Machine_Learning_Sample_Project.rmd

---
title: ""
author: ""
date: ""
output: github_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE, message = FALSE, warning = FALSE)
```


***
***


## Read data
```{r}
Data <- read.csv(file = file.choose(), header = TRUE)
head(Data)
FMdat <- Data[, c(9, 6, 7)]
head(FMdat)
nn <- dim(FMdat)[1]
nn
test_ind <- sample(1:nn, size = 400 * 30 / 100, replace = FALSE)
dtrain <- FMdat[-test_ind, ]
dtest <- FMdat[test_ind, ]
dim(dtrain)
dim(dtest)
ytest <- dtest$US

acc_error <- function(tab) {
    b <- sum(tab)
    DIM <- dim(tab)
    if (DIM[1] == 2 && DIM[2] == 2) {
    a = tab[1, 1] + tab[2, 2]
    acc = a/b; error = 1 - acc
    result <- c("Acc" = acc, "Error" = error)
    } else {
        a = sum(diag(tab))
        result = c("Acc" = a / b, "Error" = 1 - a/b)
    }
    result
}
```

***
***

# Classification -------------



***
***



## 1- SVM Model ---------------

```{r}
library(e1071)

## define Model ------------------
set.seed(444)
Model_Svm <- svm(factor(US) ~ Price + Age, data = dtrain, 
kernel = "radial", gamma = 1.5, 
cost = 1, decision.values = TRUE)
summary(Model_Svm)

pred_LABEL_svm <- predict(Model_Svm, newdata = dtest)

cfMat <- table(real = ytest, preds = pred_LABEL_svm)
cfMat
res_svm <- acc_error(cfMat)
res_svm


pred_SCORE_svm <- attributes(
    predict(Model_Svm, dtest, 
    decision.values = TRUE) )$decision.values
```


***
***

## 2-  Tree Model -------------------


```{r}
library(rpart)
library(rpart.plot)
Model_Tree <- rpart(factor(US) ~ Price + Age, 
data = dtrain, parms = list(split = "information"), 
method = "class")

# summary(Model_Tree)
##########################

pred_LABEL_tree <- predict(Model_Tree, dtest, type = "class")
pred_SCORE_tree <- predict(Model_Tree, dtest, type = "prob")[, 2]
cfMat <- table(real = ytest, pred = pred_LABEL_tree)
cfMat
res_tree <- acc_error(cfMat)
res_tree
rpart.plot(Model_Tree)
```
***
***


## 3- Boosting Model ---------------


```{r}
set.seed(123)
library(gbm)
Model_boost <- gbm(US ~ Price + Age, data = dtrain, 
distribution = "bernoulli", 
shrinkage = 0.0001)
pred_SCORE_boost <- predict(Model_boost, 
newdata = dtest, type = "response")
pred_LABEL_boost <- (pred_SCORE_boost > 0.5) / 1
cfMat <- table(real = ytest, 
preds = pred_LABEL_boost)
cfMat
res_boost1 <- acc_error(cfMat)
res_boost1
## better cutof

library(ROCR)

train_pred <- predict(Model_boost, type = "response")
yt <- dtrain$US
temp1 <- predict(Model_boost, type = "response")
temp2 <- prediction(temp1, yt) 
cost_perf = performance(temp2, "cost")
Cut_p <- cost_perf@x.values[[1]][which.min(cost_perf@y.values[[1]])] 
Cut_p


pred_LABEL_boost2 <- (pred_SCORE_boost >  Cut_p) / 1 
tab <- table(real = ytest, pred = pred_LABEL_boost2)
tab
res_boost_2 <- acc_error(tab)
res_boost_2

```
***
***


## 4-  Logistic Model -----------------

```{r}
set.seed(123)
Model_log <- glm(US ~ Price + Age, data = dtrain,
family = binomial)
pred_SCORE_log <- predict(Model_log, dtest, 
type = "response")
pred_LABEL_log <- (pred_SCORE_log > 0.5) / 1

tab <- table(real = ytest, pred = pred_LABEL_log) 
tab
res_log <- acc_error(tab)
res_log
## change cutpoint 


temp1 <- predict(Model_log, type = "response")
temp2 <- prediction(temp1, dtrain$US) 
cost_perf = performance(temp2, "cost")
Cut_p <- cost_perf@x.values[[1]][which.min(cost_perf@y.values[[1]])] 
Cut_p

pred_LABEL_log2 <- (pred_SCORE_log >  Cut_p) / 1 
tab <- table(real = ytest, pred = pred_LABEL_log2)
tab
res_log_2 <- acc_error(tab)
res_log_2

```
***
***

## 5-  QDA Model ------------------------



