Date: Friday, October 17
Time: 12-1pm
Location: Soda 510
Title: Relational Foundation Models
Abstract: Foundation Models have transformed how we interact with unstructured data enabling seamless in-context learning across text, images, and code. Yet, the structured data that drives core decisions in enterprises—transaction logs, customer journeys, events, time series—remains locked behind brittle pipelines and handcrafted machine learning models. In this talk, I will introduce Relational Foundation Models (RFMs), a new class of pre-trained models that unlock in-context learning over relational data, just as LLMs did for language. RFMs model multi-table, heterogeneous graph-structured data and can predict complex outcomes, such as user engagement, purchases, churn, fraud, and recommendations. All this without per-task supervision, feature engineering, or model training. I will describe the architecture and training objectives of RFMs, which combine table-agnostic embeddings, relational transformers, and SQL-like prompt interfaces. The result is a single general-purpose model that makes accurate, fast predictions across a broad class of tasks, often outperforming traditional supervised pipelines built over months. We will explore how RFMs reshape the paradigm of predictive AI—from model-building as a craft to model-use as querying—and what this means for the future of recommender systems, classification, regression, and more. I will argue that RFMs are not a replacement for LLMs but their structured-data complement—together forming the foundation for the next generation of enterprise AI.
Bio: Jure Leskovec <http://cs.stanford.edu/~jure> is Professor of Computer Science at Stanford University. He is affiliated with the Stanford AI Lab, the Machine Learning Group and the Center for Research on Foundation Models. In the past, he served as a Chief Scientist at Pinterest and was an investigator at Chan Zuckerberg BioHub. Most recently, he co-founded AI startup Kumo.AI. Leskovec pioneered the field of Graph Neural Networks and created PyG, the most widely-used graph neural network library. Research from his group has been used by many countries to fight COVID-19 pandemic, and has been incorporated into products at Meta, Pinterest, Uber, YouTube, Amazon, and more. His research contributions have spanned social networks, data mining and machine learning, and computational biomedicine with the focus on drug discovery. His work has won 13 best paper awards and 6 10-year test of time awards at premier venues in these research areas. Leskovec received his bachelor’s degree in computer science from University of Ljubljana, Slovenia, PhD in machine learning from Carnegie Mellon University and postdoctoral training at Cornell University