Research map / Graph learning / Graph self-supervised learning
Joint Relational Database Generation via Graph-Conditional Diffusion Models
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 GRDM, a non-autoregressive graph-conditional diffusion model that generates all tables of a relational database jointly by sampling a degree-preserving graph structure and then denoising row attributes with a GNN over K-hop neighborhoods.
Problem
Synthetic relational database (RDB) generation is typically autoregressive over tables, which limits parallelism, flexibility for downstream tasks, and captures multi-hop inter-table dependencies poorly due to compounding errors.
Approach
Represent an RDB as a directed heterogeneous graph (rows as nodes, primary-foreign key links as edges), sample the graph structure with a degree-preserving random graph model, and jointly denoise all node attributes with a Gaussian diffusion model whose denoiser is a heterogeneous GraphSAGE over the K-hop neighborhood; dimension tables are kept fixed as conditioning metadata.
Key results
- GRDM achieves the best inter-table trend scores on all databases, e.g., 95.57 vs 85.87 best baseline on Berka 2-hop and 92.95 vs 16.22 on Instacart 2-hop.
- On single-table metrics GRDM is comparable to the best methods; it is slightly below SingleTable/ClavaDDPM on some columns, e.g., Movie Lens intra-table trends 95.25 vs 98.56.
- The autoregressive variant of GRDM performed significantly worse (e.g., Berka 2-hop 52.01 vs 95.57) and was more than 2.5x slower on Berka.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use joint rather than autoregressive feature generation (ablation) | GRDM · Berka, Inter-Table Trends (1-, 2-, and 3-hop) | Inter-Table Trends: 91.41 -> 28.53; 95.57 -> 52.01; 92.43 -> 49.45 | Helped |
| use joint rather than autoregressive feature generation (ablation) | GRDM · Instacart 05, Inter-Table Trends (1- and 2-hop) | Inter-Table Trends: 88.49 -> 63.32; 92.95 -> 13.41 | Helped |
| use joint rather than autoregressive feature generation (ablation) | GRDM · RelBench-F1, Inter-Table Trends (1- and 2-hop) | Inter-Table Trends: 90.85 -> 90.82; 93.67 -> 87.73 | Mixed |
| replace the sampled graph with the ground-truth graph (ablation) | GRDM · Berka, Inter-Table Trends (1-, 2-, and 3-hop) | Inter-Table Trends: 91.41 -> 91.50; 95.57 -> 95.90; 92.43 -> 94.45 | No clear effect |
| replace the sampled graph with the ground-truth graph (ablation) | GRDM · Instacart 05, Inter-Table Trends (1- and 2-hop) | Inter-Table Trends: 88.49 -> 91.80; 92.95 -> 92.85 | Mixed |
| replace the sampled graph with the ground-truth graph (ablation) | GRDM · RelBench-F1, Inter-Table Trends (1- and 2-hop) | Inter-Table Trends: 90.85 -> 93.12; 93.67 -> 96.10 | Helped |
| use GRDM instead of the strongest baseline | relational database generation · Berka, Inter-Table Trends (3-hop) | Inter-Table Trends: +14.14% | Helped |
| use GRDM instead of the strongest baseline | relational database generation · Instacart 05, Inter-Table Trends (2-hop) | Inter-Table Trends: +473.06% | Helped |
| use GRDM instead of the strongest baseline | relational database generation · RelBench-F1, Inter-Table Trends (2-hop) | Inter-Table Trends: +12.62% | Helped |
| use GRDM instead of the strongest baseline | relational database generation · Movie Lens, Inter-Table Trends (1-hop) | Inter-Table Trends: +1.75% | Helped |
| use GRDM instead of the strongest baseline | relational database generation · CCS, Inter-Table Trends (1-hop) | Inter-Table Trends: +17.52% | Helped |
| use GRDM instead of the strongest baseline | relational database generation · California, Inter-Table Trends (1-hop) | Inter-Table Trends: +0.86% | Helped |
Surprising findings
- Single-table fidelity is not uniformly improved; GRDM underperforms some baselines on intra-table trends for Instacart 05 and Movie Lens.
- Autoregressive table-by-table variant also hurts single-table metrics, not only inter-table ones.
Limitations
- Graph generation uses a simple degree-preserving model that the authors acknowledge could be more tailored, with a gap to real-graph performance on Berka.
- K=1 chosen for efficiency; higher K might improve fidelity at more compute cost.
- Only unconditional generation is studied; privacy guarantees are not provided.
Experimental setup
Datasets: Berka Instacart 05 Movie Lens CCS California RelBench-F1
Benchmarks: Berka Instacart 05 Movie Lens CCS California RelBench-F1
Metrics: cardinality column shapes intra-table trends inter-table trends (k-hop) KS complement TV complement DCR
Baselines: SDV ClavaDDPM SingleTable Denorm
Builds on: TabDDPM DDPM GraphSAGE ClavaDDPM SDV
Closest papers by meaning
- RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.89)
- ClavaDDPM: Multi-relational Data Synthesis with Cluster-guided Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Griffin: Towards a Graph-Centric Relational Database Foundation Model (ICML 2025 · similarity 0.86)
- Position: Relational Deep Learning - Graph Representation Learning on Relational Databases (ICML 2024 · similarity 0.86)
- 4DBInfer: A 4D Benchmarking Toolbox for Graph-Centric Predictive Modeling on RDBs (NeurIPS 2024 · similarity 0.86)
- WikiDBs: A Large-Scale Corpus Of Relational Databases From Wikidata (NeurIPS 2024 · 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.