Variable importance plots (VIPs) for R: quantify and visualize which features drive a fitted model's predictions.

• One interface, many models — vi() and vip() work with dozens of model classes out of the box (randomForest, ranger, gbm, xgboost, lightgbm, glmnet, caret, tidymodels, mlr3, …)
• Model-specific and model-agnostic methods — native importance measures (method = "model"), permutation importance (method = "permute"), Shapley-based importance (method = "shap"), and variance-based importance (method = "firm")
• Works with any model — the model-agnostic methods only require a user-supplied prediction wrapper
• Lightweight plotting via tinyplot (base R graphics)
• Minimal dependencies — imports only base R packages plus tinyplot (itself dependency-free)

vip is no longer available on CRAN due to CRAN's stringent and ever-changing policies. It is now hosted on r-universe, which provides a reliable alternative for distributing R packages.

# Latest stable release (recommended)
install.packages("vip", repos = c("https://bgreenwell.r-universe.dev", "https://cloud.r-project.org"))

# Or with pak
pak::pak("bgreenwell/vip@main")  # latest stable release
pak::pak("bgreenwell/vip")       # development version (devel branch)

CRAN will archive the last CRAN release of vip (0.4.6) on 2026-07-13; all future releases live on r-universe (see above). vip 0.5.0 is also a breaking release — the main changes if you're coming from 0.4.x:

• Plots are base R graphics now, not ggplot2. vip() (and the new plot() method for "vi" objects) draws via tinyplot and invisibly returns the "vi" object, so vip(fit) + theme_bw() no longer works. Style plots with graphical parameters instead — plot(vi(fit), type = "point", col = "red") — or build your own ggplot from the tidy data frame that vi() returns.
• vi() returns a plain data frame (class "vi"), not a tibble; call tibble::as_tibble() on the result if you need one.
• aesthetics is deprecated in favor of plot_args in vip(); the ggplot2-era mapping argument is ignored with a warning; geom = "col" still works as an alias for geom = "bar".
• For package maintainers depending on vip: keep vip in Suggests, use it conditionally (requireNamespace("vip", quietly = TRUE)), and declare Additional_repositories: https://bgreenwell.r-universe.dev in your DESCRIPTION — CRAN accepts this for Suggests-level dependencies.

See NEWS.md for the complete list.

library(vip)

# Simulate Friedman 1 benchmark data; only x1-x5 are important!
trn <- gen_friedman(500, seed = 101)  # ?vip::gen_friedman

# Model-specific importance from a random forest
library(ranger)
set.seed(101)
rfo <- ranger(y ~ ., data = trn, importance = "impurity")
vi(rfo)    # data frame of variable importance scores
vip(rfo)   # variable importance plot

# Permutation importance works for ANY model; just supply a prediction wrapper
pfun <- function(object, newdata) predict(object, data = newdata)$predictions
set.seed(102)
vis <- vi(rfo, method = "permute", train = trn, target = "y", metric = "rmse",
          pred_wrapper = pfun, nsim = 10)

# "vi" objects have a plot() method; additional arguments are passed on to
# tinyplot::tinyplot()
plot(vis, type = "boxplot", all_permutations = TRUE, jitter = TRUE,
     fill = "grey90")

vi() returns a tidy data frame of importance scores, so results are easy to post-process or plot with any graphics package.

• Package website — function reference and the introductory vignette
• Greenwell, B. M., and Boehmke, B. C. (2020). "Variable Importance Plots—An Introduction to the vip Package." The R Journal, 12(1), 343–366. doi:10.32614/RJ-2020-013 (citation("vip"))
• For visualizing feature effects (which pair naturally with variable importance), see pdp

Development happens on the devel branch (the repository default); main holds stable releases, which is what r-universe builds and the website documents. Please open pull requests against devel and report bugs via the issue tracker.