glmnet (LASSO) learners capable of determining variable importance

[shudras]

LASSO implemented in glmnet and available in mlr3 (e.g. regr.glmnet) is a learner capable of extracting variable importance from the fitted model, but not yet in the mlr3learners package.

[pat-s]

Thanks.

We are tracking this in {mlr3learners} already: mlr-org/mlr3learners#28

[mllg]

There is no importance for glmnet learners, but you can extracted the features with non-zero coefficients with $selected_features(). For filtering, use FilterSelectedFeatures() (flt("selected_features")).

[albersonmiranda]

@mllg , would you mind providing reprex of selected_features method? I'm having a real hard time

[mllg]

The code is implemented in the learners, e.g. https://github.com/mlr-org/mlr3learners/blob/main/R/LearnerRegrGlmnet.R#L97 which calls functions from https://github.com/mlr-org/mlr3learners/blob/main/R/helpers_glmnet.R.

There are a lot of if-else blocks to cover all ways to call the learner which all yield slightly different objects. One way to extract the selected features is to predict the non-zero betas:

  nonzero = predict(self$model, type = "nonzero", s = lambda)

Hope this helps.