ggRandomForests: Visually Exploring Random Forests

ggRandomForests provides ggplot2-based diagnostic and exploration plots for random forests fit with randomForestSRC (>= 3.4.0) or randomForest. It keeps the data step apart from the figure step, so you can inspect, save, or reuse the tidy object on its own.

It also covers varPro (>= 3.1.0), which reaches the same questions by a different route: variable selection built from rules rather than permutation, importance for one observation rather than the whole fit, and dependency and signal detection on unsupervised fits. Eight of the nineteen gg_* families read a varPro object.

Listed in the ggplot2 extensions gallery.

Installation

# CRAN (stable)
install.packages("ggRandomForests")

# Development version from GitHub
# install.packages("remotes")
remotes::install_github("ehrlinger/ggRandomForests")

Quick start

library(randomForestSRC)
library(ggRandomForests)

# 1. Fit a forest (regression)
rf <- rfsrc(medv ~ ., data = MASS::Boston, importance = TRUE)

# 2. Check convergence: did the forest grow enough trees?
plot(gg_error(rf))

# 3. Rank predictors by importance
plot(gg_vimp(rf))

# 4. Marginal dependence for top variables
gg_v <- gg_variable(rf)
plot(gg_v, xvar = "lstat")
plot(gg_v, xvar = rf$xvar.names, panel = TRUE, se = FALSE)

# 5. Partial dependence for a single predictor
pv <- plot.variable(rf, xvar.names = "lstat", partial = TRUE, show.plots = FALSE)
pd <- gg_partial(pv)
plot(pd)

For survival forests, see the package vignette:

vignette("ggRandomForests")

For variable importance with varPro — partial dependence, importance z-scores, beta importance, individual/local importance, and isolation forests — see the dedicated vignette:

vignette("varpro", package = "ggRandomForests")

The unsupervised varPro tools — gg_udependent(), gg_beta_uvarpro(), and gg_sdependent(), which read structure off a uvarpro() fit with no outcome — have their own short vignette:

vignette("uvarpro", package = "ggRandomForests")

Function reference

Grouped by what you are trying to look at. The first column is the function you call, the second is what you hand it.

The forest itself

Partial dependence

Survival

Classification and ROC

varPro — variable priority

These read a varPro fit rather than a forest. varpro() is the supervised fit; uvarpro() is the unsupervised one, which needs no outcome.

SHAP

Each gg_* function has a matching plot() S3 method that hands back a single plottable object: a ggplot you extend with +, or a patchwork composite for the multi-panel methods. Every gg_* object also has print() and summary() methods: print() shows a short header at the REPL rather than dumping every row (use head() when you want the rows), and summary() gives you a diagnostics object you can print or keep.

Why ggRandomForests?

The package is built on one decision: keep the data step and the figure step apart. The gg_* functions pull a tidy data object out of the forest; the plot() methods turn that object into a ggplot2 figure. Two things follow from that split.

First, the data object stands on its own. It carries everything its plot needs, so you can save it, inspect it, or come back to it later without keeping the original forest — which can be large — in memory.

Second, you are never locked into the default figure. Because a plot() method returns a single plottable object (a ggplot, or a patchwork composite for the multi-panel methods), you can add layers, swap scales, or apply a theme; and if the default is not what you want, you can ignore it entirely and build the figure from the tidy data yourself.

Recent changes

See NEWS.md for the full changelog. Recent highlights:

• v3.5.1 gg_roc() on an rfsrc forest now honors the documented which_outcome = 0, which had been returning an unusable two-row object; gg_partial_rfsrc() rejects a non-forest with a real error instead of “argument is of length zero”. Also a test-only fix for the gcc-UBSAN report filed against 3.5.0.
• v3.5.0 varPro fixes: plot.gg_varpro() no longer draws a phantom “NA” category, gg_partial_varpro() warns when you name a variable the fit cannot reach, and scale = "chf" now honors xvar.names instead of computing every variable. Vignette figures render with ragg, which cut the source tarball from 4.7 MB to 2.3 MB.
• v3.4.0 Unsupervised varPro wrappers (gg_beta_uvarpro(), gg_sdependent()) with their own vignette; gg_partial_rfsrc() now handles factor predictors correctly.
• v3.3.0 varPro partial plots default to interpretable scales — probability for classification, survival S(τ) for survival.
• v3.1.0 varPro integration: release-rule importance, partial dependence, local importance, anomaly scores, and the dependency graph.

References

Breiman, L. (2001). Random forests, Machine Learning, 45:5–32.

Ishwaran H. and Kogalur U.B. randomForestSRC: Random Forests for Survival, Regression and Classification. R package version >= 3.4.0. https://cran.r-project.org/package=randomForestSRC

Ishwaran H. and Kogalur U.B. (2007). Random survival forests for R. R News 7(2), 25–31.

Ishwaran H., Kogalur U.B., Blackstone E.H. and Lauer M.S. (2008). Random survival forests. Ann. Appl. Statist. 2(3), 841–860.

Liaw A. and Wiener M. (2002). Classification and Regression by randomForest. R News 2(3), 18–22.

Wickham H. (2009). ggplot2: Elegant Graphics for Data Analysis. Springer New York.

mirror server hosted at Truenetwork, Russian Federation.