Hello Joshua. Thanks for responding and apologies for a late reply. I am not sure of the version that we have used earlier but it was prior to making some changes in Conditional Generator.

Is it simply that you initialize CTGAN with a high epsilon, and the call to fit hangs infinitely?
It is the call to fit that hangs infinitely. We are using it with PytorchDPSynthesizer (as per the documentation). But the same works for MWEM.

Or is there some other message?
No, there is no other message in regards to this

What leads you to infer there is an issue with category_epsilon_percent?
We used a brute method to at least make the synthesiser run. So, we looked at the source code and started iterating if any of them gets executed.
class PATECTGAN(CTGANSynthesizer):
def init(
self,
embedding_dim=128,
generator_dim=(256, 256),
discriminator_dim=(256, 256),
generator_lr=2e-4,
generator_decay=1e-6,
discriminator_lr=2e-4,
discriminator_decay=1e-6,
batch_size=500,
discriminator_steps=1,
log_frequency=False,
verbose=False,
epochs=300,
pac=1,
cuda=True,
epsilon=1,
binary=False,
regularization=None,
loss="cross_entropy",
teacher_iters=5,
student_iters=5,
sample_per_teacher=1000,
delta=None,
noise_multiplier=1e-3,
preprocessor_eps=1,
moments_order=100,
category_epsilon_pct=0.1,
)
We went one-by-one while keeping others constant and the default value of 0.1 (for category_epsilon_pct) was not working for higher epsilons and when we increased this, we saw that the code was working.

Also, can you confirm that A) your data are categorical, and B) roughly how many categorical columns exist in the source data?
Yes, we converted every column into categorical and these synthesisers work for us at lower epsilons. Approximately, on an average, we feed in 15 categorical columns

You are also correct that in general, lower epsilon should be less utility with higher privacy. If that pattern isn't holding for you, there may be a bug. Is this a separate issue from the issue with large epsilon?
Yes. The pattern was holding for us previously when we used our own metrics, but now we are using sdvmetrics to assess privacy gain and accuracy. At times, this was holding for some of the datasets but it was not holding for some as well! I expect this pattern to be data-agnostic. We started assessing synthetic datasets, from eps=0.2 to 1000 to see how the pattern holds and it was behaving weirdly for GANs and for some datasets.

is it the case that large epsilon has unexpected behavior in one way, such as hanging infinitely, and smaller epsilon (below what range?)work fine, but accuracy doesn't seem to be impacted by epsilon?
True, it was hanging infinitely for higher epsilons. We are not even able to run the codes with epsilon = 5.0. Accuracy seems to be increasing for most of the datasets across epsilons. This is particularly visible in MWEM. When it comes to GANs, they are not showing monotonous behaviour in most of the cases.

Just wanted to let you know thatI have also started working on the new version and it seems that it is still failing at higher epsilons. Will let you know if there is something we realise!!

Thank you for your support, Joshua. Have a great day!

Thanks,
Dileep

Hi Joshua,

Please let me know if you need more clarification on this!