Abstract
Relational databases (RDBs) are ubiquitous in enterprise and real-worldapplications. Flattening the database poses challenges for deep learning modelsthat rely on fixed-size input representations to capture relational semanticsfrom the structured nature of relational data. Graph neural networks (GNNs)have been proposed to address this, but they often oversimplify relationalstructures by modeling all the tuples as monolithic nodes and ignoringintra-tuple associations. In this work, we propose a novel hypergraph-basedframework, that we call rel-HNN, which models each unique attribute-value pairas a node and each tuple as a hyperedge, enabling the capture of fine-grainedintra-tuple relationships. Our approach learns explicit multi-levelrepresentations across attribute-value, tuple, and table levels. To address thescalability challenges posed by large RDBs, we further introduce asplit-parallel training algorithm that leverages multi-GPU execution forefficient hypergraph learning. Extensive experiments on real-world andbenchmark datasets demonstrate that rel-HNN significantly outperforms existingmethods in both classification and regression tasks. Moreover, oursplit-parallel training achieves substantial speedups -- up to 3.18x forlearning on relational data and up to 2.94x for hypergraph learning -- comparedto conventional single-GPU execution.