How to plot a ROC curve from Classification Tree p

2019-07-22 15:42发布

问题:

I am attempting to plot a ROC curve with classification trees probabilities. However, when I plot the curve, it is absent. I am trying to plot the ROC curve and then find the AUC value from the area under the curve. Does anyone know how to fix this? Thank you if you can. The binary column Risk stands for risk misclassification, which I presume is my label. Should I be applying the ROC curve equation at a different point in my code?

Here is the data frame:

   library(ROCR)

   data(Risk.table)

   pred = prediction(Risk.table$Predicted.prob, Risk.table2$Risk)
   perf = performance(pred, measure="tpr", x.measure="fpr")
   perf
   plot(perf)

   Predicted.prob Actual.prob   predicted actual Risk
  1       0.5384615   0.4615385        G8     V4    0
  2       0.1212121   0.8787879        V4     V4    1
  3       0.5384615   0.4615385        G8     G8    1
  4       0.9000000   0.1000000        G8     G8    1
  5       0.1212121   0.8787879        V4     V4    1
  6       0.1212121   0.8787879        V4     V4    1
  7       0.9000000   0.1000000        G8     G8    1
  8       0.5384615   0.4615385        G8     V4    0
  9       0.5384615   0.4615385        G8     V4    0
  10      0.1212121   0.8787879        V4     G8    0
  11      0.1212121   0.8787879        V4     V4    1
  12      0.9000000   0.1000000        G8     V4    0
  13      0.9000000   0.1000000        G8     V4    0
  14      0.1212121   0.8787879        G8     V4    1
  15      0.9000000   0.1000000        G8     G8    1
  16      0.5384615   0.4615385        G8     V4    0
  17      0.9000000   0.1000000        G8     V4    0
  18      0.1212121   0.8787879        V4     V4    1
  19      0.5384615   0.4615385        G8     V4    0
  20      0.1212121   0.8787879        V4     V4    1
  21      0.9000000   0.1000000        G8     G8    1
  22      0.5384615   0.4615385        G8     V4    0
  23      0.9000000   0.1000000        G8     V4    0
  24      0.1212121   0.8787879        V4     V4    1

Here is the ROC curve this code outputs, but the curve is missing:

I tried again and this ROC curve is just wrong

I constructed the above data frame using the code below:

The initial data frame containing all the data is called shuffle.cross.validation2

  #Split data 70:30 after shuffling the data frame

  index<-1:nrow(LDA.scores1)
  trainindex.LDA3=sample(index, trunc(length(index)*0.70),replace=FALSE)      

  LDA.70.trainset3<-shuffle.cross.validation2[trainindex.LDA3,]

  LDA.30.testset3<-shuffle.cross.validation2[-trainindex.LDA3,]

Run classification tree using package rpart()

 tree.split3<-rpart(Family~., data=LDA.70.trainset3, method="class")
 tree.split3
 summary(tree.split3)
 print(tree.split3)
 plot(tree.split3)
 text(tree.split3,use.n=T,digits=0)
 printcp(tree.split3)
 tree.split3

Predict the predicted and actual data

 res3=predict(tree.split3,newdata=LDA.30.testset3)
 res4=as.data.frame(res3)

Create two columns with NA's (Actual and predicted classification rate)

 res4$predicted<-NA
 res4$actual<-NA


 for (i in 1:length(res4$G8)){

 if(res4$R2[i]>res4$V4[i]) {
 res4$predicted[i]<-"G8"
 }

 else {
 res4$predicted[i]<-"V4"
 }

  print(i)
 }

 res4

 res4$actual<-LDA.30.testset3$Family
 res4
 Risk.table$Risk<-NA
 Risk.table

Create the binary predictor column

  for (i in 1:length(Risk.table$Risk)){

  if(Risk.table$predicted[i]==res4$actual[i]) {
  Risk.table$Risk[i]<-1
  }

  else {
  Risk.table$Risk[i]<-0
  }

  print(i)
  }

Creation of the predicted and actual probabilities for the two families V4 and G8 above

    #Confusion Matrix

    cm=table(res4$actual, res4$predicted)

    names(dimnames(cm))=c("actual", "predicted")

Naive Bayes

  index<-1:nrow(significant.lda.Wilks2)
  trainindex.LDA.help1=sample(index, trunc(length(index)*0.70), replace=FALSE)                                     
  sig.train=significant.lda.Wilks2[trainindex.LDA.help1,]
  sig.test=significant.lda.Wilks2[-trainindex.LDA.help1,]


    library(klaR)
    nbmodel<-NaiveBayes(Family~., data=sig.train)
    prediction<-predict(nbmodel, sig.test)
    NB<-as.data.frame(prediction)
    colnames(NB)<-c("Actual", "Predicted.prob", "acual.prob")

