In bcjaeger/aorsf: Accelerated Oblique Random Forests

Examples

library(aorsf)
library(magrittr) # for %>%

orsf() is the entry-point of the aorsf package. It can be used to fit classification, regression, and survival forests.

More than one way to grow a forest

You can use orsf(no_fit = TRUE) to make a specification to grow a forest instead of a fitted forest.

orsf_spec <- orsf(pbc_orsf, 
                  formula = time + status ~ . - id,
                  no_fit = TRUE)

orsf_spec

Why would you do this? Two reasons:

For very computational tasks, you may want to check how long it will take to fit the forest before you commit to it:

orsf_spec %>% 
 orsf_update(n_tree = 10000) %>%
 orsf_time_to_train()

If fitting multiple forests, use the blueprint along with orsf_train() and orsf_update() to simplify your code:

orsf_fit <- orsf_train(orsf_spec)
orsf_fit_10 <- orsf_update(orsf_fit, leaf_min_obs = 10)
orsf_fit_20 <- orsf_update(orsf_fit, leaf_min_obs = 20)

orsf_fit$leaf_min_obs
orsf_fit_10$leaf_min_obs
orsf_fit_20$leaf_min_obs

tidymodels

tidymodels includes support for aorsf as a computational engine:

library(tidymodels)
library(censored)
library(yardstick)

pbc_tidy <- pbc_orsf %>% 
 mutate(event_time = Surv(time, status), .before = 1) %>% 
 select(-c(id, time, status)) %>% 
 as_tibble()

split  <- initial_split(pbc_tidy)

orsf_spec <- rand_forest() %>% 
 set_engine("aorsf") %>% 
 set_mode("censored regression")

orsf_fit <- fit(orsf_spec, 
                formula = event_time ~ ., 
                data = training(split))

Prediction with aorsf models at different times is also supported:

time_points <- seq(500, 3000, by = 500)

test_pred <- augment(orsf_fit, 
                     new_data = testing(split), 
                     eval_time = time_points)

brier_scores <- test_pred %>% 
  brier_survival(truth = event_time, .pred)

brier_scores

roc_scores <- test_pred %>% 
  roc_auc_survival(truth = event_time, .pred)

roc_scores