```{r}
library(MASS)
Model_qda <- qda(US ~ Price + Age, data = dtrain)
pred <- predict(Model_qda, newdata = dtest, 
type = "class")

pred_LABEL_qda <- pred$class
pred_SCORE_qda <- pred$posterior[, 2]
tab <- table(real = ytest, pred = pred_LABEL_qda)
tab
res_qda <- acc_error(tab)
res_qda
```
```
***
***


## 6-  RandomForest Model -----------


```{r}
set.seed(111)
library(randomForest)

Model_rf <- randomForest(factor(US) ~ Price + Age, 
data = dtrain, 
ntree = 300)
pred_LABEL_rf <- predict(Model_rf, 
dtest, type = "response")
pred_SCORE_rf <- predict(Model_rf, 
dtest, type = "prob")[, 2]

tab <- table(real = ytest, 
preds = pred_LABEL_rf)
tab
res_rf <- acc_error(tab)
res_rf
```


***
***


## 7-  LDA Model -------------------

```{r}

library(MASS)
Model_lda <- lda(US ~ Price + Age, data = dtrain)
pred <- predict(Model_lda, newdata = dtest, type = "class")
pred_LABEL_lda <- pred$class
pred_SCORE_lda <- pred$posterior[, 2]
tab <- table(pred_LABEL_lda, ytest)
tab
res_lda <- acc_error(tab)

res_lda
```

***
***

## 8- bagging -------------------



```{r}
library(caret)
cvcontrol <- trainControl(method="repeatedcv", number = 5,
                          allowParallel=TRUE)
Model_bag <- train(factor(US) ~ Price + Age, data = dtrain, 
method = "treebag", 
trControl = cvcontrol, 
importance = TRUE) 

Model_bag


pred_SCORE_bag <- predict(Model_bag, dtest, 
type = "prob")[, '1']

pred_LABEL_bag <- predict(Model_bag, dtest, type = "raw")
tab <- table(real = ytest, pred = pred_LABEL_bag)
tab
res_bag <- acc_error(tab)
res_bag
```
***
***


## Results ---------------

```{r fig.height = 9, fig.width = 9}
library(pROC)
roc_bag <- roc(ytest, pred_SCORE_bag)
roc_log <- roc(ytest, pred_SCORE_log)
roc_lda <- roc(ytest, pred_SCORE_lda)
roc_qda <- roc(ytest, pred_SCORE_qda)
roc_rf <- roc(ytest, pred_SCORE_rf)
roc_boost <- roc(ytest, pred_SCORE_boost)
roc_svm <- roc(ytest, pred_SCORE_svm)
roc_tree <- roc(ytest, pred_SCORE_tree)
plot(roc_bag, main = "ROC", col = 1)
plot(roc_log, col = 2, add = T)
plot(roc_lda, col = 3, add = T)
plot(roc_qda, col = 4, add = T)
plot(roc_rf, col = 5, add = T)
plot(roc_boost, col = 6, add = T)
plot(roc_svm, col = 7, add = T)
plot(roc_tree, col = 8, add = T)
legend("bottomright", 
legend = c("Bagging", "Logistic", "LDA", 
"QDA", "Randomforest", "SVM", "LDA", "QDA"), 
col = 1:8, lwd = 2, lty = 1, bty = "n")
Models <- c("bagging", "logistic", "LDA", "QDA", 
"Randomforest", "Boosting", "SVM", "TREE")

AUC <- c(auc(roc_bag), auc(roc_log), auc(roc_lda), 
auc(roc_qda), auc(roc_rf), auc(roc_boost), auc(roc_svm), 
auc(roc_tree))

ERROR = c(res_bag[2], res_log_2[2], res_lda[2], 
res_qda[2], res_rf[2], res_boost_2[2], res_svm[2], res_tree[2])

data_result <- data.frame(Model = Models, 
AUC = AUC, Error = ERROR)

knitr :: kable(data_result, caption = "RESULTS", align = "c")


```


***
***

# clustering ---------------


```{r}
library(ISLR)
names(USArrests)
dd <- USArrests[c('Assault', "UrbanPop")]
dd2 <- scale(dd)
tail(dd2, 4)
distt <- dist(dd2, method = "euclidean")

average_model <- hclust(distt, method = "average")

plot(average_model, hang = -1, 
main = "Dendogram Average Linkage")

```