    NB$actual2 = NA
    NB$actual2[NB$Actual=="G8"] = 1
    NB$actual2[NB$Actual=="V4"] = 0
    NB2<-as.data.frame(NB)

    plot(fit.perf, col="red"); #Naive Bayes
    plot(perf, col="blue", add=T); #Classification Tree
    abline(0,1,col="green")

Original Naive Bayes code using the caret package

     library(caret)
     library(e1071)

  train_control<-trainControl(method="repeatedcv", number=10, repeats=3)
  model<-train(Matriline~., data=LDA.scores, trControl=train_control,    method="nb")
  predictions <- predict(model, LDA.scores[,2:13])
  confusionMatrix(predictions,LDA.scores$Family)

Results

               Confusion Matrix and Statistics

                        Reference
                Prediction V4 G8
                        V4 25  2
                        G8  5 48

                  Accuracy : 0.9125         
                    95% CI : (0.828, 0.9641)
       No Information Rate : 0.625          
       P-Value [Acc > NIR] : 4.918e-09      

                    Kappa : 0.8095         
   Mcnemar's Test P-Value : 0.4497         

              Sensitivity : 0.8333         
              Specificity : 0.9600         
           Pos Pred Value : 0.9259         
           Neg Pred Value : 0.9057         
               Prevalence : 0.3750         
           Detection Rate : 0.3125         
     Detection Prevalence : 0.3375         
        Balanced Accuracy : 0.8967         

         'Positive' Class : V4         

回答1:

I have various things to point out:

1) I think your code has to be Family ~ . inside your rpart command.

2) In your initial table I can see a value W3 in your predicted column. Does that mean you don’t have a binary dependent variable? ROC curves work with binary data, so check it.

3) Your predicted and actual probabilities in your initial table always sum to 1. Is that reasonable? I think they represent something else, so you might consider changing names in case they confuse you in the future.

4) I think you’re confused about how ROC works and what inputs it needs. Your Risk column uses 1 to represent a correct prediction and 0 to represent a wrong prediction. However, the ROC curve needs 1 to represent one class and 0 to represent the other class. In simple words, the command is prediction(predictions, labels) where predictions are your predicted probabilities and labels are the true class/levels of your dependent variable. Check the following code:

dt = read.table(text="
Id Predicted.prob Actual.prob   predicted actual Risk
1       0.5384615   0.4615385        G8     V4    0
2       0.1212121   0.8787879        V4     V4    1
3       0.5384615   0.4615385        G8     G8    1
4       0.9000000   0.1000000        G8     G8    1
5       0.1212121   0.8787879        V4     V4    1
6       0.1212121   0.8787879        V4     V4    1
7       0.9000000   0.1000000        G8     G8    1
8       0.5384615   0.4615385        G8     V4    0
9       0.5384615   0.4615385        G8     V4    0
10      0.1212121   0.8787879        V4     G8    0
11      0.1212121   0.8787879        V4     V4    1
12      0.9000000   0.1000000        G8     V4    0
13      0.9000000   0.1000000        G8     V4    0
14      0.1212121   0.8787879        W3     V4    1
15      0.9000000   0.1000000        G8     G8    1
16      0.5384615   0.4615385        G8     V4    0
17      0.9000000   0.1000000        G8     V4    0
18      0.1212121   0.8787879        V4     V4    1
19      0.5384615   0.4615385        G8     V4    0
20      0.1212121   0.8787879        V4     V4    1
21      0.9000000   0.1000000        G8     G8    1
22      0.5384615   0.4615385        G8     V4    0
23      0.9000000   0.1000000        G8     V4    0
24      0.1212121   0.8787879        V4     V4    1", header=T)

library(ROCR)

roc_pred <- prediction(dt$Predicted.prob, dt$Risk)
perf <- performance(roc_pred, "tpr", "fpr")
plot(perf, col="red")
abline(0,1,col="grey")

The ROC curve is :

When you create a new column actual2 where you have 1 instead of G8 and 0 instead of V4:

dt$actual2 = NA
dt$actual2[dt$actual=="G8"] = 1
dt$actual2[dt$actual=="V4"] = 0

roc_pred <- prediction(dt$Predicted.prob, dt$actual2)
perf <- performance(roc_pred, "tpr", "fpr")
plot(perf, col="red")
abline(0,1,col="grey")

5) As @eipi10 mentioned above, you should try to get rid of the for loops in your code.