Partial Dependence Plots

A general framework for constructing partial dependence (i.e., marginal effect) plots from various types of machine learning models in R.

pdp

Partial dependence plots (PDPs) and individual conditional expectation (ICE) curves for R: visualize how a fitted model's predictions depend on a subset of its features.

Model-agnostic — works with dozens of model classes out of the box (randomForest, ranger, gbm, xgboost, caret, e1071, …) and with any model via a user-supplied prediction function (pred.fun)
PDPs and ICE/c-ICE curves for regression and classification
Lightweight plotting via tinyplot (base R graphics), with lattice displays (3-D surfaces, paneled three-predictor plots) behind plot(..., lattice = TRUE)
Fast: batched predictions (batch.size), parallel execution via foreach, training-data subsampling (frac), and Friedman's exact weighted tree-traversal method for gbm models
Minimal dependencies: imports only base R packages plus lattice and tinyplot

Installation

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

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

Quick start

library(pdp)
library(randomForest)

data(boston)  # ships with pdp
set.seed(101)
rfo <- randomForest(cmedv ~ ., data = boston)

# Partial dependence of cmedv on lstat
pd <- partial(rfo, pred.var = "lstat", train = boston)
plot(pd, rug = TRUE, train = boston)

# Two predictors (false color level plot), restricted to the convex hull
partial(rfo, pred.var = c("lstat", "rm"), chull = TRUE, train = boston,
        plot = TRUE)

# ICE curves colored by another feature
ice <- partial(rfo, pred.var = "rm", ice = TRUE, train = boston)
plot(ice, alpha = 0.1, color.by = "lstat", train = boston)

partial() returns a plain data frame, so results are easy to post-process or plot with any graphics package.

Documentation

Package website — function reference and vignettes (getting started, ICE curves, performance tips)
Greenwell, B. M. (2017). "pdp: An R Package for Constructing Partial Dependence Plots." The R Journal, 9(1), 421–436. doi:10.32614/RJ-2017-016 (citation("pdp"))
For variable importance (which pairs naturally with feature effects), see vip

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.