svm.Rmd

---
title: "R Notebook"
output:
  html_document:
    df_print: paged
  word_document:
    toc: yes
  html_notebook:
    number_sections: yes
    toc: yes
---

# Model SVM
## Set-up
```{r}
# clear the workspace 
rm(list = ls())     

# install all the packages
library(readxl)
library(tidyverse)
library(ggplot2)
library(e1071)
library(caTools)
library(caret)
library(pROC)
```

##Read the "Don't get kicked" dataset 
```{r}
rawdata<-read.csv("TRAIN_Numeric.csv")
```

##Set up for holdout validation, randomly select 70% of the dataset as trainning dataset and saved the rest as testing dataset
```{r}
set.seed(11)   # set a random seed 
sample <- sample.int(n = nrow(rawdata), size = floor(.7*nrow(rawdata)), replace = F)
train <- rawdata[sample, ]
test  <- rawdata[-sample, ]
```

##Select variables used for svm model [Variables were selected based on maximazing information gain]
```{r}
sixVariables <- c("VehicleAge","VehBCost","BPER","PRIMEUNIT","VehYear","VehOdo")
x <- train[sixVariables]
y <- train["IsBadBuy"]
```




#first svm model 
##inital model with linear kernel
```{r}
model_svm<-svm(IsBadBuy ~ VehicleAge+VehBCost+BPER+PRIMEUNIT+VehYear+VehOdo,
               data=train, 
               type='C-classification',
               kernel="linear", 
               cost=0.01)         
model_svm
```

##predict and test the model on the testing data 
```{r}
test$svm_pred_class <- predict(model_svm,test)

result <- test %>%
  select(IsBadBuy,svm_pred_class)

result$IsBadBuy_factor <- result$IsBadBuy %>%
  as.factor() 
```

##confusion matrix
```{r}
confusionMatrix(data = result$svm_pred_class, reference = result$IsBadBuy_factor)
```

##ROC Curve
```{r}
test$svm_dv<-as.numeric(attr(predict(model_svm, test, decision.values = TRUE),"decision.values"))
svm_roc<-roc(test$IsBadBuy,test$svm_dv,auc=TRUE)
plot(svm_roc,print.auc=TRUE,col="blue")
```



#Second model Testing out "radial" kernal 
```{r}
model_svm2<-svm(IsBadBuy ~ VehicleAge+VehBCost+BPER+PRIMEUNIT+VehYear+VehOdo,
               type='C-classification',
               data=train,                   
               kernel="radial",
               cost=0.01)
print(model_svm2)
```

##Predict and test the model on the testing data 
```{r}
#predict and test the model on the testing data 
test$svm_pred_class2 <- predict(model_svm2,test)

result <- test %>%
  select(IsBadBuy,svm_pred_class2)

result$IsBadBuy_factor <- result$IsBadBuy %>%
  as.factor() 

```

##Confusion table 
```{r}
confusionMatrix(data = result$svm_pred_class2, reference = result$IsBadBuy_factor)
```

##ROC Curve
```{r}
test$svm_dv2<-as.numeric(attr(predict(model_svm2, test, decision.values = TRUE),"decision.values"))
svm_roc_2<-roc(test$IsBadBuy,test$svm_dv2,auc=TRUE)
plot(svm_roc_2,print.auc=TRUE,col="red")
```



#Third model using "sigmoid"
```{r}
model_svm3<-svm(IsBadBuy ~ VehicleAge+VehBCost+BPER+PRIMEUNIT+VehYear+VehOdo,
               type='C-classification',
               data=train,                   
               kernel="sigmoid",
               cost=0.01)
print(model_svm3)
```

##Predict and test the model on the testing data 
```{r}
#predict and test the model on the testing data 
test$svm_pred_class3 <- predict(model_svm3,test)

result <- test %>%
  select(IsBadBuy,svm_pred_class3)

result$IsBadBuy_factor <- result$IsBadBuy %>%
  as.factor() 

```

##Confusion table 
```{r}
confusionMatrix(data = result$svm_pred_class3, reference = result$IsBadBuy_factor)
```

##ROC Curve
```{r}
test$svm_dv3<-as.numeric(attr(predict(model_svm3, test, decision.values = TRUE),"decision.values"))
svm_roc_3<-roc(test$IsBadBuy,test$svm_dv3,auc=TRUE)
plot(svm_roc_3,col="black")
plot(svm_roc_2,col="green",add=TRUE)
plot(svm_roc,col="blue",add=TRUE)
```


#Comput probilility 
```{r}
#build the inital model using training data
model_svm_prob<-svm(IsBadBuy ~ VehicleAge+VehBCost+BPER+PRIMEUNIT+VehYear+VehOdo,
               type='C-classification',
               data=train,                   
               kernel="linear",
               cost=0.01,
               probability=TRUE)
model_svm_prob
print(model_svm_prob)

#predict and test the model on the testing data 
test$svm_pred_class_prob <- predict(model_svm_prob,test,probability=TRUE,type="response")

#predict probabilities
prob_table<-attr(predict(model_svm_prob,test,probability=TRUE), "probabilities") 

#compute a proability table
prob_table <- data.frame(prob_table)

#define if probability of a badbuy is less than 0.12, than it is a badbuy. 
prob_table$badbuy=ifelse(prob_table$X1<0.11, "1", "0") 
prob_table$badbuy <- as.factor(prob_table$badbuy)

test$IsBadBuy <- as.factor(test$IsBadBuy)

#confusion matrix
confusionMatrix(data=prob_table$badbuy, reference = test$IsBadBuy)

test$svm_dv4<-attr(predict(model_svm_prob, test,probability=TRUE ),"probabilities")
```