Relational databases contain multiple tables whose connections can provide valuable information for prediction tasks. By representing rows as nodes and primary–foreign key relationships as edges, graph neural networks can learn across tables while retaining their relational structure.

In order to theoretically advance expressiveness, we introduce inner-table connections based on similarity of the embedded item (e.g. connecting rows with comparable transaction times or concerts in the same city). We assume that the effects are largest on not fully-normalized databases which are common in practice, but not regularly used in benchmarking. Since initial experiments suggested that the difference is not huge, explaining why additional expressivity does not help in such cases can also be part of a possible thesis.

Starting literature:

• Relational Deep Learning (ICML 2024)
• RelBench (NeurIPS 20249
• RelGNN (2025)