Any way to fix one or more categorical variable value during the data generating stage? about ctgan

[kevinykuo]

Sampling is cheap so you could just sample a bunch of records then filter for the ones you're interested in.

[csala]

Sampling is cheap so you could just sample a bunch of records then filter for the ones you're interested in.

Exactly what @kevinykuo says: right now there is no way to fix variables when sampling, so the only option is to discard the invalid rows and resample until you have as many as you need.

Another option is, is you know what you need in advance, to create a subset of the original dataset with only the rows that contain the value(s) that you need and then fitting a specific CTGAN instance with that subset of rows so all the newly generated ones have those values fixed.

[pawelkuk]

Sampling is cheap so you could just sample a bunch of records then filter for the ones you're interested in.

Exactly what @kevinykuo says: right now there is no way to fix variables when sampling, so the only option is to discard the invalid rows and resample until you have as many as you need.

Another option is, is you know what you need in advance, to create a subset of the original dataset with only the rows that contain the value(s) that you need and then fitting a specific CTGAN instance with that subset of rows so all the newly generated ones have those values fixed.

What about setting condvec by hand while sampling? Shouldn't we then get the desired class most of the time?

[onacrame]

This is exactly the feature I’m looking for. Would be a great addition. More specifically, would like the ability to fix any value, discrete or continuous. I’m in financial services where certain combination of high level parameters in a given month (eg interest rate) we would want to fix but have other variables generated such as credit profile of the customer. We may have a high level plan for lending but the loans that get written will vary significantly due to variation in customer profiles.

I can’t just generate synthetic credit profiles because only certain financial products would accept certain credit profiles and it would be inefficient to build a model for each combination of features I want to treat as fixed.

I would think this could be handled by the conditioning vector?

[csala]

This feature has been covered in the PR #68