fastshap

fastshap computes fast, approximate Shapley values for any supervised learning model in R, using ideas from game theory to attribute a prediction to its input features.

• Model-agnostic — works with any fitted model via a user-supplied prediction function (pred_wrapper)
• Exact Shapley values for linear models (lm()), and Tree SHAP for xgboost and lightgbm (exact = TRUE)
• Main/total/interaction effect decomposition (explain_effects()) — a deterministic, Monte-Carlo-free diagnostic that is exact for models with at most pairwise interactions
• Fast: batched predictions (few large pred_wrapper() calls rather than many small ones), an efficient C++ (Rcpp/RcppArmadillo) core, and optional parallel execution across features via foreach
• Plays nicely with shapviz for waterfall, force, importance, and dependence plots

Installation

fastshap 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("fastshap", repos = c("https://bgreenwell.r-universe.dev", "https://cloud.r-project.org"))

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

Quick start

library(fastshap)

# Fit a projection pursuit regression model
fit <- ppr(mpg ~ ., data = mtcars, nterms = 5)

# Prediction wrapper: takes `object` and `newdata`, returns a numeric vector
pfun <- function(object, newdata) predict(object, newdata = newdata)

# Approximate Shapley values via Monte Carlo simulation
set.seed(101)
shap <- explain(fit, X = subset(mtcars, select = -mpg), nsim = 50,
                pred_wrapper = pfun)
head(shap)

# Deterministic main/total/interaction effect decomposition (no simulation)
eff <- explain_effects(fit, X = subset(mtcars, select = -mpg),
                       pred_wrapper = pfun)
head(eff$shapley_values)

Documentation

• Package website — function reference and the introductory vignette
• Štrumbelj, E., and Kononenko, I. (2014). "Explaining prediction models and individual predictions with feature contributions." Knowledge and Information Systems, 41(3), 647–665.
• Lundberg, S. M., et al. (2020). "From local explanations to global understanding with explainable AI for trees." Nature Machine Intelligence, 2(1), 56–67.

Development

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.