Synthetic Data Generation
Parent: Data Analysis · Published reference · snapshot 2026-09-08 · skill da-applied-and-communication/references/da-35-synthetic-data-generation.md
↓ Facts as markdownall context files
Synthetic data is artificial data produced by a model fit to real data, designed to reproduce the real data's statistical properties (marginals, correlations, joint structure) without being a copy of
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Synthetic Data Generation
- Synthetic data is artificial data produced by a model fit to real data, designed to reproduce the real data's statistical properties (marginals, correlations, joint structure) without being a copy of any real record. The goal is to make synthetic records useful for a downstream task while breaking the one-to-one link to real individuals. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#synthetic-data-generation]
- This skill treats synthetic data generation as a discipline: when to use it, which generator to reach for, how to make it private, and most importantly how to prove it is good enough on the three axes that always compete: fidelity, utility, privacy. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#synthetic-data-generation]
When to use this skill
- Use it when the task is to create data rather than analyze existing data: share data you cannot share in raw form, augment a too-small or imbalanced training set, populate a test/dev/CI environment without production PII, or release a public benchmark from a sensitive source. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#when-to-use-this-skill]
- Do not use it for plain class-rebalancing with off-the-shelf SMOTE (data-prep step, see da-4), abstract differential-privacy/k-anonymity theory (da-11), or feature encoding (da-17). Those skills are adjacent; this one owns the generation pipeline and its evaluation. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#when-to-use-this-skill]
1. Why synthetic data
- Privacy / data sharing. Replace a protected dataset with a synthetic surrogate. Gartner projected synthetic data would let organizations avoid 70% of privacy-violation sanctions by 2025, and estimated ~60% of AI/analytics data in 2024 would be synthetic (Gartner 2024; MIT Sloan 2023). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#1-why-synthetic-data]
- Augmentation. Fill gaps, enlarge small datasets, balance rare classes, simulate scarce scenarios (fraud, failure). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#1-why-synthetic-data]
- Testing / dev / CI. Populate non-prod at scale with realistic-but-fake records (Gartner 2024 Data Masking Market Guide). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#1-why-synthetic-data]
- Class rebalancing. Generate minority-class examples to de-skew a classifier. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#1-why-synthetic-data]
2. Tabular synthesis methods
- Tabular is the hard case: mixed types, non-Gaussian/multimodal continuous columns, imbalanced categoricals. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#2-tabular-synthesis-methods]
- Gaussian copula. Transform each column to standard-normal via its CDF, fit a multivariate Gaussian for correlation structure, sample, invert to original margins. Fast, transparent, stable on small data; weak on highly non-linear dependencies (SDV docs). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#2-tabular-synthesis-methods]
- CTGAN (Conditional Tabular GAN). Mode-specific normalization for multimodal continuous columns + conditional generator + training-by-sampling for imbalanced categoricals. Beat Bayesian-network and prior GAN baselines on >=87% of test datasets (Xu et al., NeurIPS 2019). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#2-tabular-synthesis-methods]
- TVAE (Tabular VAE). VAE from the same NeurIPS 2019 paper; strong on smaller datasets, trains more stably than a GAN. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#2-tabular-synthesis-methods]
- CopulaGAN. Hybrid applying the copula transform before GAN modeling (experimental in SDV). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#2-tabular-synthesis-methods]
- CART / sequential synthesis. Non-deep classical approach (R synthpop): synthesize one column at a time conditioned on already-synthesized columns via CART trees. Interpretable, fast, official-statistics default. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#2-tabular-synthesis-methods]
3. Deep generative methods
- GANs. Adversarial generator vs discriminator; high fidelity but training instability and mode collapse. CTGAN is the tabular workhorse. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#3-deep-generative-methods]
- VAEs. Encoder/decoder with latent prior; smoother, more stable, good diversity. TVAE is the tabular instance. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#3-deep-generative-methods]
- Diffusion. Iterative denoising from a Gaussian prior. TabDDPM (ICML 2023) handles numerical + categorical features and beat prior SOTA on several benchmarks; concurrent STaSy, CoDi; later TabSyn, TabDiff (Kotelnikov et al., ICML 2023). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#3-deep-generative-methods]
4. Class imbalance: resampling vs generative
- SMOTE interpolates minority points along segments to k nearest minority neighbors (local info only). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#4-class-imbalance-resampling-vs-generative]
- ADASYN adapts the number of synthetic points per minority sample by density, generating more in harder-to-learn regions. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#4-class-imbalance-resampling-vs-generative]
