Créer et optimiser un modèle de random forest

Machine Learning dans le tidyverse

Dmitriy (Dima) Gorenshteyn

Lead Data Scientist, Memorial Sloan Kettering Cancer Center

Performance en validation croisée

Modèle de régression linéaire

MAE de validation :

1,5 ans

Un autre modèle

Atouts de la random forest

Gère les relations non linéaires
Gère les interactions

Outils de base pour la random forest

Modèle

rf_model <- ranger(formula = ___, data = ___, seed = ___)

Prédiction

prediction <- predict(rf_model, new_data)$predictions

Construire des modèles random forest basiques

library(ranger)
cv_models_rf <- cv_data %>% 
 mutate(model = map(train, ~ranger(formula = life_expectancy~., 
                                    data = .x, seed = 42)))

cv_prep_rf <- cv_models_rf %>% 
 mutate(validate_predicted = map2(model, validate, 
                                  ~predict(.x, .y)$predictions))

Hyperparamètres de ranger

Modèle

rf_model <- ranger(formula, data, seed, mtry, num.trees)

Hyperparamètres

nom	plage	valeur par défaut
mtry	$1:number\ of\ features$	$\sqrt{number\ of\ feat}$
num.trees	$1:\infty$	$500$

Ajuster les hyperparamètres

cv_tune <- cv_data %>% 
  crossing(mtry = 1:5)

cv_tune

# A tibble: 25 x 5
   splits       id    train                validate            mtry
   <list>       <chr> <list>               <list>             <int>
 1 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [601 × 7]>     1
 2 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [601 × 7]>     2
 3 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [601 × 7]>     3
 4 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [601 × 7]>     4
 5 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [601 × 7]>     5
 6 <S3: rsplit> Fold2 <tibble [2,402 × 7]> <tibble [601 × 7]>     1
 7 <S3: rsplit> Fold2 <tibble [2,402 × 7]> <tibble [601 × 7]>     2

Ajuster les hyperparamètres

cv_model_tunerf <- cv_tune %>% 
  mutate(model = map2(train, mtry, ~ranger(formula = life_expectancy~., 
                                           data = .x, mtry = .y)))

cv_model_tunerf

# A tibble: 25 x 6
   splits       id    train                validate      mtry  model       
 * <list>       <chr> <list>               <list>        <int> <list>      
 1 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [60…   1    <S3: ranger>
 2 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [60…   2    <S3: ranger>
 3 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [60…   3    <S3: ranger>
 4 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [60…   4    <S3: ranger>
 5 <S3: rsplit> Fold1 <tibble [2,402 × 7]> <tibble [60…   5    <S3: ranger>
 6 <S3: rsplit> Fold2 <tibble [2,402 × 7]> <tibble [60…   1    <S3: ranger>
 7 <S3: rsplit> Fold2 <tibble [2,402 × 7]> <tibble [60…   2    <S3: ranger>

Passons à la pratique !

Machine Learning dans le tidyverse