Oblique Random Forest, using aorsf
In tidypredict: Run Predictions Inside the Database

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(dplyr)
library(tidypredict)
library(parsnip)
library(aorsf)
set.seed(100)

| Function |Works| |---------------------------------------------------------------|-----| |tidypredict_fit(), tidypredict_sql(), parse_model() | ✔ | |tidypredict_to_column() | ✗ | |tidypredict_test() | ✔ | |tidypredict_interval(), tidypredict_sql_interval() | ✗ | |parsnip | ✔ |

Only regression models with numeric predictors are supported. Classification requires a voting mechanism that cannot be expressed as a single formula, and oblique splits on categorical predictors operate on an internal encoding that is not reproduced here.

How it works

Here is a simple orsf() model using the mtcars dataset:

library(dplyr)
library(tidypredict)
library(aorsf)

model <- orsf(mtcars, mpg ~ ., n_tree = 5)

Under the hood

Unlike axis-aligned forests, each split in an oblique random forest is a linear combination of (standardized) predictors compared against a cutpoint. tidypredict reads the trees stored in the fitted forest, folds the forest's centering and scaling into the split coefficients, and turns each tree into a nested dplyr::case_when() statement. The trees are then averaged.

tidypredict_fit(model)

From there, the Tidy Eval formula can be used anywhere it can be evaluated. tidypredict provides three paths:

Use directly inside dplyr, mutate(mtcars, !! tidypredict_fit(model))
Use tidypredict_to_column(model) to add it to a piped command set
Use tidypredict_sql(model) to retrieve the SQL statement

A note on split boundaries

aorsf uses observed linear-combination values from the training data as split cutpoints. A training row can therefore land exactly on a split boundary, where floating-point differences between aorsf's internal traversal and the generated formula may send it down a different branch. This affects only rows that coincide with a training cutpoint; on new data the formula reproduces predict() exactly.

parsnip

tidypredict also supports aorsf model objects fitted via the parsnip package (using the bonsai extension).

library(parsnip)
library(bonsai)

parsnip_model <- rand_forest(mode = "regression", trees = 5) %>%
  set_engine("aorsf") %>%
  fit(mpg ~ ., data = mtcars)

tidypredict_fit(parsnip_model)