domain adaptation for regression datasets

[mrticker]

I see examples of supervised optimal transfer for domain adaptation for classification. What would you guys recommend when ys are continuous, that is, in a regression setting, assuming that ys in the target domain are unknown?

[rflamary]

Hello I would recommend to use JDOT or deepJDOT :
https://arxiv.org/pdf/1705.08848.pdf
https://arxiv.org/pdf/1803.10081.pdf
JDOT has been used fo both classification and rergression, DeepJDOT has not been tested but should work fine.

They are quite easy to implement with POT you juste need to code the JDOT loss (in torch) and then you can use any optimizer.

I was planning on implementing a clean deepjdot loss in POT but other things got in the way.