Synthetic data glossary
Short definitions of the metrics, privacy tests and generator terms used in synthetic data work. Each one links to the post that goes deeper.
TSTR · DCR · NNDR · Membership inference · Differential privacy · Fidelity · Utility · Mode collapse · CTGAN · Copula · SDV
TSTR (train-synthetic, test-real)
TSTR is a utility test: train a model on synthetic data, evaluate it on held-out real data, and compare the score with a model trained on real data. If the synthetic-trained model comes within a few points, the synthetic data carries the signal that matters for modelling.
DCR (distance to closest record)
DCR is a privacy metric. For each synthetic row it measures the distance to the nearest real training row. If synthetic rows are systematically closer to the training data than genuinely new real rows are, the generator is copying records rather than generalising.
NNDR (nearest-neighbor distance ratio)
NNDR is the distance from a synthetic row to its closest real record divided by the distance to the second-closest. A ratio near 1 means the row is not tied to one individual; a very small ratio means it sits almost on top of one person's record. It can catch cases that DCR alone misses.
Membership inference
A membership-inference attack tries to tell, from the synthetic output alone, whether a specific real record was in the generator's training data. Testing it means running the attack yourself on members and held-out non-members; a score near chance (0.5) is the goal, and a materially higher one is evidence of leakage.
Differential privacy
Differential privacy is a formal, mathematical bound on how much any single person's record can change what a generator outputs. It is a guarantee about leakage, not a legal conclusion, but it is strong evidence in a privacy assessment, and the parameters used should be reported with it.
Fidelity
Fidelity is how closely synthetic data matches the statistics of the real data: per-column distributions, means and cardinalities, and also the relationships between columns such as correlations and joint distributions. Good fidelity matches not just each column but how the columns move together.
Utility
Utility is whether synthetic data is still worth using for its intended job, such as training a model, forecasting or analytics. It is measured with tests like TSTR and with the downstream tasks you care about, and it trades off against privacy.
Mode collapse
Mode collapse is a failure of GAN-based generators in which the model produces only a narrow set of the patterns in the training data and misses the rest, so rare categories and minority groups go underrepresented. Comparing per-column and joint distributions against the real data is how it shows up.
CTGAN
CTGAN (Conditional Tabular GAN) is a generative adversarial network designed for tabular data, introduced by Xu et al. at NeurIPS 2019 alongside TVAE. It conditions generation on category values to handle mixed column types and imbalanced categories.
Copula
A copula is a statistical model of how variables depend on each other, separate from each variable's own distribution. Gaussian and vine copulas are used as cheap, fast, explainable synthetic data generators, but a copula sees correlation, not complex nonlinear structure.
SDV (Synthetic Data Vault)
SDV is an open-source Python library for generating synthetic tabular and relational data. It is mature and well documented, but most of the SDV ecosystem moved to the Business Source License in 2025, which restricts production use without a commercial agreement; its SDMetrics evaluation library remains MIT-licensed.