replacing vip::vi() call with internal functionality

[amcim]

The vip package was removed from CRAN and that is problematic for Bioconductor submission. We made the decision to replace the use of the package with our own functionality.

The relevant call to vip was vip::vi() in extractTopFeats(), which returns the features a model found most important for predicting AMR phenotype. Its output is what gets written to the *_top_features.tsv files and plotted by plotTopFeatsVI(). As such, this was replaced with with an internal .viGlmnet().

That call goes through a few functions in vip:

• vi(): wrapper; computes importance, drops NAs, sorts by decreasing importance
• vi_model(): S3 generic that picks the extraction method for the model type
• vi_model.model_fit(): unwraps the parsnip fit to the underlying engine object
• vi_model.glmnet(): pulls the coefficients out of the glmnet fit and returns them as Variable, Importance (absolute coefficient), and Sign (POS/NEG)

This implemntation takes the fitted glmnet engine, drops the intercept, and returns a tibble of Variable, Importance (|coefficient|), and Sign, sorted by decreasing importance.

I also added some tests which check that .viGlmnet() returns what vip::vi() did(the three columns, sorted by decreasing importance, with Importance equal to the absolute coefficient at the minimum lambda), and that multi-class fits return per-class columns instead of erroring.

One thing about this approach is that this currently only works for glmnet, which is the only engine we use currently. If we want to add more models later we will need to expand this functionality.

[eboyer221]

Approving - verified .viGlmnet() against vip's actual source and the math checked out.

[jananiravi]

Hey @eboyer221, could you check these out -- I was going to post these from earlier this afternoon, but it got merged in the meantime.

1. workflowsets and extract...parsnip calls: These two @importFrom workflowsets lines are now likely orphaned — nothing calls extract_fit_parsnip or extract_spec_parsnip after vip::vi() was removed. Will produce R CMD check NOTEs; safe to delete.
2. min(glmnet_fit$lambda): This does not use the tuned penalty. The old vip::vi() used the penalty selected by cross-validation (fit$fit$fit$spec$args$penalty). The multi-class branch does use the tuned penalty, so the two paths are inconsistent. Intentional? (@AbhirupaGhosh)
3. Sign = ifelse(coefs > 0, "POS", "NEG"): Zero coefficients (common with LASSO/elastic net) get labeled "NEG". Suggest ifelse(coefs > 0, "POS", ifelse(coefs < 0, "NEG", NA_character_)) to avoid mislabeling features that were zeroed out. @AbhirupaGhosh may also be important for feature filtering in your other PR.
4. filter(cum_imp < cum_vi_upper ...): Check strict upper inequality; excludes the last feature when prop_vi_top_feats = c(0, 1) (the default), because the final cumsum equals the total. Change < to <= so the full range is inclusive on the right. (if correct, this is a good catch by C!)

[eboyer221]

Thank you for flagging those issues, @jananiravi.
I have opened another PR to address them: #40
@AbhirupaGhosh and I are reviewing your second question together to reach a conclusion about how to resolve that discrepancy with the tuned penalty.