- Generative (GAN/VAE/diffusion) captures complex non-linear joint structure SMOTE/ADASYN miss; hybrids like SMOTE->GAN refine unrealistic SMOTE points (MDPI Mathematics 2023; SMOTified-GAN 2021). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#4-class-imbalance-resampling-vs-generative]
- Rule of thumb. Low-dim numeric, small budget -> SMOTE/ADASYN. High-cardinality categoricals, non-linear interactions, or privacy need -> generative. SMOTE before the train/test split leaks information; resample inside the CV fold only. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#4-class-imbalance-resampling-vs-generative]
5. Differentially private synthesis
- Plain synthetic data is not automatically private; formal guarantees require DP built into training. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#5-differentially-private-synthesis]
- DP-GAN / DP-SGD GANs. Calibrated noise + gradient clipping on the discriminator; post-processing theorem makes the generator DP. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#5-differentially-private-synthesis]
- PATE-GAN. PATE ensemble of teacher discriminators trains a student via noisy aggregated votes; DP by post-processing, often better utility than DP-GAN at equal epsilon (Jordon, Yoon, van der Schaar, ICLR 2019). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#5-differentially-private-synthesis]
- Marginal-based (PrivBayes, MST). Privately measure low-order marginals (Laplace/Gaussian mechanism), then reconstruct a distribution: a Bayesian network (PrivBayes, Zhang et al. 2017) or a graphical model over a maximum-spanning-tree of marginals (MST). MST won the 2018 NIST DP Synthetic Data Challenge; both strong against membership inference with high utility (McKenna, Miklau, Sheldon 2021; NIST). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#5-differentially-private-synthesis]
- SmartNoise (OpenDP). Microsoft + Harvard toolkit; smartnoise-synth exposes DP synthesizers (MST, PrivBayes, AIM) with a uniform fit()/sample() API (Microsoft 2021). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#5-differentially-private-synthesis]
- Caveat: high epsilon hollows out the guarantee. MST/PrivBayes at high epsilon still leak (arXiv 2402.06699, 2024). DP is only as strong as the epsilon you actually set. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#5-differentially-private-synthesis]
6. Tools & frameworks
- SDV (Synthetic Data Vault) Python one-stop: single-table (GaussianCopula, CTGAN, TVAE, CopulaGAN), multi-table (HMA), sequential (PAR); metadata-driven fit()/sample(). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#6-tools-frameworks]
- SDMetrics SDV evaluation: Quality Report (column shapes, column-pair trends), Diagnostic Report (validity/structure), privacy metrics. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#6-tools-frameworks]
- CTGAN (sdv-dev/ctgan) standalone CTGAN + TVAE. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#6-tools-frameworks]
- synthcity van der Schaar lab; tabular/time-series/survival + privacy benchmarks. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#6-tools-frameworks]
- SmartNoise-synth (OpenDP) DP synthesizers with formal guarantees. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#6-tools-frameworks]
- imbalanced-learn SMOTE/ADASYN variants. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#6-tools-frameworks]
- synthpop (R) CART/sequential synthesis. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#6-tools-frameworks]
7. Evaluation: fidelity vs utility vs privacy
- The three axes trade off (especially privacy vs fidelity/utility under DP). Measure all three. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#7-evaluation-fidelity-vs-utility-vs-privacy]
- Fidelity (looks like real): per-column shape (KS continuous, total-variation/chi-sq categorical), pairwise correlation/contingency similarity, validity/structure (SDMetrics). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#7-evaluation-fidelity-vs-utility-vs-privacy]
- Utility (works like real): TSTR (Train on Synthetic, Test on Real) vs TRTR (train-real-test-real); TSTR ~= TRTR means high utility (AWS ML blog 2023). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#7-evaluation-fidelity-vs-utility-vs-privacy]
- Privacy (attacker recovery): Distance to Closest Record (DCR) flags near-copies; Membership Inference Attack (MIA) tests whether an adversary can tell a record was in training (Frontiers Digital Health 2025). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#7-evaluation-fidelity-vs-utility-vs-privacy]
8. Text & image synthesis (overview)
- Text. Instruction-tuned LLMs generate synthetic text/labels (DataGen, MagPie); quality/diversity and label noise are limits; DP-via-API (Aug-PE/Private Evolution) exists (2023-24). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#8-text-image-synthesis-overview]
- Images. GANs gave limited augmentation benefit; diffusion era (Stable Diffusion, DALL-E, Imagen) sharply improved fidelity and dropped per-image cost (2024-25 reviews). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#8-text-image-synthesis-overview]
9. Regulatory context
- Synthetic data is not automatically anonymous or out of GDPR scope. Fully synthetic data meeting the anonymisation bar escapes GDPR; partially synthetic data usually remains personal data. ICO's March 2025 anonymisation guidance: effective anonymisation is a high bar requiring a documented re-identification-risk assessment (GDPR Local; RPC 2025; NIST). Treat a privacy claim as something you must measure (DCR + MIA, ideally under a DP budget), not something the word "synthetic" grants. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#9-regulatory-context]
