Research map / Image generation / Fast diffusion sampling
ClavaDDPM: Multi-relational Data Synthesis with Cluster-guided 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 ClavaDDPM, a multi-table synthesizer that uses relationship-aware GMM clustering labels as latent intermediaries for foreign-key parent-child dependencies, with a Gaussian diffusion backbone, classifier guidance, and nearest-neighbor matching for multi-parent children.
Problem
Synthesizing multi-relational (multi-table) relational databases that preserve foreign-key group structure and long-range inter-table correlations, which existing methods (SDV, PrivLava) fail to scale to or capture.
Approach
Learn cluster latent variables on the joint child-parent space (weighted GMM), augment parent tables with these labels, model the augmented tables with a single Gaussian diffusion backbone using label encoding, generate children via classifier-guided diffusion conditioned on latent clusters, and resolve multi-parent divergence by approximate nearest-neighbor matching and interpolation.
Key results
- On Instacart 05, ClavaDDPM outperforms the best baseline by 58.29% on 2-hop correlations (avg 2-way 76.02 vs. 56.19-66.66 for baselines).
- On Berka, ClavaDDPM achieves avg 2-way score 89.21 versus 82.33 for the best single-table baseline (ST-ClavaDDPM), with 3-hop correlations exceeding the best baseline by 20.24%.
- PrivLava is strongest on California (avg 2-way 97.97 vs 96.41 for ClavaDDPM) but fails to converge on the other datasets; TabDDPM-based pipelines hit the 7-day time limit on several datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ClavaDDPM instead of the strongest reported baseline | Multi-table synthesis · California | avg 2-way: 97.97 -> 96.41 | Hurt |
| Use ClavaDDPM instead of the strongest reported baseline | Multi-table synthesis · Instacart 05 | avg 2-way: 66.66 -> 76.02 | Helped |
| Use ClavaDDPM instead of the strongest reported baseline | Multi-table synthesis · Berka | avg 2-way: 82.33 -> 89.21 | Helped |
| Use ClavaDDPM instead of the strongest reported baseline | Multi-table synthesis · Movie Lens | avg 2-way: 94.87 -> 97.11 | Helped |
| Use ClavaDDPM instead of the strongest reported baseline | Multi-table synthesis · CCS | avg 2-way: 68.51 -> 87.33 | Helped |
| Use k=1 cluster instead of the default k=20 (ablation) | ClavaDDPM · Berka | avg 2-way: 89.21 -> 81.64 | Hurt |
| Use k=1000 clusters instead of the default k=20 (ablation) | ClavaDDPM · Berka | avg 2-way: 89.21 -> 87.77 | Hurt |
| Set parent scale λ=0 instead of λ=1.5 (ablation) | ClavaDDPM · Berka | avg 2-way: 89.21 -> 84.01 | Hurt |
| Set parent scale λ=10 instead of λ=1.5 (ablation) | ClavaDDPM · Berka | avg 2-way: 89.21 -> 87.52 | Hurt |
| Set parent scale λ=100 instead of λ=1.5 (ablation) | ClavaDDPM · Berka | avg 2-way: 89.21 -> 86.36 | Hurt |
| Set classifier gradient scale η=0 instead of η=1 (ablation) | ClavaDDPM · Berka | avg 2-way: 89.21 -> 84.69 | Hurt |
| Set classifier gradient scale η=2 instead of η=1 (ablation) | ClavaDDPM · Berka | avg 2-way: 89.21 -> 87.57 | Hurt |
Surprising findings
- PrivLava, a dedicated multi-relational method, achieves the best single-table-style scores on California yet did not converge on four of five datasets.
- Setting the number of clusters k=1 yields perfect agree rate (100%) but degrades long-range correlations drastically, showing higher agreement does not imply better quality.
Limitations
- Latent learning uses a simple diagonal GMM; more complex latent models (e.g., TabSyn-style VAE) are left for future work.
- Multi-parent matching is a heuristic approximate nearest-neighbor interpolation without one-to-one guarantees and sampled only once.
- The long-range metric is based on pairwise column correlations and does not directly assess downstream task utility on multi-table data.
Experimental setup
Datasets: California Instacart 05 Berka Movie Lens CCS
Benchmarks: California Instacart 05 Berka Movie Lens CCS
Metrics: cardinality 1-way density k-hop pair-wise column correlation avg 2-way alpha-precision beta-recall C2ST machine learning efficacy (R2, F1)
Baselines: PrivLava SDV ST-CTGAN ST-TabDDPM D-CTGAN D-TabDDPM
Builds on: TabDDPM DDPM classifier guidance PrivLava SDV
Closest papers by meaning
- TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · similarity 0.89)
- Joint Relational Database Generation via Graph-Conditional Diffusion Models (NeurIPS 2025 · similarity 0.88)
- CTSyn: A Foundation Model for Cross Tabular Data Generation (ICLR 2025 · similarity 0.87)
- Mixed-Type Tabular Data Synthesis with Score-based Diffusion in Latent Space (ICLR 2024 · similarity 0.86)
- WikiDBs: A Large-Scale Corpus Of Relational Databases From Wikidata (NeurIPS 2024 · similarity 0.86)
- Continuous Diffusion for Mixed-Type Tabular Data (ICLR 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.