Brandon Greenwell

Would be nice to include options for including surrogates and primary splitters (caret does this and the code used in rpart is pretty straight forward).

Would be convenient to be able to plot all VI scores using something like ```r vip(mod, scale = TRUE, num_features = "all") ``` ```r vip(mod, scale = TRUE, num_features =...

https://arxiv.org/pdf/2006.04750.pdf

https://stackoverflow.com/questions/61606337/computing-importance-measure-using-vip-package-on-a-parsnip-model/62064736#62064736 The `Sign` column should not be `POS` or `NEG` for zero coefficients (i.e., in **glmnet** models).

`vi()` could return a tibble by default, if installed), whereas the`vi_*()` family of functions should return a data frame.

Example from **nnet** package: ```r library(MASS) example(birthwt) bwt.mu

"...This option can only for the permutation-based..."

TODO: - [ ] Should accept `type = c("relative_importance", "scaled_importance", "percentage")`. - [ ] Update `"type"` attribute (for plotting with `include_type = TRUE`).

For example, provide default prediction wrappers for easier permutation importance for keras, xgboost, catboost, lightgbm, and other models. @bradleyboehmke