MetaRTL Scales Relational Table Learning via Meta-path Attention
September 18, 2026
MetaRTL replaces deep GNN/HGNN stacks with a two-stage framework using lightweight pre-training and non-parametric message passing. It uses a MetaAttn module to aggregate meta-path features, providing higher efficiency and scalability on large real-world relational databases.
HOW THIS AFFECTS YOU
●
researcherYou can achieve better scalability on relational datasets by shifting from deep message passing to meta-path aggregation.