Improving Relational Deep Learning via Language Model-based Graph Augmentation

Abstract

Graph Neural Networks (GNNs) have recently been adopted in relational deep learning (RDL) to enable machine learning directly on relational databases (RDBs) composed of multiple tables linked via primary-foreign key (PFK) relationships. However, existing RDL methods that construct graphs solely from PFK relationships face two major challenges: sparsity in the resulting graph structure and indirect, multi-hop connections between semantically related nodes. These limitations hinder effective information propagation and reduce the utility of the relational data. To address these challenges, we propose a Language model-based Graph Augmentation (LGA) framework that enriches relational graphs by inferring additional semantic links between nodes using textual attributes commonly found in RDBs. These augmented edges depict implicit relationships and improve connectivity, particularly for new or sparsely linked nodes. Additionally, LGA introduces a dual-branch GNN architecture that separates message passing along structural (PFK-based) and semantic (text-based) edges, enabling finer control over their respective influence. Experiments on benchmark datasets from RelBench demonstrate that our approach consistently outperforms existing RDL methods in predictive performance.

Ähnliche Arbeiten