Can TabPFN synthesize data with missing values?

[AwesomeLemon]

Hi,

I'm wondering whether TabPFN is capable of synthesizing data with missing values.

I'm unable to make it work, and since the implementation of synthesis is based on imputation, I'm assuming that it is not capable of generating nans (because imputation replaces all nans)

[spreeni]

I have a similar use case and am interested in this, also I reckon this is a very common use case.

For categorical values it is easy to just introduce another category, but for continuous numerical distributions, I am not sure what the workaround would be.

[LennartPurucker]

Heyho, we do not support this so far. The embedding of TabPFN does not naturally recover nan values, so that we could go from the embedding to nan values. There might be ways to support this, by imputing before and then reversing the imputation. In general, this topic would require more research first.

Once we revisit the code for this, we can add more documentation on this topic to the extensions!

[klemens-floege]

@AwesomeLemon I would close the issue here, feel free to re-open :)