rangertoolkit

Rangertoolkit is intended to be a collection of convenient utilities for working with ranger random forest models. It currently provides the score_oob function to generate out-of-bag curves for regression, classification, and survival forests. It uses the yardstick package for evaluation in order to integrate nicely with tidymodels and tidyverse work flows.

The OOBCurve package is an alternative resource for generating out-of-bag curves for ranger models; however, it is based around the mlr package and does not support survival models.

Installation

You can install the development version of rangertoolkit with devtools.

devtools::install_github("https://github.com/AFC-AI2C/rangertoolkit/")

Examples

Regression

This is an example of evaluating the out-of-bag error for a regression model.

library(rangertoolkit)
library(ggplot2)
set.seed(0)

regression_model <- ranger::ranger(
  mpg ~ .,
  data = mtcars,
  num.trees = 200,
  keep.inbag = TRUE
)
regression_score <- score_oob(
  regression_model,
  mtcars[, -1],
  mtcars[, 1]
)

print(head(regression_score))
#> # A tibble: 6 × 4
#>   num_trees .metric .estimator .estimate
#>       <dbl> <chr>   <chr>          <dbl>
#> 1         1 rmse    standard       3.50 
#> 2         1 rsq     standard       0.405
#> 3         2 rmse    standard       3.91 
#> 4         2 rsq     standard       0.611
#> 5         3 rmse    standard       3.59 
#> 6         3 rsq     standard       0.685

ggplot(regression_score,
       aes(x = num_trees,
           y = .estimate)) +
  geom_point() +
  geom_line() +
  facet_grid(rows = vars(.metric),
             scales = "free_y") +
  ggtitle("Regression Score") +
  xlab("Number of Trees") +
  ylab("Metric") +
  theme_bw()

Classification

This is an example of evaluating the out-of-bag error for a classification model.

classification_model <- ranger::ranger(
  Species ~ .,
  data = iris,
  num.trees = 200,
  keep.inbag = TRUE
)
classification_score <- score_oob(
  classification_model,
  iris[, -5],
  iris[, 5]
)

print(head(classification_score))
#> # A tibble: 6 × 4
#>   num_trees .metric  .estimator .estimate
#>       <dbl> <chr>    <chr>          <dbl>
#> 1         1 accuracy multiclass     0.946
#> 2         1 roc_auc  hand_till      0.959
#> 3         2 accuracy multiclass     0.942
#> 4         2 roc_auc  hand_till      0.956
#> 5         3 accuracy multiclass     0.963
#> 6         3 roc_auc  hand_till      0.978

ggplot(classification_score,
       aes(x = num_trees,
           y = .estimate)) +
  geom_point() +
  geom_line() +
  facet_grid(rows = vars(.metric),
             scales = "free_y") +
  ggtitle("Classification Score") +
  xlab("Number of Trees") +
  ylab("Metric") +
  theme_bw()

Survival

This is an example of evaluating the out-of-bag error for a survival model. This might take some time to run for models with a large number of unique event times.

lung_clean <- na.omit(survival::lung)
survival_model <- ranger::ranger(
  survival::Surv(time, status) ~ .,
  data = lung_clean,
  num.trees = 500,
  keep.inbag=TRUE
)
survival_score <- score_oob(
  survival_model,
  lung_clean[, -c(2, 3)],
  survival::Surv(lung_clean$time, lung_clean$status)
)
#> Progress: 1%
#> Progress: 60%

print(head(survival_score))
#> # A tibble: 6 × 4
#>   num_trees .metric              .estimator .estimate
#>       <dbl> <chr>                <chr>          <dbl>
#> 1         1 concordance_survival standard       0.619
#> 2         2 concordance_survival standard       0.627
#> 3         3 concordance_survival standard       0.621
#> 4         4 concordance_survival standard       0.585
#> 5         5 concordance_survival standard       0.599
#> 6         6 concordance_survival standard       0.590

ggplot(survival_score,
       aes(x = num_trees,
           y = 1 - .estimate)) +
  geom_point() +
  geom_line() +
  ggtitle("Survival Score") +
  xlab("Number of Trees") +
  ylab("1 - C-Index") +
  theme_bw()