Synthetic Data Generation Vocabulary Exercises

1 / 5

A colleague says: "We trained the model on synthetic data because we didn't have access to real patient records."

What makes data "synthetic" by definition?

Synthetic data = artificially generated data that was never directly observed or collected from real people or systems:

Data type	Definition	Derived from real data?
Synthetic data	Algorithmically generated to mirror real-data statistics	No
Anonymised data	Real data with identifiers removed or generalised	Yes
Pseudonymised data	Real data with identifiers replaced by tokens	Yes

Key vocabulary: fully synthetic, partially synthetic, hybrid synthetic, statistical properties, data distribution.

2 / 5

A paper describes a GAN (Generative Adversarial Network) architecture for synthetic tabular data.

Which statement correctly describes the roles of the generator and discriminator?

GAN architecture — adversarial training loop:

Component	Role	Training objective
Generator (G)	Creates synthetic samples from noise	Fool the discriminator
Discriminator (D)	Classifies real vs. generated samples	Correctly identify fakes

Tabular GAN variants: CTGAN (Conditional Tabular GAN), TVAE, CopulaGAN. Key vocabulary: adversarial training, mode collapse, generator loss, discriminator loss, Nash equilibrium.

3 / 5

A data scientist uses a VAE (Variational Autoencoder) to generate synthetic electronic health records.

What are the encoder, latent space, and decoder in a VAE?

VAE architecture:

Component	Function
Encoder	Maps input → latent distribution (μ, σ)
Latent space	Compressed probabilistic representation of data
Decoder	Samples from latent space → reconstructs/generates data

VAE advantage over GAN: more stable training, smooth latent space interpolation. Key vocabulary: reparameterisation trick, KL divergence, ELBO (Evidence Lower Bound), latent vector, reconstruction loss.

4 / 5

A team uses rule-based synthesis to generate synthetic customer records for testing.

Which statement best describes the rule-based approach to synthetic data?

Rule-based synthesis — explicit specification of data constraints:

Approach	How it generates data	Best suited for
Rule-based	Hand-coded constraints + random sampling within rules	Structured test data, compliance use cases
GAN/VAE	Learned generative model from real examples	ML training data augmentation
Simulation-based	Physics/agent-based simulation	Robotics, autonomous vehicles

Key vocabulary: constraint specification, Faker library, referential integrity rules, domain constraints.

5 / 5

An engineer proposes using an LLM to generate synthetic training data for a text classification task.

Which concern is most specific to LLM-generated synthetic data?

LLM-generated synthetic data risks:

Risk	Explanation
Bias amplification	LLM reflects its own training data biases
Model collapse	Downstream models trained on LLM data can degrade if loops repeat
Memorisation leakage	LLM may reproduce near-verbatim text from its training set
Distribution mismatch	Generated data may not match the target domain's real distribution

Key vocabulary: LLM-generated synthetic data, data flywheel, self-distillation, data quality filtering, prompt-driven generation.