search

shabbychef / ohenery

Modeling of Ordinal Random Variables via Softmax
GNU Lesser General Public License v3.0
6 stars 0 forks source link

Could we add a variable selection method? #6

Open BenoitLondon opened 11 months ago

BenoitLondon commented 11 months ago

Either a penalty term (lasso L1? or elastic-net) or a step method based on AIC or a "select" argument to remove noisy covariates (high p-values)

I know step/select methods are not ideal for variable selection but that could be helpful and easy to implement.

shabbychef commented 1 month ago

I don't think "variable selection" itself is a viable addon, but we can think about:

  1. how this package could better interact with variable selection packages. I am thinking of anything based on AIC/BIC could be of use.
  2. potentially adding regularization terms to the optimization. I'm afraid we would lose the MLE-ness of the output, so I would want to remove the statistical capabilities of the resultant fits.
  3. adding constraints to the MLE estimation?
BenoitLondon commented 1 month ago

I think elastic net penalty term added to the likelihood is the best option I tried that in my own package and it works as expected, not sure how it fits wih maxLik package (I use optim on nll+penalty term)