vip: Variable Importance Plots

Make your ML models more interpretable with beautiful variable importance plots

Overview

vip provides a unified framework for constructing variable importance plots from virtually any machine learning model in R. Instead of juggling different importance() functions across packages, vip gives you one consistent interface for interpretable ML.

Key features

• Universal interface that works with 40+ model types from different packages
• Multiple methods: model-specific, permutation, SHAP, and variance-based importance
• Publication-ready visualizations with ggplot2
• Optimized algorithms with parallel processing support
• Easy to add support for new model types
• Comprehensive guides and academic backing

Installation

# Install from CRAN (stable)
install.packages("vip")

# Install development version (latest features)
# install.packages("pak")
pak::pak("koalaverse/vip")

Quick start

library(vip)
library(randomForest)

# Fit a model
model <- randomForest(Species ~ ., data = iris)

# Get importance scores
vi_scores <- vi(model)
print(vi_scores)
#> # A tibble: 4 × 2
#>   Variable     Importance
#>   <chr>             <dbl>
#> 1 Petal.Length      32.4 
#> 2 Petal.Width       31.3 
#> 3 Sepal.Length       9.51
#> 4 Sepal.Width        6.75

# Create a beautiful plot
vip(model)

Supported methods

Supported models

Tree-based models

• randomForest • ranger • xgboost • lightgbm • gbm • C50 • Cubist • rpart • party • partykit

Linear models

• glmnet • earth (MARS) • Base R (lm, glm)

Neural networks

• nnet • neuralnet • h2o • RSNNS

Meta-frameworks

• caret • tidymodels • parsnip • workflows • mlr • mlr3 • sparklyr

Specialized models

• pls • mixOmics (Bioconductor) • And many more...

Advanced examples

Permutation importance with custom metrics

library(ranger)

# Fit model
rf_model <- ranger(mpg ~ ., data = mtcars, importance = "none")

# Permutation importance with custom metric
vi_perm <- vi(
  rf_model, 
  method = "permute",
  train = mtcars,
  target = "mpg",
  metric = "rmse",
  nsim = 50,        # 50 permutations for stability
  parallel = TRUE   # Speed up with parallel processing
)

# Create enhanced plot
vip(vi_perm, num_features = 10, geom = "point") +
  labs(title = "Permutation-based Variable Importance",
       subtitle = "RMSE metric, 50 permutations") +
  theme_minimal()

SHAP values for detailed attribution

library(xgboost)

# Prepare data
X <- data.matrix(subset(mtcars, select = -mpg))
y <- mtcars$mpg

# Fit XGBoost model
xgb_model <- xgboost(data = X, label = y, nrounds = 100, verbose = 0)

# SHAP-based importance
vi_shap <- vi(
  xgb_model, 
  method = "shap",
  train = X,
  nsim = 30
)

# Beautiful SHAP plot
vip(vi_shap, geom = "col", aesthetics = list(fill = "steelblue", alpha = 0.8)) +
  labs(title = "SHAP-based Variable Importance") +
  theme_light()

Contributing

We welcome contributions! Here's how to get involved:

Development setup

# Clone the repo
git clone https://github.com/koalaverse/vip.git
cd vip

# Open in RStudio or your favorite editor

Testing framework

We use tinytest for lightweight, reliable testing:

# Run all tests
tinytest::test_package("vip")

# Test specific functionality
tinytest::run_test_file("inst/tinytest/test_vip.R")

Development workflow

1. Check issues: Look for good first issues
2. Create branch: git checkout -b feature/awesome-feature
3. Write tests: Follow TDD principles
4. Run checks: R CMD check and tests
5. Submit PR: With clear description

Adding model support

Adding support for new models is straightforward:

# Add S3 method to R/vi_model.R
vi_model.your_model_class <- function(object, ...) {
  # Extract importance from your model
  importance_scores <- your_model_importance_function(object)
  
  # Return as tibble
  tibble::tibble(
    Variable = names(importance_scores),
    Importance = as.numeric(importance_scores)
  )
}

Getting help

• Bug reports: GitHub Issues

Citation

If you use vip in your research, please cite:

@article{greenwell2020variable,
  title={Variable Importance Plots—An Introduction to the vip Package},
  author={Greenwell, Brandon M and Boehmke, Bradley C},
  journal={The R Journal},
  volume={12},
  number={1},
  pages={343--366},
  year={2020},
  doi={10.32614/RJ-2020-013}
}

License

GPL (>= 2) © Brandon M. Greenwell, Brad Boehmke

Built by the koalaverse team