent relational databases as graphs of linkages, then apply specialized attention-based architectures (relational transformers) to learn patterns directly from raw tables for prediction.

In practice

• Improve fraud detection accuracy.
• Optimize customer churn prediction.
• Generate natural language explanations for predictions.

Topics

• Relational Deep Learning
• Graph Neural Networks
• Foundation Models
• Predictive Modeling
• Enterprise AI
• Feature Engineering
• PyTorch Geometric

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

Related on AIssential

• Is XGBoost gone: How Relational Foundation Models Conquered 500 Billion Row Enterprise Data — The dominance of gradient boosted trees on tabular data was an illusion created by pre-processed Kaggle datasets.
• Is Fixing Schema Graphs Necessary? Full-Resolution Graph Structure Learning for Relational Deep Learning — FROG enables learnable graph structure for Relational Deep Learning by modeling table roles and enforcing semantic consistency.
• Learning to Extrapolate to New Tasks: A Relational Approach to Task Extrapolation — Extrapolation to unseen tasks fundamentally requires learning the relational transformations between tasks.
• Relational Foundation Models for Enterprise Data with Jure Leskovec - #768 — Relational Foundation Models enable accurate, no-training predictions on multi-table enterprise data by learning directly from graph-structured relationships.
• Beyond the lakehouse: Fundamental's NEXUS bypasses manual ETL with a native foundation model for tabular data — NEXUS is a Large Tabular Model designed to natively understand and predict from complex, non-sequential business data.
• The "Final Boss" of Deep Learning — Intrinsic architectural improvements, not just tool use, are crucial for robust algorithmic reasoning in language models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Software Engineering Radio - the podcast for professional software developers.