Ac/mcnnm

[achenzion]

initial MC commit + refactor to centralize synthetic outcome creation for ease of integration with MC

[apoorvalal]

this looks good (and surprisingly simple!); would be good to parametrize whether to partial out FEs before fitting MC though.

[achenzion]

Ya. Can add the parameters, but got to decide the right way to pull them out. I couldn't come up with an elegant way that nested the intercept inclusion for the other methods, but maybe that is a problem for synth.py. Originally I was thinking maybe that can but upstream, but realized the FE will need to be iterative within each SVT iteration (if I recall correctly) so will need a mcnnm specific implementation.

[apoorvalal]

Thanks for taking a crack at this; I've been down the LLM finetuning/evals mines lately and haven't been able to develop this much.

Also I remember Susan and co's package does the iterative SVT work in C++; might be feasible to do the same via pyensmallen (or even directly use the c++ code in their library via pybind11).

[apoorvalal]

This and this are best factor model libraries I've found in python; could be useful to benchmark against them / potentially wrap them.

[achenzion]

Thanks for taking a crack at this; I've been down the LLM finetuning/evals mines lately and haven't been able to develop this much.

Also I remember Susan and co's package does the iterative SVT work in C++; might be feasible to do the same via pyensmallen (or even directly use the c++ code in their library via pybind11).

Ya, I started going down that rabbit hole. Here is the code explicitly https://github.com/susanathey/MCPanel/blob/master/src/mcnnm_R.cpp

After trying a bit I did some research to try and figure it out, but got a bit stuck. It was suggested that a python binding might not be that much better for the base SVD which makes sense. We'd probably only benefit from binding the whole iterative SVT. Agree worth exploring those alternatives (alt packages AND porting the implementation from that package) for benchmarking.

[achenzion]

@apoorvalal how were you thinking about generalizing the syntax for Synth to MC, Vertical Reg and Horizintal Reg? I pulled out the _get_synthetic which seemed necessary given MC is less "weight" focused, but other comments before this merge? I figure we can connect MC with Synth in the next PR.

[achenzion]

there was a use of synth instead of matching_synth so the matching section wasn't actually using matching estimator. It threw an error when corrected, but I did not address it.

[achenzion]

@apoorvalal

[achenzion]

For now added lambda that performed reasonable on ADH data. Need to add cross validation for lambda selection or consider alternative intelligent selection (printed std of Y above to consider, but did not use)