Synthetic Data Generation Types Vocabulary Exercises

1 / 5

A ML engineer says: "We used a GAN-based approach to synthesise the tabular financial data."

Which statement best describes what distinguishes GAN-based synthesis from rule-based synthesis?

GAN-based vs. rule-based synthesis:

Approach	How it works	Requires real data?
GAN-based	Generator + discriminator compete; learns real distribution	Yes (training set)
Rule-based	Hand-coded constraints, random sampling within rules	No
Agent-based	Simulated agents interact to produce emergent data	No

Key vocabulary: adversarial training, CTGAN, mode collapse, discriminator, generator, domain constraints.

2 / 5

A research paper describes an agent-based simulation used to generate synthetic mobility data for an urban traffic model.

What is the defining characteristic of agent-based synthetic data generation?

Agent-based simulation for synthetic data:

Concept	Explanation
Agent	Software entity with state, behaviour rules, and ability to interact
Emergent behaviour	Complex patterns arising from simple agent interactions
Use cases	Traffic, epidemiology, financial markets, urban planning

Key vocabulary: agent-based modelling (ABM), emergent phenomena, micro-simulation, calibration, scenario generation.

3 / 5

A privacy engineer explains: "We applied differential privacy with an epsilon of 0.5 when training the generative model."

What does the term "epsilon budget" mean in the context of differentially private synthetic data generation?

Epsilon (ε) — the privacy budget:

ε value	Privacy level	Utility impact
ε → 0	Very strong privacy	High noise, low utility
ε = 1	Strong privacy	Moderate utility
ε = 10	Weaker privacy	Higher utility

Key vocabulary: epsilon budget, privacy-utility trade-off, DP-SGD, Gaussian mechanism, composition theorem, privacy accountant.

4 / 5

During a sprint review, a team discusses whether to use GAN-based, rule-based, or agent-based synthesis for generating synthetic e-commerce transaction data.

Which factor most strongly favours choosing rule-based synthesis in this scenario?

Choosing the right synthesis approach:

Scenario	Best approach	Why
Need referential integrity, no real data	Rule-based	Explicit constraints guaranteed by design
Learn statistical correlations from real data	GAN/VAE	Learns distribution from examples
Model agent interactions over time	Agent-based	Captures emergent dynamics

Key vocabulary: referential integrity, constraint specification, Faker, data contract, schema-driven generation.

5 / 5

A team applies differentially private synthetic data generation using DP-SGD during GAN training.

A stakeholder asks: "If we lower epsilon from 10 to 1, what happens to the synthetic data?" Which answer is most accurate?

Lowering epsilon (stronger DP) effects:

Effect	Explanation
Privacy improves	Smaller ε = harder to distinguish individuals in real data
Noise increases	DP-SGD clips gradients and adds Gaussian noise more aggressively
Fidelity decreases	Model learns a blurrier, less accurate data distribution
Utility decreases	Downstream tasks trained on noisy synthetic data may perform worse

Key vocabulary: DP-SGD, gradient clipping, Gaussian noise, epsilon budget, privacy-utility trade-off, sensitivity.