Analyse en composantes principales (ACP)

Réduction de dimension en R

Matt Pickard

Owner, Pickard Predictives, LLC

Réaliser une ACP

pca_res <- prcomp(attrition_df %>% select(-Attrition), scale. = TRUE)

summary(pca_res)

Importance of components:
                          PC1    PC2    PC3     PC4     PC5
Standard deviation     1.4259 1.3295 0.8618 0.48401 0.47138
Proportion of Variance 0.4067 0.3535 0.1485 0.04685 0.04444
Cumulative Proportion  0.4067 0.7602 0.9087 0.95556 1.00000

Charges factorielles des CP

pca_res

Standard deviations (1, .., p=5):
[1] 1.43 1.33 0.86 0.48 0.47

Rotation (n x k) = (5 x 5):
                            PC1    PC2     PC3    PC4    PC5
MonthlyIncome            0.6244 -0.024  0.3665 -0.280 -0.630
TotalWorkingYears        0.6390 -0.011  0.2674  0.293  0.659
YearsSinceLastPromotion  0.4488  0.018 -0.8902 -0.061 -0.047
PercentSalaryHike       -0.0018  0.707  0.0426 -0.647  0.284
PerformanceRating        0.0210  0.707 -0.0033  0.643 -0.294

Charges factorielles des CP

zoom sur les charges des deux premières composantes principales

Charges factorielles des CP

charges des variables pour CP1

Charges factorielles des CP

charges des variables pour CP2

ACP avec tidymodels

pca_recipe <- recipe(Attrition ~ . , data = train) %>% 
  step_normalize(all_numeric_predictors()) %>% 
  step_pca(all_numeric_predictors(), num_comp = 2)


attrition_fit <- workflow(preprocessor = pca_recipe, spec = logistic_reg()) %>% 
  fit(train)


attrition_pred_df <- predict(attrition_fit, test) %>% 
  bind_cols(test %>% select(Attrition))


f_meas(attrition_pred_df, Attrition, .pred_class)

Voir les CP dans les détails du modèle

attrition_fit

Call:  stats::glm(formula = ..y ~ ., family = stats::binomial, data = data)

Coefficients:
(Intercept)          PC1          PC2  
   -2.42067      0.80493     -0.03429  

Degrees of Freedom: 1339 Total (i.e. Null);  1337 Residual
Null Deviance:        951.8 
Residual Deviance: 870.6     AIC: 876.6

Passons à la pratique !

Réduction de dimension en R