Skip to contents

Exploring Random Forests with ggRandomForests

John Ehrlinger

2026-03-04

Source: vignettes/ggRandomForests.qmd

The ggRandomForests package extracts tidy data objects from either randomForestSRC or randomForest fits and feeds them into familiar ggplot2 workflows. This vignette highlights the most common objects— gg_error, gg_variable, and gg_vimp—along with a small helper for building balanced conditioning intervals.

Error trajectories with gg_error() 

library(randomForest)
set.seed(42)
rf_iris <- randomForest(Species ~ ., data = iris, ntree = 200, keep.forest = TRUE)
err_df <- ggRandomForests::gg_error(rf_iris, training = TRUE)
head(err_df)
         OOB setosa versicolor  virginica ntree      train
1 0.06349206      0 0.08695652 0.13333333     1 0.02666667
2 0.04255319      0 0.03225806 0.10714286     2 0.02000000
3 0.04761905      0 0.05714286 0.09375000     3 0.02666667
4 0.04098361      0 0.07500000 0.05263158     4 0.02000000
5 0.05426357      0 0.06976744 0.10256410     5 0.01333333
6 0.05970149      0 0.08888889 0.09756098     6 0.01333333

The gg_error() object stores the cumulative OOB error rate for each outcome column plus the ntree counter. When training = TRUE, the function reconstructs the original model frame and appends the in-bag error trajectory (train). Plotting overlays both curves by default:

plot(err_df)

Marginal dependence via gg_variable() 

set.seed(99)
boston <- MASS::Boston
rf_boston <- randomForest(medv ~ ., data = boston, ntree = 150)
var_df <- ggRandomForests::gg_variable(rf_boston)
str(var_df[, c("lstat", "yhat")])
Classes 'gg_variable', 'regression' and 'data.frame':   506 obs. of  2 variables:
 $ lstat: num  4.98 9.14 4.03 2.94 5.33 ...
 $ yhat : num  29.2 22.5 35.1 36.4 33.4 ...

Because the original training data are recovered from the model call, gg_variable() works even when the forest was trained within helper functions or against a subset() expression. The output keeps the raw predictors plus either a continuous yhat column (regression) or per-class probabilities (yhat.<class> for classification). Plotting a single variable is straightforward:

plot(var_df, xvar = "lstat")
`geom_smooth()` using method = 'loess' and formula = 'y ~ x'

Survival forests can request multiple horizons using the time argument; non-OOB predictions are available by setting oob = FALSE.

Variable importance with gg_vimp() 

vimp_df <- ggRandomForests::gg_vimp(rf_boston)
head(vimp_df)
     vars           set      vimp positive
1   lstat IncNodePurity 13004.646     TRUE
2      rm IncNodePurity 11661.671     TRUE
3     dis IncNodePurity  2848.850     TRUE
4   indus IncNodePurity  2751.109     TRUE
5 ptratio IncNodePurity  2697.541     TRUE
6    crim IncNodePurity  2645.701     TRUE
plot(vimp_df)
Warning in ggplot2::geom_bar(ggplot2::aes(y = msr, x = "vars", color = "positive"), : All aesthetics have length 1, but the data has 13 rows.
ℹ Please consider using `annotate()` or provide this layer with data containing
  a single row.

If a randomForest object lacks stored importance scores, gg_vimp() tries to compute them on the fly. When the forest truly cannot provide the information (for example when importance = FALSE and the predictors are no longer accessible), the function emits a warning and returns NA placeholders so plots still render.

Balanced conditioning cuts with quantile_pts() 

rm_breaks <- ggRandomForests::quantile_pts(boston$rm, groups = 6, intervals = TRUE)
rm_groups <- cut(boston$rm, breaks = rm_breaks)
table(rm_groups)
rm_groups
(3.56,5.76] (5.76,5.99] (5.99,6.21] (6.21,6.44] (6.44,6.85] (6.85,8.78]
         85          84          84          85          84          84

The helper wraps stats::quantile() to produce evenly populated strata that drop directly into cut() when building coplots or facet labels.

Next steps