RelGT-AC: A Relational Graph Transformer for Autocomplete Tasks in Relational Databases

In three linesRelGT-AC extends RelGT for autocomplete tasks on relational databases. The model combines column masking, a unified task head (classification/regression), and a TF-IDF encoder for free text. On 7 RelBench v2 tasks, it outperforms GraphSAGE on 3 regression tasks and gains +10 AUROC on text-heavy eligibility tasks.

Summary generated by Claude — human-verified