randomPlantedForest

randomPlantedForest implements Random Planted Forest (Hiabu, Mammen & Meyer), a tree ensemble whose predictions you can read off directly, without post-hoc explanation methods.

Like a random forest, it averages many tree-based models grown on bootstrap samples. Unlike a random forest, each of these models is a sum of trees that each split on a fixed set of predictors, and max_interaction bounds how many predictors that can be. With max_interaction = 1 the forest is an additive model; with 2 it adds pairwise interactions, and so on. As a result, the fitted model decomposes exactly into an intercept, main effects and interactions up to that order:

$$\hat m(x) = \hat m_0 + \sum_k \hat m_k(x_k) + \sum_{k < l} \hat m_{kl}(x_k, x_l) + \dots$$

Each component can be inspected and plotted on its own, and together they sum to the prediction.

Installation

Install the development version from r-universe with

install.packages("randomPlantedForest", repos = "https://plantedml.r-universe.dev")

or from GitHub with

# install.packages("pak")
pak::pak("PlantedML/randomPlantedForest")

Example

rpf() takes a formula, x/y data or a recipe, and predict() returns a tibble as in tidymodels:

library(randomPlantedForest)

mtcars$cyl <- factor(mtcars$cyl)
rpfit <- rpf(mpg ~ cyl + wt + hp, data = mtcars, ntrees = 25, max_interaction = 2)
rpfit
#> ── Regression Random Planted Forest ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
#> Formula: `mpg ~ cyl + wt + hp`
#> 25 tree families with 30 splits each on 3 predictors, interactions to degree 2.
#> ℹ Forest is not purified.
#> 
#> ── Tree growing 
#>    split_structure: leaves
#>          split_try: 10
#>              t_try: 0.4
#>     max_candidates: 50
#>   split_decay_rate: 0.1
#>      delete_leaves: TRUE
#> 
#> ℹ Fit using 1 thread, also the default for `predict()` and `purify()`.

head(predict(rpfit, new_data = mtcars))
#> # A tibble: 6 × 1
#>   .pred
#>   <dbl>
#> 1  21.4
#> 2  21.0
#> 3  24.4
#> 4  20.9
#> 5  17.7
#> 6  19.1

predict_components() returns the decomposition: one column per main effect and interaction, plus the intercept.

components <- predict_components(rpfit, new_data = mtcars)
head(components$m)
#>          cyl         wt         hp     cyl:wt      cyl:hp        hp:wt
#>        <num>      <num>      <num>      <num>       <num>        <num>
#> 1: 2.8275105  0.2569457  0.2970626  0.3448129  0.06873138  0.006002504
#> 2: 2.8275105 -0.2438784  0.2970626  0.3929089  0.06873138  0.076817727
#> 3: 4.8022694  1.7754646  1.3572422 -0.4610986 -0.41379471 -0.231709254
#> 4: 2.8275105 -0.4727413  0.2970626  0.3538112  0.06873138  0.256989969
#> 5: 0.7661282 -1.0143147 -0.3952336  0.4528133  0.02336632  0.241715900
#> 6: 2.8275105 -1.1525180  0.5397470 -0.4590129 -0.03247695 -0.199217461

The glex package plots these components:

library(glex)
library(ggplot2)
library(patchwork)

(autoplot(components, "wt") + autoplot(components, "hp")) /
  (autoplot(components, "cyl") + autoplot(components, c("wt", "hp")))

The Get started guide covers classification, recipes, saving models and more on interpreting components, and the Bikesharing decomposition article works through a larger example.