RelBench
Benchmark for deep learning directly on relational databases (Robinson et al., 2024), cited in Relational Reranking - Scoring Search Results with Structured Facts as precedent for conditioning models on relational structure rather than hand-engineered tabular features. Per that article, relational deep learning systems that learn across primary/foreign-key links matched or beat tuned tabular-model baselines on 11 of RelBench’s 15 tasks, because the relational models gain access to fields outside a task’s target table — the architectural motivation behind RelativeDB’s Relational Transformer (RT) reranker.
Related Concepts
- Relational Transformer — RT’s Cell-attention mechanism extends the cross-table structure RelBench demonstrates value for, into reranking