Transformations de variables courantes

Feature engineering en R

Jorge Zazueta

Research Professor and Head of the Modeling Group at the School of Economics, UASLP

Deux familles de transformations

Box-Cox

Transforme une variable non normale pour la rapprocher de la normale
Famille incluant comme cas particuliers l'inverse, le log, les racines carrée et cubique
S'applique aux valeurs strictement positives

Équation de la transformation Box-Cox.

Yeo-Johnson

Propriétés proches de Box-Cox
Gère les valeurs nulles et négatives
Pour $y$ positif, équivaut à Box-Cox de $y+1$

Équation de la transformation Yeo-Johnson.

Le jeu de données loans_num

glimpse(loans_num)

Rows: 480
Columns: 6
$ Loan_Status       <fct> N, Y, Y, Y, Y, Y, N, Y, N, Y, Y, N, Y, Y, N...
$ ApplicantIncome   <dbl> 4583, 3000, 2583, 6000, 5417, 2333, 3036, 4...
$ CoapplicantIncome <dbl> 1508, 0, 2358, 0, 4196, 1516, 2504, 1526, 1...
$ LoanAmount        <dbl> 128, 66, 120, 141, 267, 95, 158, 168, 349, ...
$ Loan_Amount_Term  <dbl> 360, 360, 360, 360, 360, 360, 360, 360, 360...
$ Credit_History    <fct> 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0...

Appliquer des transformations

Recette simple

lr_recipe_plain <- # Define recipe
  recipe(Loan_Status ~., data = train)
lr_workflow_plain <- # Bundle workflows
  workflow() %>%
  add_model(lr_model) %>%
  add_recipe(lr_recipe_plain)
lr_fit_plain <- # fit and augment
  lr_workflow_plain %>%
  fit(train)

Évaluer les performances

lr_aug_plain %>% # Assess
  class_evaluate(truth = Loan_Status,
                 estimate = .pred_class,
                 .pred_N)

# A tibble: 2 × 3
  .metric  .estimator .estimate
  <chr>    <chr>          <dbl>
1 accuracy binary         0.817
2 roc_auc  binary         0.641

Appliquer des transformations

Recette Box-Cox

lr_recipe_BC <- # Define recipe
  recipe(Loan_Status ~., data = train) %>%
  step_BoxCox(all_numeric())
lr_workflow_BC <- # Bundle workflows
  workflow() %>%
  add_model(lr_model) %>%
  add_recipe(lr_recipe_BC)
lr_fit_BC <- # fit and augment
  lr_workflow_BC %>%
  fit(train)

Message d'avertissement

Box-Cox ne peut pas traiter des valeurs non positives

Warning messages:
1: Non-positive values in selected
variable. 
2: No Box-Cox transformation could be 
estimated for: `CoapplicantIncome`

Appliquer des transformations

Recette Box-Cox (2e essai)

Désélectionnons maintenant CoappliantIncome pour éviter l'avertissement.

lr_recipe_BC <- # Define recipe
  recipe(Loan_Status ~., data = train) %>%
  step_BoxCox(all_numeric(), 
              -CoapplicantIncome)
lr_workflow_BC <- # Bundle workflows
  workflow() %>%
  add_model(lr_model) %>%
  add_recipe(lr_recipe_BC)
lr_fit_BC <- # fit and augment
  lr_workflow_BC %>%
  fit(train)

Évaluer les performances

lr_aug_BC %>% # Assess
  class_evaluate(truth = Loan_Status,
                 estimate = .pred_class,
                 .pred_N)

# A tibble: 2 × 3
  .metric  .estimator .estimate
  <chr>    <chr>          <dbl>
1 accuracy binary         0.817
2 roc_auc  binary         0.599

Appliquer des transformations

Recette Yeo-Johnson

lr_recipe_YJ <- # Define recipe
  recipe(Loan_Status ~., data = train) %>%
  step_YeoJohnson(all_numeric())
lr_workflow_YJ <- # Bundle workflows
  workflow() %>%
  add_model(lr_model) %>%
  add_recipe(lr_recipe_YJ)
lr_fit_YJ <- # fit and augment
  lr_workflow_YJ %>%
  fit(train)

Évaluer les performances

lr_aug_YJ %>% # Assess
  class_evaluate(truth = Loan_Status,
                 estimate = .pred_class,
                 .pred_N)

# A tibble: 2 × 3
  .metric  .estimator .estimate
  <chr>    <chr>          <dbl>
1 accuracy binary         0.817
2 roc_auc  binary         0.700

Passons à la pratique !

Feature engineering en R