[FEATURE] export the propensity score model for predicting new weights

[talgalili]

Currently running the model on many millions of observations is computationally costly.
A better solution would be to run the model a sample of observations (E.g., 10K-50K observations for sample and target). And then export the model, and use it to predict the weights of other observations.

This should probably be done only after issue #30 is resolved.

[talgalili]

As a temporary hack: One option is to use balance to fit weights to a sample (with, say, 20K users).
And then use another model (say, XGboost/RandomForest, or something else) on this sample to predict (based on the same covariates), what the weights are.
With that model, you could then predict the weights to all of your users.

[talgalili]

FYI: @neuralsorcerer interesting issue for you (complex).

[talgalili]

At least for ipw, can think about creating some method using this the ['model']['fit'] output:
https://github.com/facebookresearch/balance/blob/main/balance/weighting_methods/ipw.py#L685

[neuralsorcerer]

At least for ipw, can think about creating some method using this the ['model']['fit'] output: https://github.com/facebookresearch/balance/blob/main/balance/weighting_methods/ipw.py#L685

Yup, need to look into it.