Research map / Graph learning / Knowledge and temporal graphs
Griffin: Towards a Graph-Centric Relational Database Foundation Model
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.
Griffin is a graph-centric foundation model for relational databases that unifies encoders and decoders, adds cross-attention and hierarchical MPNN aggregation, and is pretrained on single-table and RDB data.
Problem
Foundation models for relational databases (RDBs) remain underexplored; existing RDB models are task-specific GNNs or single-table models that cannot capture inter-table structure while generalizing across schemas and tasks.
Approach
Unify the data encoder (pretrained text encoder for categorical/text cells, pretrained float encoder for numerical cells, metadata embeddings) and task decoder (label-text-embedding classification head, shared float decoder for regression), with a cross-attention module over cells within a row and a hierarchical MPNN that averages within each relation type before max-aggregating across relations;
Key results
- Griffin-pretrained achieves the best average rank (3.042) across 24 tasks versus 4.792 for the best GNN baseline (SAGE) in the raw results table.
- Griffin-unpretrained reaches average rank 3.708, showing architecture gains without pretraining.
- On Amazon/rating RMSE, Griffin-pretrained scores 0.5994 versus 0.9639 for SAGE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Griffin's architecture without pretraining instead of the strongest listed baseline | RDB prediction tasks · All tasks | average rank: 4.792 -> 3.708 | Helped |
| pretrain on single-table datasets | Griffin · RDB prediction tasks | average rank: 3.708 -> 3.042 | Helped |
| replace cross-attention with average attention (ablation) | Griffin · All tasks | average rank: 1.4 -> 2.7 | Hurt |
| replace hierarchical aggregation with mean-only aggregation (ablation) | Griffin · All tasks | average rank: 1.4 -> 1.9 | Hurt |
| use full SFT rather than limited-sample or mixed SFT (ablation) | Griffin-RDB-SFT · Commerce-2 to Commerce-1; Commerce-1 to Commerce-2; Others-1 to Others-2 | transfer performance | Helped |
| fine-tune on similar-domain commerce data rather than no pretraining | Griffin-RDB-SFT · Commerce-1 to Commerce-2 and Commerce-2 to Commerce-1 | transfer performance | Helped |
| fine-tune on commerce data before transferring to non-commerce tasks | Griffin-RDB-SFT · Commerce to non-commerce | transfer performance | Hurt |
| fine-tune on diverse Others-domain data before transferring to commerce | Griffin-RDB-SFT · Others to Commerce | transfer performance | Helped |
| use Others-1 SFT before transferring to Others-2 | Griffin-RDB-SFT · Others-1 to Others-2 | transfer performance | Helped |
| use Others-2 SFT before transferring to Others-1 | Griffin-RDB-SFT · Others-2 to Others-1 | transfer performance | Hurt |
| restrict SFT to 4096 samples instead of full SFT (ablation) | Griffin-RDB-SFT · Cross-domain transfer | transfer performance | Hurt |
Surprising findings
- Transfer from commerce to non-commerce domains often underperforms the no-pretrain baseline, so pretraining helps only when domains are similar or diverse.
- In the original RelBench regression results, values were not comparable to aligned results after normalization (e.g., rel-f1/position 4.022 original vs 0.5945 aligned), so reported baselines were re-evaluated.
Limitations
- Baseline results were re-run or modified for standardized preprocessing, so numbers may differ from original benchmark reports.
- Few-shot comparison with TabPFNv2 shows Griffin and TabPFNv2 each excel under different conditions rather than a clear winner.
- Pretraining corpus excluded large-scale RDB datasets that diverged in distribution, limiting the foundation-model scope.
Experimental setup
Datasets: TPBerta single-table pretraining datasets CARTE datasets 4DBInfer RDBs RelBench RDBs
Benchmarks: 4DBInfer RelBench
Metrics: ROC-AUC accuracy RMSE MAE logloss average rank
Baselines: SAGE GAT PNA HGT DFS+MLP DFS+DeepFM DFS+FTTransformer DFS+XGBoost TabPFNv2
Builds on: 4DBInfer RelBench TPBerta CARTE Deep Feature Synthesis
Closest papers by meaning
- RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.90)
- 4DBInfer: A 4D Benchmarking Toolbox for Graph-Centric Predictive Modeling on RDBs (NeurIPS 2024 · similarity 0.88)
- Position: Relational Deep Learning - Graph Representation Learning on Relational Databases (ICML 2024 · similarity 0.87)
- A Pre-training Framework for Relational Data with Information-theoretic Principles (NeurIPS 2025 · similarity 0.86)
- Joint Relational Database Generation via Graph-Conditional Diffusion Models (NeurIPS 2025 · similarity 0.86)
- RelGNN: Composite Message Passing for Relational Deep Learning (ICML 2025 · similarity 0.86)
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.