How to Generate Synthetic Data with a GAN While Preserving Row Constraints and Outcome Patterns?

[WilsimanEvangelista]

How can I generate synthetic data based on an existing table with 6 columns (including the outcome column), where the outcome can be 0, 1, or 2, using a GAN, and considering the following constraints?

• The cell values must be between 0 and 1 (inclusive), with up to two decimal places.

• The sum of all cells in a row, except the last column, must be exactly 1.

Additionally:

• New rows should follow the same "pattern" as the original rows based on the "outcome" (last column). For instance, if rows with "outcome" equal to 0 have a specific "pattern" in the values of the other columns, the newly generated rows for that "outcome" should preserve that pattern.

• The process should allow generating different amounts of synthetic data for each "outcome," while maintaining the above constraints.

How can I implement this using a GAN in Python? If possible, provide examples of libraries and code to set up and train the GAN to meet these constraints and the requested pattern.

[npatki]

Hi @WilsimanEvangelista, nice to meet you. Have you been able to try using the SDV library with your data already? I think it would be helpful to run through the resources below with your dataset -- as the SDV's synthetic data is designed to meet almost all the points that you have inputted above.

General SDV Resources:

• SDV docs site
• SDV demos
• CTGANSynthesizer docs page -- use GANs to learn patterns from data and recreate it in the synthetic data
• CTGAN demo notebook

Possibly the only pattern that will be difficult for a GAN learn out-of-the-box is this one:

The sum of all cells in a row, except the last column, must be exactly 1.

To achieve this, you can apply constraints

Constraints resources:

• Blog article intro to constraints
• Documentation for constraints
• Demo notebook for adding constraints

If you are able to run through this and have any specific questions, we request that you please share the code and any output(s) you are getting. Thanks.