MetaRTL is a two‑stage framework for relational table learning. In the first stage lightweight pre‑training yields initial table embeddings; in the second stage non‑parametric message passing constructs meta‑path features, which are aggregated by the attention module MetaAttn. By shifting the heavy computation of deep message passing to efficient meta‑path aggregation, the model captures rich relational semantics while remaining highly efficient. Experiments on ten real‑world datasets covering twenty‑four tasks demonstrate that the approach outperforms traditional deep GNN/HGNN‑based methods in both accuracy and runtime.
Review: MetaRTL achieves a strong trade‑off between performance and scalability for large relational databases, making it a promising candidate for deployment in real systems.