Random forest

Machine Learning with Tree-Based Models in R

Sandro Raabe

Data Scientist

Random forest

Suited for high-dimensional data
Easy to use
Out-of-the-box performance
Implemented in a variety of packages: ranger, randomForest
tidymodels interface to these packages: rand_forest() (contained in parsnip package)

Idea

Basic idea (identical to bagging): train trees on bootstrap samples
Key difference: random predictors across trees $\rightarrow$ random forest

Intuition

random forest sketch

Coding: Specify a random forest model

Function name: rand_forest()

Hyperparameters:

mtry: predictors seen at each node, default:
$$\left\lfloor\sqrt\text{num predictors}\right\rfloor$$
trees: number of trees in the forest
min_n: smallest node size allowed

rand_forest(

    mtry = 4,

    trees = 500,

    min_n = 10) %>%

  # Set the mode
  set_mode("classification") %>%

  # Use engine ranger or randomForest
  set_engine("ranger")

Coding: Specify a random forest model

spec <- rand_forest(trees = 100) %>%

            set_mode("classification") %>%

            set_engine("ranger")

Random Forest Model Specification

(classification)


Main Arguments:
  trees = 100

Computational engine: ranger

Training a forest

spec %>% fit(still_customer ~ ., data = customers_train)

parsnip model object

Fit time:  631ms 
Ranger result

Number of trees:                  100 
Sample size:                      9116 
Number of independent variables:  19 
Mtry:                             4 
Target node size:                 10

Variable importance

rand_forest(mode = "classification") %>%
    set_engine("ranger", importance = "impurity") %>%

    fit(still_customer ~ ., data = customers_train) %>%

    vip::vip()

vip plot

Let's plant a random forest!

Machine Learning with Tree-Based Models in R