Relational Hypergraph Transformer Unifies Multi-Table Learning

Summary

Researchers introduce the Relational Hypergraph Transformer (RHT), a hypergraph-based architecture for complex multi-table machine learning. It combines pentadimensional embeddings with sparse relational attention whose complexity scales with average relational degree. On Synthea electronic health records, RHT produced the most semantically coherent embeddings among the evaluated baselines, while XGBoost achieved the highest rare-code recall. The study includes ablations, open-source code, and planned MIMIC-IV validation.