SESSION 02Draft
Methods for generating synthetic data
29 slides in three parts, with speaker notes.
Download slides PPTX01Slides
Downloads
Session slides
Both decks are drafts and include speaker notes.
SESSION 01Draft
Fundamentals of synthetic data: contributed slides
Six optional slides for Session 1.
Download slides PPTX02Software
Tools
Open-source tools used in the course
Start with one tool, run it on a small dataset, then evaluate the output before trusting it.
Synthetic Data Vault (SDV)
Python library with Gaussian copula, CTGAN, TVAE and evaluation reports.
synthpop
R package for sequential CART and parametric synthesis.
simstudy
R package for simulating study and trial data from assumptions.
Synthea
Open-source generator of synthetic patient records.
03Glossary
Key terms
The words you will hear today
Fifteen terms, in the order they come up in the session.
- Synthetic data
- Generated data that reproduces the patterns of real data without being a record of any real person.
- Marginal distribution
- The shape of a single variable on its own, such as the spread of ages.
- Copula
- A way of joining separate variable shapes together with a chosen dependence structure.
- CART
- Classification and regression trees; used in synthpop to generate each variable from those before it.
- GAN
- Generative adversarial network: a generator and a discriminator trained against each other.
- VAE
- Variational autoencoder: compresses records into a small summary and learns to rebuild them.
- Diffusion model
- Learns to reverse a step-by-step noising process to create new records.
- Mode collapse
- When a generator produces only common patterns and misses rare ones.
- Fidelity
- How closely synthetic data resemble the real data.
- Utility
- Whether analyses on synthetic data give the same answers as on real data.
- TSTR
- Train on synthetic, test on real: a utility check for prediction models.
- Distance to closest record
- How near each synthetic record is to its nearest real record; very small distances suggest copies.
- Membership inference
- An attack that tries to tell whether a specific person was in the training data.
- Differential privacy
- A mathematical guarantee that limits how much any one person can influence the output.
- Censoring
- In survival data, when a patient's event time is not observed, for example because follow-up ended.