CSE Colloquium – Towards Relational Foundation Models: Zero-Shot Forecasting over Relational Databases
Presenter: Charilaos I. Kanatsoulis, Stanford University
Abstract: Foundation models have transformed unstructured domains such as language and vision, yet relational datasets, where most enterprise knowledge lives, still rely on brittle, task-specific ML pipelines. I will begin by introducing Relational Deep Learning (RDL), a general framework for learning directly from heterogeneous multi-table data, capturing structure across entities, attributes, and relationships without handcrafted schemas or features.
Building on this paradigm, I will present the Relational Transformer (RT), a schema-invariant model pretrained across diverse relational databases that performs structural learning with in-context information and transfers zero-shot to new databases and predictive tasks. By modeling both inter- and intra-table dependencies and reframing prediction as pattern recognition inside a unified latent relational space, RT represents a concrete step toward relational foundation models that can be prompted, reused, and generalized for new problems.
Bio: Charilaos I. Kanatsoulis is a Research Scientist in the Department of Computer Science at Stanford University. He previously was a Postdoctoral Researcher in the Department of Electrical and Systems Engineering at the University of Pennsylvania and received his Ph.D. in Electrical and Computer Engineering from the University of Minnesota, Twin Cities. His research lies at the intersection of machine learning and signal processing, with a focus on Transformer and foundation model design for structured data, graph representation learning, tensor analysis, and explainable AI. His work has been recognized with the Best Paper Award at the KDD Temporal Graph Learning Workshop (2025) and the Best Student Paper Award at IEEE CAMSAP (2023). He co-instructs CS246 and CS224W at Stanford and previously taught ESE 5140 at Penn. He has organized several community events, including the Graph Signal Processing short course at IEEE ICASSP 2023, the Stanford Graph Learning Workshop (2024–2025), the Relational Deep Learning tutorial at ACM KDD 2025, and the New Perspectives in Advancing Graph Machine Learning Workshop at NeurIPS 2025.
Hosted by: Professor Nikos Tziavelis
Location: Engineering 2, Room E2-180 (Refreshments such as coffee, pastries, and fruit will be provided.)
Zoom: https://ucsc.zoom.us/j/93445911992?pwd=YkJ2TQtF79h0PcNXbEcpZLbpK0coiY.1&jst=3
