Tune about CTGAN about ctgan HOT 3 CLOSED

Hi there @zealscott 👋 A few suggestions to explore!

Gaussian Copulas
Because GAN's in general are hard to tune, we created a gaussian copula based alternative that's faster, easier to tune, and is more interpretable in general. That synthesizer is part of SDV, which is our framework that's situated one level above our individual model libraries like CTGAN and copulas. You can take it for a spin in just a few lines of code so something to explore!

CTGANSynthesizer
We have a CTGANSynthesizer class in SDV and some people have achieved better model quality and better synthetic data quality just by switching from the CTGAN library to this class in SDV. We do some extra data pre-processing so that can help CTGAN converge on better results!

Tuning CTGAN
Tuning GAN's is a bit of an art. You could use a hyperparameter exploration library like BTB. It's unaffiliated with DataCebo and we're not experts in the library, but something to explore as well.

from ctgan.

Thank for your suggestion @srinify !

from ctgan.

Sure thing @zealscott let us know if these end up working for you!

from ctgan.