Methodology (end-to-end pipeline)
- Define the goal first (privacy share / augmentation / test data / rebalancing) - sets which axis you optimize and which gates you must pass. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
- Profile and build metadata (types, keys, datetime formats, constraints). Bad metadata is the #1 cause of bad output. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
- Split before you fit - hold out a real test set before training so TSTR and MIA are honest. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
- Pick a generator by data + constraint - copula baseline; escalate to CTGAN/TVAE; diffusion (TabDDPM) for highest fidelity; switch to a DP synthesizer (PATE-GAN, MST/PrivBayes via SmartNoise) the moment a formal privacy guarantee is required. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
- Fit, then enforce constraints - apply business rules / valid ranges; reject or post-process invalid rows. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
- Evaluate on all three axes - fidelity, utility (TSTR vs TRTR), privacy (DCR + MIA). For DP, report epsilon. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
- Iterate against the binding constraint - privacy fails: lower epsilon/regularize; utility fails: more capacity/epochs or change family. Expect to trade. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
- Document generator, hyperparameters, epsilon, seed, metrics, and the re-identification-risk assessment (required for any GDPR/anonymisation claim). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#methodology-end-to-end-pipeline]
Practical patterns
- Copula-first ladder; don't start with a GAN. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#practical-patterns]
- Resample inside the fold (never before the split). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#practical-patterns]
- DP by construction, not post-hoc filtering. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#practical-patterns]
- Conditional sampling (CTGAN/SDV) for rare classes. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#practical-patterns]
- Report TSTR next to TRTR; the gap is the signal. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#practical-patterns]
- Keep metadata in version control; regenerate from it. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#practical-patterns]
Anti-patterns
- Calling synthetic data "anonymous" with no DCR + MIA test. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#anti-patterns]
- Evaluating fidelity only (a perfect KS can still memorize records or be useless downstream). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#anti-patterns]
- SMOTE before the split / on test data (leakage). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#anti-patterns]
- Treating high epsilon as "private" (MST/PrivBayes leak at high epsilon). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#anti-patterns]
- GAN by default on small/simple data (mode collapse; copula or TVAE is faster/better). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#anti-patterns]
- Ignoring constraints/keys (synthesizers emit out-of-range values and broken keys unless constrained). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#anti-patterns]
- One synthetic draw as ground truth (sample multiple; metrics vary, especially for GANs). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#anti-patterns]
Troubleshooting
- Mode collapse / low diversity (GAN): switch to TVAE/diffusion; add conditional generator; tune batch size / PacGAN packing. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#troubleshooting]
- Categorical cardinality blows up training: group rare categories or use copula/CART. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#troubleshooting]
- DP output useless (utility floor): epsilon too small or marginals too high-order; raise epsilon within policy, lower marginal order (MST), or swap PATE-GAN <-> marginal-based. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#troubleshooting]
- TSTR far worse than TRTR: fidelity gap in relied-on columns; inspect per-column shape + correlation reports. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#troubleshooting]
- Synthetic rows duplicate real rows (DCR ~0): overfitting; reduce epochs/capacity, add DP, or add training data. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#troubleshooting]
- Invalid rows (out of range, broken keys): add SDV constraints/metadata; post-process and re-validate. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#troubleshooting]
References
- Xu et al. - Modeling Tabular Data using Conditional GAN (CTGAN), NeurIPS 2019. https://proceedings.neurips.cc/paper_files/paper/2019/file/254ed7d2de3b23ab10936522dd547b78-Paper.pdf [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- Jordon, Yoon, van der Schaar - PATE-GAN, ICLR 2019. https://openreview.net/pdf?id=S1zk9iRqF7 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- Kotelnikov et al. - TabDDPM, ICML 2023. https://proceedings.mlr.press/v202/kotelnikov23a/kotelnikov23a.pdf [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- McKenna, Miklau, Sheldon - Winning the NIST Contest: MST, 2021. https://arxiv.org/pdf/2301.08844 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- Zhang et al. - PrivBayes, ACM TODS 2017. https://dl.acm.org/doi/10.1145/3134428 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- SDV docs & synthesizers. https://docs.sdv.dev/sdv | https://github.com/sdv-dev/SDV | https://github.com/sdv-dev/CTGAN [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- SmartNoise / OpenDP. https://opensource.microsoft.com/blog/2021/02/18/create-privacy-preserving-synthetic-data-for-machine-learning-with-smartnoise/ | https://docs.smartnoise.org/synth/index.html [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- NIST Privacy Collaborative Research Cycle. https://pages.nist.gov/privacy_collaborative_research_cycle/pages/techniques.html [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- AWS ML Blog - fidelity/utility/privacy. https://aws.amazon.com/blogs/machine-learning/how-to-evaluate-the-quality-of-the-synthetic-data-measuring-from-the-perspective-of-fidelity-utility-and-privacy/ [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- Frontiers Digital Health - synthetic tabular health eval, 2025. https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1576290/full [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- High Epsilon Vulnerabilities in MST and PrivBayes, 2024. https://arxiv.org/html/2402.06699v1 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- MDPI Mathematics - GANs vs SMOTE, 2023. https://www.mdpi.com/2227-7390/11/16/3605 | SMOTified-GAN 2021. https://arxiv.org/pdf/2108.03235 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- imbalanced-learn SMOTE & ADASYN. https://machinelearningmastery.com/smote-oversampling-for-imbalanced-classification/ [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- Gartner - Safeguarding Privacy with Synthetic Data, 2024. https://www.gartner.com/en/newsroom/press-releases/2024-06-27-safeguarding-privacy-with-synthetic-data [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- ICO anonymisation guidance, RPC 2025. https://www.rpclegal.com/snapshots/data-protection/summer-2025/ico-publishes-new-guidance-on-anonymisation-and-pseudonymisation/ [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- GDPR Local - Synthetic Data Under GDPR. https://gdprlocal.com/synthetic-data-under-gdpr/ [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- Synthetic Data in 2024 review. https://www.timlrx.com/blog/synthetic-data-in-2024-progress-opportunities-challenges/ [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
- synthcity. https://arxiv.org/pdf/2301.07573 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-35-synthetic-data-generation/#references]
Where this helps
- Sharing a dataset externally or across teams when the raw data contains PII that can't leave its original boundary, by generating a synthetic surrogate instead. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Augmenting a too-small or class-imbalanced training set, such as rare fraud cases, where SMOTE/ADASYN's local interpolation isn't capturing the real joint structure. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Populating a realistic test/dev/CI environment at scale without exposing production PII to lower environments. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Deciding whether a synthetic dataset needs formal differential-privacy guarantees, like DP-GAN or PATE-GAN, versus a plain generative model, based on how sensitive the source data is. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Project ideas
- Build a tabular synthesis pipeline starting with a Gaussian copula baseline, and only escalate to CTGAN or a diffusion model like TabDDPM if the copula's fidelity metrics fall short on multimodal or highly imbalanced columns. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Implement a TSTR (Train on Synthetic, Test on Real) evaluation harness that compares a model trained on synthetic data against one trained on the real data, holding out a genuine real test set. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Build a PATE-GAN or DP-SGD synthesis pipeline for a sensitive dataset and validate its privacy guarantee empirically with a membership-inference-attack test, not fidelity metrics alone. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Prototype a hybrid class-imbalance pipeline (SMOTE feeding a GAN, or vice versa) and compare it against plain SMOTE and a pure generative baseline on a rare-event classification task. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Common mistakes
- Calling synthetic data "anonymous" without running a distance-to-closest-record and membership-inference-attack test — fully synthetic data can still leak information about real records. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Evaluating only fidelity, how well the synthetic data statistically resembles the real data, when a perfect fidelity score can coexist with the model having memorized real records or the data being useless for the downstream task. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Running SMOTE or other resampling before the train/test split, or on test data, which leaks information across the split and inflates apparent model performance. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Reaching straight for a GAN when a Gaussian copula would do — copulas are faster, more transparent, and more stable on small data, and should be the first rung of the ladder, not the last resort. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Known issues
