UDA for Regression Problem

[diliprk]

Hi, thanks for this great research paper. I have a physics based regression problem (~6 input features and 1 response variable) with only 35 data points from the real world lab test results. We have a 1D simulation tool with which we can generate any number of low fidelity artificial data points. Unfortunately the low fidelity artificial synthesised data points are not having the same distribution as the real world lab test results and there are differences. I understand the concept of labels may not apply to a regression problem, but Can we use your UDA technique to make the low fidelity data (unlabeled data) to appear more like the real world lab test (labelled ) data?

[michaelpulsewidth]

Hi, if you only have 35 data points in the real world domain, you might need to design some prior knowledge to filter out synthesised data points. If you have more data from the real world domain, style transfer might be helpful if you are trying to transform the synthetic data to real data. Take a look at CycleGAN: https://junyanz.github.io/CycleGAN/

To use UDA on regression problem, you can simply replace the KL loss to MSE.