Classification

V-cramer

cramer  <- matrix(NA,ncol(credit2),3)
for (i in (1:ncol(credit2)))
{     cramer[i,1] <- names(credit2[i])
cramer[i,2] <- sqrt(chisq.test(table(credit2[,i],credit$Cible))$statistic/
                      (length(credit2[,i])))
cramer[i,3] <- chisq.test(table(credit2[,i],credit$Cible))$p.value
}
colnames(cramer) <- c("variable","V de Cramer","p-value chi2")

# affichage des variables par V de Cramer décroissants
vcramer <- cramer[order(cramer[,2], decreasing=T),]
vcramer

par(mar = c(8, 4, 4, 0))
barplot(as.numeric(vcramer[,2]),col=gray(0:nrow(vcramer)/nrow(vcramer)),
        names.arg=vcramer[,1], ylab='V de Cramer', ylim=c(0,0.35),cex.names = 0.8, las=3)
abline(h=0.1, lty=2)
abline(h=0.2, lty=2)

• \(\gt 0.2\) :
• \(\gt 0.1\) :

Performances

y_pred_class <- predict(abre_complet, dtf_test)y_proba <- pred_class[,2]
y_predit <- as.factor(as.integer(pred_class[,2]>0.5))
y_vrai <- Y_test$drugg

Indicateurs

library(caret)

• confusionMatrix(data = y_predit, reference = y_vrai) matrice de confusion et les indicateurs associés.

MLmetrics

library(MLmetrics)

Warning

Il faut spécifier la valeur qu'on considère comme étant positive.

• Accuracy(y_pred = y_predit, y_true = y_vrai, positive = 1) précision.
• Precision(y_pred = y_predit, y_true = y_vrai, positive = 1).
• Sensitivity(y_true = y_vrai, y_pred = y_predit, positive = 1).
• Recall(y_true = y_vrai, y_pred = y_predit, positive = 1) rappel.
• Specificity(y_true = y_vrai, y_pred = y_predit, positive = 1) spécificité
• F1_Score(y_true = y_vrai, y_pred = y_predit, positive = 1) F score.
• AUC(y_pred = y_proba, y_true = y_vrai) AUC pour la courbe de ROC.

Courbe de ROC (Receiver Operating Characteristic)

library(plotROC)

1. rocdata <- data.frame(D = y_vrai, M = y_proba) définir les valeurs de la courbe.

2. ggplot(rocdata, aes(m = M, d = D) + geom_roc() afficher la courbe de ROC. Paramètres :

   • color = model colorer en fonction du modéle.

---

Courbe de Lift (courbe de gain)

library(dplyr)

nb.cible = length(which(test$cibke==1))
ref = c(rep(1,nb.cible), rep(0,length(test$FlagASV)-nb.cible))

lift.rf = data.frame( cible = test$FlagASV, proba = pred.rf)
lift.rf = lift.rf %>% arrange(desc(proba)) %>%
  mutate( cumden = cumsum(cible)/sum(cible)*100,
          perpop = (seq(n())/n()*100),
          parfait = cumsum(ref)/sum(ref)*100 )

library(ggplot2)

fonc1 = function(x){
    classe = nb.cible/nrow(test)*100
    ifelse(x<classe, (x/classe)*100, 100)
  }
fonc = function(x){x}
lift = ggplot() + 
    stat_function(fun = fonc1, data = data.frame(x = c(0, 100))) +
  stat_function(fun = fonc) + xlab("% population") + ylab("% cible")
lift + geom_line( aes(x= lift$perpop, y=lift$cumden), colour = "blue")

Courbe de Lift

library(BCA)

lift.chart( c("CCS.glm", "CCS.rpart"), data = CCSVal, targLevel = "1", 
            trueResp = 0.01,
            type = "cumulative",
            sub = "Validation")

• type = :

  • cumulative
  • incremental