- Plain synthetic data is not automatically private; formal guarantees require differential privacy built into the training process itself, not applied as a post-hoc filter. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Fidelity, utility, and privacy trade off against each other, especially under differential privacy — pushing harder on one, say a tighter privacy epsilon, measurably costs on the others. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Mode collapse and low diversity are known GAN failure modes on tabular data, often requiring a switch to TVAE or diffusion-based methods, or tuning batch size and packing to recover. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Synthetic data's regulatory status is not settled by default — fully synthetic data that clears the anonymization bar escapes GDPR, but partially synthetic data usually remains classified as personal data, and that line depends on the specific dataset and generation method. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Where this helps
- Sharing a dataset externally or across teams when the raw data contains PII that can't leave its original boundary, by generating a synthetic surrogate instead. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Augmenting a too-small or class-imbalanced training set, such as rare fraud cases, where SMOTE/ADASYN's local interpolation isn't capturing the real joint structure. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Populating a realistic test/dev/CI environment at scale without exposing production PII to lower environments. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Deciding whether a synthetic dataset needs formal differential-privacy guarantees, like DP-GAN or PATE-GAN, versus a plain generative model, based on how sensitive the source data is. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Project ideas
- Build a tabular synthesis pipeline starting with a Gaussian copula baseline, and only escalate to CTGAN or a diffusion model like TabDDPM if the copula's fidelity metrics fall short on multimodal or highly imbalanced columns. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Implement a TSTR (Train on Synthetic, Test on Real) evaluation harness that compares a model trained on synthetic data against one trained on the real data, holding out a genuine real test set. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Build a PATE-GAN or DP-SGD synthesis pipeline for a sensitive dataset and validate its privacy guarantee empirically with a membership-inference-attack test, not fidelity metrics alone. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Prototype a hybrid class-imbalance pipeline (SMOTE feeding a GAN, or vice versa) and compare it against plain SMOTE and a pure generative baseline on a rare-event classification task. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Common mistakes
- Calling synthetic data "anonymous" without running a distance-to-closest-record and membership-inference-attack test — fully synthetic data can still leak information about real records. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Evaluating only fidelity, how well the synthetic data statistically resembles the real data, when a perfect fidelity score can coexist with the model having memorized real records or the data being useless for the downstream task. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Running SMOTE or other resampling before the train/test split, or on test data, which leaks information across the split and inflates apparent model performance. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Reaching straight for a GAN when a Gaussian copula would do — copulas are faster, more transparent, and more stable on small data, and should be the first rung of the ladder, not the last resort. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Known issues
- Plain synthetic data is not automatically private; formal guarantees require differential privacy built into the training process itself, not applied as a post-hoc filter. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Fidelity, utility, and privacy trade off against each other, especially under differential privacy — pushing harder on one, say a tighter privacy epsilon, measurably costs on the others. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Mode collapse and low diversity are known GAN failure modes on tabular data, often requiring a switch to TVAE or diffusion-based methods, or tuning batch size and packing to recover. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
- Synthetic data's regulatory status is not settled by default — fully synthetic data that clears the anonymization bar escapes GDPR, but partially synthetic data usually remains classified as personal data, and that line depends on the specific dataset and generation method. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Children
- Why synthetic data (privacy, augmentation, testing, rebalancing) (frontier)
- Tabular synthesis methods (Gaussian copula, CTGAN, TVAE, CopulaGAN, CART/sequential) (frontier)
- Deep generative methods (GANs, VAEs, diffusion/TabDDPM) (frontier)
- Class imbalance: SMOTE/ADASYN vs generative (frontier)
- Differentially private synthesis (DP-GAN, PATE-GAN, PrivBayes, MST, SmartNoise) (frontier)
- Tools and frameworks (SDV, SDMetrics, synthcity, synthpop) (frontier)
- Fidelity vs utility vs privacy evaluation (TSTR, DCR, membership inference) (frontier)
- Text and image synthesis overview (frontier)
- Regulatory context (GDPR, ICO, NIST) (frontier)
Frontier under this node: Class imbalance: SMOTE/ADASYN vs generative, Deep generative methods (GANs, VAEs, diffusion/TabDDPM), Differentially private synthesis (DP-GAN, PATE-GAN, PrivBayes, MST, SmartNoise), Fidelity vs utility vs privacy evaluation (TSTR, DCR, membership inference), Regulatory context (GDPR, ICO, NIST), Tabular synthesis methods (Gaussian copula, CTGAN, TVAE, CopulaGAN, CART/sequential), Text and image synthesis overview, Tools and frameworks (SDV, SDMetrics, synthcity, synthpop), Why synthetic data (privacy, augmentation, testing, rebalancing)