The README lists GPU support as an open issue, but the library is built on top of LightGBM which IIRC has GPU support. Is there any room for additional GPU computation?
I was wondering the same thing! It looks like in the setup.py the python package actually has CUDA and LightGBM options built-in already. I didn't dig deep enough to understand how fully implemented they are, but I'm certainly curious.
Not entirely sure I got the point of this paper/library. Is it simply that it improves accuracy, or is the user supposed to be able to (after training) leverage the RW components of the model, say, in a later forward/backward induction?
Since it incorporates mixed effects models my guess is that it allows you to leverage structure in your data to improve the efficiency (in the statistical sense) of the estimation.
That is, you may have a panel of groups. So you might assume observations have in-group and out-group variances (or some other clustering of errors) and furthermore, observations come in time dimensions.
Then, obviously, making use of that information will improve your "fit" in the sense that the regression error will be lower when you estimate across the whole sample, since the dependencies are taken into account.
5 comments
[ 3.1 ms ] story [ 24.9 ms ] threadReally interesting work.
That is, you may have a panel of groups. So you might assume observations have in-group and out-group variances (or some other clustering of errors) and furthermore, observations come in time dimensions.
Then, obviously, making use of that information will improve your "fit" in the sense that the regression error will be lower when you estimate across the whole sample, since the dependencies are taken into account.
Something like that.