A Fair Benchmarking of Deep Relational Database Learning Models

This study systematically benchmarks recently released deep learning methods for relational databases under a consistent experimental protocol. The evaluation covers five relational databases, with one classification and one regression task per database. Relational Transformer (RT) achieves the strongest overall results against graph-based approaches. Extending context from a single table to neighboring tables improves performance, although gains diminish at higher hops as computational cost increases. The benchmark also reports that deep relational database models outperform TabPFN 2.5 on single-table tasks. Refactored implementations and source code are publicly available.