Publications
Selected publications from the group on foundation models, datasets, and benchmarks for tabular and relational data. Papers, code, datasets, and model weights are openly available.
A framework for synthesizing multi-table relational databases. The resulting synthetic data yields power-law scaling during pretraining and improves performance on real-world RelBench tasks.
An architecture that transfers across relational databases without fine-tuning. A 22M-parameter model attains 93% of fully-supervised AUROC in a single forward pass, exceeding a 27B-parameter language model.
A graph transformer architecture designed for the structure of relational entity graphs.
A large-scale benchmark and repository for relational data, providing standardized datasets and tasks for evaluation.
A comprehensive review of relational deep learning, surveying its challenges, foundations, and next-generation architectures.
Rel-LLM, an architecture that pairs a graph neural network encoder with a large language model through retrieval-augmented generation, bringing the reasoning of LLMs to relational databases.
A composite message-passing scheme for relational deep learning that addresses many-to-many relationships, achieving state-of-the-art results on RelBench with improvements of up to 25%.
A benchmark for deep learning on relational databases, comprising seven databases and thirty predictive tasks across diverse domains.
A position paper introducing relational deep learning, which represents relational databases as graphs to enable end-to-end learning without manual feature engineering.