Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
On Consistent Bayesian Inference from Synthetic Data
Ossi Räisä, Joonas Jälkö, Antti Honkela
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enlarge synthetic data sets to 5 times the real data size (ablation) | Mixture of synthetic data posteriors (Gaussian mean, known variance) · Gaussian mean estimation (n_X=100) | match to analyst posterior (qualitative): n_X* = n_X -> n_X* = 5n_X | Improved |
| enlarge synthetic data sets to 20 times the real data size (ablation) | Mixture of synthetic data posteriors (Gaussian mean, known variance) · Gaussian mean estimation (n_X=100) | match to analyst posterior (qualitative): n_X* = 5n_X -> n_X* = 20n_X | Improved |
| increase number of synthetic data sets m together with their size (ablation) | Mixture of synthetic data posteriors (Gaussian mean) · Gaussian mean estimation (n_X=100) | smoothness of mixture posterior (qualitative) | Improved |
| analyst uses too small known variance (1 instead of 4), uncongenial (ablation) | Mixture of synthetic data posteriors (Gaussian mean) · Gaussian mean estimation (n_X=100) | convergence target (qualitative): analyst variance 4 -> 1 | No clear effect |
| analyst uses incorrect known mean in variance estimation, uncongenial (ablation) | Mixture of synthetic data posteriors (Gaussian variance) · Gaussian variance estimation with known mean | convergence to data provider posterior (qualitative) | Hurt |
| add mean correction to the mixture of posteriors (ablation) | Mixture of synthetic data posteriors (Gaussian variance) · Gaussian variance estimation with known mean | variance and shape match to provider posterior (qualitative) | Improved |
| use mixture of synthetic data posteriors instead of DP-GLM | DP Bayesian logistic regression · Toy data logistic regression (eps 0.5 and 1) | credible interval width: n_X*/n_X = 20, m = 400 | Improved |
| lower privacy budget to epsilon 0.1 with fixed m and n_X* (ablation) | Mixture of synthetic data posteriors · Toy data logistic regression | credible interval coverage: eps 1 and 0.5 -> eps 0.1 | Hurt |
| increase only n_X* or only m (ablation) | Mixture of synthetic data posteriors · Toy data logistic regression | total variation distance | Mixed |
| use mixture of synthetic data posteriors instead of DPVI | DP Bayesian logistic regression · UCI Adult logistic regression (eps 0.25, 0.5, 1) | credible interval coverage: n_X*/n_X = 10, m = 100 | Improved |
| use mixture of synthetic data posteriors instead of DPVI | DP Bayesian logistic regression · UCI Adult logistic regression (eps 0.25, 0.5, 1) | credible interval width | Hurt |
| use KL divergence instead of posterior-mean distance for tuning DPVI (ablation) | DPVI · UCI Adult logistic regression | posterior width | Hurt |
Closest papers by meaning
- Analysis of Differentially Private Synthetic Data: A Measurement Error Approach (AAAI 2024 · similarity 0.87)
- On Inherent Privacy of Posterior Sampling: A Unified R\'enyi-Divergence Framework (NeurIPS 2026 · similarity 0.87)
- Privacy Amplification Through Synthetic Data: Insights from Linear Regression (ICML 2025 · similarity 0.86)
- Privacy Guarantees in Posterior Sampling under Contamination (NeurIPS 2026 · similarity 0.86)
- Debiasing Synthetic Data Generated by Deep Generative Models (NeurIPS 2024 · similarity 0.86)
- Valid Inference with Imperfect Synthetic Data (NeurIPS 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.
