The tidysynthesis R package
summary
Describes creating fully synthetic tax data (SynPUF and a supplemental SynPUF) that mirror confidential files while protecting privacy, using a sequential approach that fits one predictive model per variable and handles >200 variables and hundreds of thousands of records. Implements tidysynthesis built on tidymodels (parsnip, recipes) to run many models, sample from terminal nodes (sample_rpart), add noise, enforce logical constraints, generate replicates in parallel, and compress a workflow that otherwise would be hundreds of lines into ~70 lines.
AI-written summary · about
chapters (9)
- 0:00The problem of data access
- 1:23Value of administrative data
- 2:31Challenges of traditional access
- 4:26Introduction to synthetic data
- 6:00The tidy synthesis framework
- 7:28Demonstrating data synthesis
- 10:27Running the synthesis process
- 12:01Validating synthetic results
- 13:33Conclusion and call to action
related talks
18mtidymodels: Adventures in Rewriting a Modeling Pipeline
22mModelling, Tidyverse Way
14mLearning the Ropes of Synthetic Data
21mEmpowering Reproducible Finance through Tidy Finance with R and Python
57mGood Practices for Applied Machine Learning
29mDo I need to be Dr. Frankenstein to create real-ish synthetic data?
51mProgramming in the Tidyverse
35mTowards Interpretable Data Science