1. Importer les données

library(kernlab)
data(spam)
summary(spam[,56:58])

  capitalLong       capitalTotal          type     
 Min.   :   1.00   Min.   :    1.0   nonspam:2788  
 1st Qu.:   6.00   1st Qu.:   35.0   spam   :1813  
 Median :  15.00   Median :   95.0                 
 Mean   :  52.17   Mean   :  283.3                 
 3rd Qu.:  43.00   3rd Qu.:  266.0                 
 Max.   :9989.00   Max.   :15841.0                 

set.seed(5678)
perm <- sample(4601,3000)
app <- spam[perm,]
valid <- spam[-perm,]

2. Construire et analyser une forêt aléatoire

library(randomForest)
set.seed(1234)
foret <- randomForest(type~.,data=app)
foret

Call:
 randomForest(formula = type ~ ., data = app) 
               Type of random forest: classification
                     Number of trees: 500
No. of variables tried at each split: 7

        OOB estimate of  error rate: 4.73%
Confusion matrix:
        nonspam spam class.error
nonspam    1764   54  0.02970297
spam         88 1094  0.07445008

3. Sélectionner les paramètres de la forêt

plot(foret)

tail(foret$err.rate)

              OOB    nonspam       spam
[495,] 0.04733333 0.02915292 0.07529611
[496,] 0.04666667 0.02915292 0.07360406
[497,] 0.04700000 0.02915292 0.07445008
[498,] 0.04733333 0.02915292 0.07529611
[499,] 0.04733333 0.02970297 0.07445008
[500,] 0.04733333 0.02970297 0.07445008

grille.mtry <- data.frame(mtry=seq(1,57,by=3))
library(caret)
ctrl <- trainControl(method="oob")
library(doParallel)    # pour paralléliser
cl <- makePSOCKcluster(4)
registerDoParallel(cl)     
set.seed(12345)
sel.mtry <- train(type~.,data=app,method="rf",trControl=ctrl,tuneGrid=grille.mtry)
sel.mtry

Random Forest 

3000 samples
  57 predictor
   2 classes: 'nonspam', 'spam' 

No pre-processing
Resampling results across tuning parameters:

  mtry  Accuracy   Kappa    
   1    0.9233333  0.8356570
   4    0.9506667  0.8960131
   7    0.9536667  0.9025254
  10    0.9540000  0.9032122
  13    0.9536667  0.9025543
  16    0.9540000  0.9033846
  19    0.9503333  0.8955749
  22    0.9506667  0.8962295
  25    0.9496667  0.8941418
  28    0.9493333  0.8935198
  31    0.9470000  0.8885002
  34    0.9476667  0.8899682
  37    0.9480000  0.8906528
  40    0.9460000  0.8864809
  43    0.9453333  0.8849769
  46    0.9466667  0.8878157
  49    0.9460000  0.8864134
  52    0.9470000  0.8885665
  55    0.9453333  0.8849769

Accuracy was used to select the optimal model using the largest value.
The final value used for the model was mtry = 10.

stopCluster(cl)        ## fermeture des clusters

4. Faire de la prévision

set.seed(5432)
foret1 <- randomForest(type~.,data=app,mtry=10)
prev.valid <- predict(foret1,newdata=valid)
prev.valid[1:10]

      2       6       9      11      15      16      18      19      21      22 
   spam    spam    spam    spam    spam    spam    spam    spam nonspam    spam 
Levels: nonspam spam

prob.valid <- predict(foret1,newdata=valid,type="prob")
prob.valid[1:10,]

   nonspam  spam
2    0.000 1.000
6    0.306 0.694
9    0.056 0.944
11   0.180 0.820
15   0.086 0.914
16   0.036 0.964
18   0.024 0.976
19   0.206 0.794
21   0.542 0.458
22   0.026 0.974

5. Estimer les performances de la forêt

set.seed(5432)
foret2 <- randomForest(type~.,data=app,xtest=valid[,-58],ytest=valid[,58],keep.forest=TRUE)
set.seed(891)
foret3 <- randomForest(type~.,data=app,mtry=10,xtest=valid[,-58],ytest=valid[,58],keep.forest=TRUE)
foret2

Call:
 randomForest(formula = type ~ ., data = app, xtest = valid[,      -58], ytest = valid[, 58], keep.forest = TRUE) 
               Type of random forest: classification
                     Number of trees: 500
No. of variables tried at each split: 7

        OOB estimate of  error rate: 4.63%
Confusion matrix:
        nonspam spam class.error
nonspam    1760   58  0.03190319
spam         81 1101  0.06852792
                Test set error rate: 5.5%
Confusion matrix:
        nonspam spam class.error
nonspam     941   29  0.02989691
spam         59  572  0.09350238

foret3

Call:
 randomForest(formula = type ~ ., data = app, mtry = 10, xtest = valid[,      -58], ytest = valid[, 58], keep.forest = TRUE) 
               Type of random forest: classification
                     Number of trees: 500
No. of variables tried at each split: 10

        OOB estimate of  error rate: 4.5%
Confusion matrix:
        nonspam spam class.error
nonspam    1765   53  0.02915292
spam         82 1100  0.06937394
                Test set error rate: 5.68%
Confusion matrix:
        nonspam spam class.error
nonspam     939   31  0.03195876
spam         60  571  0.09508716

library(pROC)
prev2 <- predict(foret2,newdata=valid,type="prob")[,2]
roc2 <- roc(valid$type,prev2)
prev3 <- predict(foret3,newdata=valid,type="prob")[,2]
roc3 <- roc(valid$type,prev3)
plot(roc2,print.auc=TRUE,print.auc.cex=0.5,print.auc.x=0.4,print.auc.y=0.3)
plot(roc3,add=TRUE,col="red",print.auc=TRUE,print.auc.cex=0.5,print.auc.col="red",print.auc.x=0.4,print.auc.y=0.2)

library(rpart)
set.seed(12345)
arbre <- rpart(type~.,data=app,cp=0.0001)
library(tidyverse)
cp_opt <- arbre$cptable %>% as.data.frame() %>% filter(xerror==min(xerror)) %>% select(CP) %>% max() %>% as.numeric() 
arbre_sel <- prune(arbre,cp=cp_opt)
prev.arbre <- predict(arbre_sel,newdata=valid,type="prob")[,2]
roc.arbre <- roc(valid$type,prev.arbre)
plot(roc.arbre,add=TRUE,col="blue",print.auc=TRUE,print.auc.cex=0.5,print.auc.col="blue",print.auc.x=0.4,print.auc.y=0.1)

6. Interpréter la forêt aléatoire

var.imp <- foret2$importance
ord <- order(var.imp,decreasing=TRUE)
barplot(sort(var.imp,decreasing = TRUE)[1:10],names.arg=rownames(var.imp)[ord][1:10],cex.names=0.4)