Research map / Graph learning / GNN architectures
RelGNN: Composite Message Passing for Relational Deep Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RelGNN, a GNN for relational deep learning that uses composite message passing over automatically derived atomic routes (single-hop src→mid→dst paths) with graph attention, evaluated on 30 RelBench tasks.
Problem
Standard heterogeneous GNNs treat primary-foreign key edges of relational entity graphs as semantic edges, causing redundant two-hop message passing and over-smoothing through bridge and hub nodes that arise from many-to-many junction tables.
Approach
Define atomic routes as simple paths derived from primary-foreign key relations (source→intermediate→destination for tables with multiple foreign keys) and perform composite single-step message passing along them, using a linear FUSE of intermediate and source embeddings and multi-head attention for AGGR.
Key results
- RelGNN surpasses the heterogeneous GraphSAGE baseline on 27 of 30 RelBench tasks while performing comparably on the remaining three (per the introduction).
- Gains reach up to 25% on the site-success regression task in rel-trial.
- Ablation shows RelGNN without attention still improves over baselines, suggesting atomic routes drive most of the gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RelGNN instead of the standard heterogeneous GNN | Entity classification on RelBench · RelBench entity classification tasks | ROC-AUC: 10 out of 12 tasks | Helped |
| Use RelGNN instead of the standard heterogeneous GNN | Entity regression on RelBench · RelBench entity regression tasks | MAE: 8 out of 9 tasks | Helped |
| Use RelGNN instead of the baselines | Recommendation on RelBench · RelBench recommendation tasks | MAP: all 9 tasks | Helped |
| Replace standard heterogeneous GNN message passing with atomic routes, using GraphSAGE-style aggregation (ablation) | RelGNN · RelBench classification, regression, and recommendation tasks | task performance | Helped |
| Remove attention and use GraphSAGE-style aggregation (ablation) | RelGNN · RelBench classification, regression, and recommendation tasks | task performance | No clear effect |
| Add relative positional encoding over the schema graph | RelGNN on rel-stack · rel-stack classification tasks | ROC-AUC: 2% | Helped |
Surprising findings
- Gains are limited on rel-stack, which contains self-loop primary-foreign key links between posts nodes.
- The ablation with attention removed performs only modestly worse than full RelGNN, indicating attention contributes little beyond atomic routes.
Limitations
- Weak performance on rel-stack due to self-loop tables, left for future work.
- ID-GNN recommendation head restricts candidates to sampled subgraph, hurting low-locality tasks.
- Regression head may limit performance on unbounded targets such as study-adverse.
Experimental setup
Datasets: RelBench rel-amazon RelBench rel-avito RelBench rel-event RelBench rel-f1 RelBench rel-hm RelBench rel-stack RelBench rel-trial
Benchmarks: RelBench
Metrics: ROC-AUC MAE MAP@K
Baselines: LightGBM Hetero-GNN (heterogeneous GraphSAGE) GAT GIN
Builds on: RelBench Relational Deep Learning GraphSAGE meta-paths PyTorch Frame
Closest papers by meaning
- Position: Relational Deep Learning - Graph Representation Learning on Relational Databases (ICML 2024 · similarity 0.89)
- Locality-Aware Graph Rewiring in GNNs (ICLR 2024 · similarity 0.86)
- TopoTune: A Framework for Generalized Combinatorial Complex Neural Networks (ICML 2025 · similarity 0.86)
- Recurrent Graph Neural Networks and Their Connections to Bisimulation and Logic (AAAI 2024 · similarity 0.86)
- Griffin: Towards a Graph-Centric Relational Database Foundation Model (ICML 2025 · similarity 0.86)
- RelAgent: LLM Agents as Data Scientists for Relational Learning (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.