Tree and Random Forest.Rmd

---
title: "Don't get kicked"
output:
  word_document: 
    toc: yes
  html_notebook: 
    number_sections: yes
    toc: yes
  html_document:
    df_print: paged
---
# TREE MODEL & RANDOM FORESTS
## Set-up
Clear the workspace and load packages. 
```{r}
rm(list = ls())
library(readxl)
library(tidyverse)
library(ggplot2)
```

Load dataset
```{r}
car <-read.csv("TRAIN_Numeric.csv", header=TRUE)
View(car)
```

## Set up for holdout validation
We select 30% of dataset. Using these indices, we will create a test and a training dataset. 
```{r}
set.seed(1) 
index <- sample(nrow(car), nrow(car)*0.3) 
test <- car[index,]      
training <-car[-index,]   
```

## Tree model 1-VehicleAge, VehBCost, BPER, PRIMEUNIT, VehYear
load rpart package
```{r}
library(rpart)
library(rpart.plot)
```

(1) Build the tree model with cp=0
```{r}
ct_model<-rpart(IsBadBuy~VehicleAge+VehBCost+VehOdo,data=training,method="class",control=rpart.control(cp=0))  
rpart.plot(ct_model)
```

(2) Check the cross-validation table using `printcp()`. Prune the tree using the cp value with the minimum xerror. 
```{r}
printcp(ct_model)
plotcp(ct_model)
min_xerror<-ct_model$cptable[which.min(ct_model$cptable[,"xerror"]),]
min_xerror
min_xerror_tree<-prune(ct_model, cp=min_xerror[1])
rpart.plot(min_xerror_tree)
```

(3) Apply the model (`min_xerror_tree`) to the test dataset to get the predicted probabilities. Save the result as a new variable `ct_pred_prob` in test.  
```{r}
test$ct_pred_prob<-predict(min_xerror_tree,test)[,2]  
```

(4) Using the 50% cut-off, generate class prediction. Save the result as a new variable `ct_pred_class` in test.  
```{r}
test$ct_pred_class=ifelse(test$ct_pred_prob>0.5,"1","0")
```

(5) Calculate the error rate of this model when we use the 50% cut-off. 
```{r}
table(test$ct_pred_class==test$IsBadBuy)
```

(6) Generate a confusion table of this model. 
```{r}
table(test$ct_pred_class,test$IsBadBuy, dnn=c("predicted","actual")) 
```

## Tree model2-VehicleAge, VehBCost, VehOdo, WheelTypeID, BodyType
(1) Build the tree model with cp=0
```{r}
ct_model1<-rpart(IsBadBuy~VehicleAge+VehBCost+VehOdo+WheelTypeID+VehYear+BodyType, data=training,control=rpart.control(cp=0),method="class")
rpart.plot(ct_model1)
```

(2)Check the cross-validation table using `printcp()`. Prune the tree using the cp value with the minimum xerror. 
```{r}
printcp(ct_model1)
plotcp(ct_model1)

min_xerror1<-ct_model1$cptable[which.min(ct_model1$cptable[,"xerror"]),]
min_xerror1
min_xerror_tree1<-prune(ct_model1, cp=min_xerror1[1],method="class")
rpart.plot(min_xerror_tree1)
```

(3) Apply the model (`min_xerror_tree1`) to the test dataset to get the predicted probabilities. Using 50% cut-off and check the performance with the confusion matrix. 
```{r}
test$ct_pred_prob1<-predict(min_xerror_tree1,test)[,2] 
test$ct_pred_class1=ifelse(test$ct_pred_prob1>0.5,"1","0")
table(test$ct_pred_class1==test$IsBadBuy)
table(test$ct_pred_class1,test$IsBadBuy, dnn=c("predicted","actual")) 
```

## Random forests-VehicleAge, VehBCost, VehOdo, WheelTypeID, BodyType

(1)Build random forest with 50% cut-off  
```{r}
library(randomForest)
set.seed(2)
rf_training_model<-randomForest(as.factor(IsBadBuy)~VehicleAge+VehBCost+VehOdo+WheelTypeID+BodyType, data=training,ntree=500,cutoff=c(0.5,0.5),mtry=2,importance=TRUE, method="class")
rf_training_model
```

(2) Using 50% cut-off and check the performance with the confusion matrix. 
```{r}
test$rf_pred_prob2<-predict(rf_training_model,test,type="prob")[,2]
test$rf_pred_class2<-predict(rf_training_model,test,type="class")
table(test$rf_pred_class2,test$IsBadBuy, dnn=c("predicted","actual")) 
```


## 4. Performance Visualization with ROC

Plot ROC curves of all of the models developed. 
```{r}
library(pROC)

ct_roc<-roc(test$IsBadBuy,test$ct_pred_prob1,auc=TRUE)
plot(ct_roc,print.auc=TRUE,col="blue",print.auc.y=.6)

ct_roc<-roc(test$IsBadBuy,test$ct_pred_prob,auc=TRUE)
plot(ct_roc,print.auc=TRUE,col="red",print.auc.y=.5,add=TRUE)

rf_roc<-roc(test$IsBadBuy,test$rf_pred_prob2,auc=TRUE)
plot(rf_roc,print.auc=TRUE,col="green",print.auc.y=1, add=TRUE)

```