Projection approach to variable selection

Hello! Has anyone tried to implement posterior predictive projection for variable selection in PyMC3?

There is an R-package for this, called projpred. An example of how it can be used is presented here:

Seems pretty cool!

I loved Aki’s project on this, I started to trying to port some of the example to PyMC3 awhile ago and completely forgot about it at some point :sweat_smile:

FWIW, many of the effort had since been put into Arviz, which provide feature like the khat plot. We dont currently have variable selection yet, but maybe that could be one of the roadmap for Bambi.

cc @aloctavodia @ahartikainen


That would be awesome! I’ve found that variable selection in bayesian analysis is a bit more complicated than I expected, having used LASSO (glmnet) previously…