---
title: "Inclass 11-08"
author: "Nino & Theo"
date: "8 novembre 2018"
output: html_document
---

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

```{r}


spor <- as.data.frame(read.table("spor.csv", header = T, sep = ","))

spor$Medu <- as.factor(spor$Medu)
levels(spor$Medu) <- c("none", "4th grade", "5th to 9th grade", "secondary eduction", "higher education")
spor$Fedu <- as.factor(spor$Fedu)
levels(spor$Fedu) <- c("none", "4th grade", "5th to 9th grade", "secondary eduction", "higher education")
spor$studytime <- as.factor(spor$studytime)
levels(spor$studytime) <- c("<2", "2 to 5", "5 to 10", ">10")

linmod <- glm(alc ~., data = spor, family = "binomial")
summary(linmod)

```

```{r}
library(kernlab)
library(caret)
library(caretEnsemble)
library(forecast)

spor <- as.data.frame(read.table("spor.csv", header = T, sep = ","))

spor$age2 <- 0
spor$age2[spor$age >= 19] <- 1
spor$age <- NULL



data <- spor


set.seed(11)


intrain <- createDataPartition(y = data$alc, p= 0.8, list = FALSE)
data_train <- data[intrain,]
data_test <- data[-intrain,]


trctrl <- trainControl(method = "repeatedcv", number=10, repeats=5)


#alc_fit <- train(as.factor(alc) ~., data = data_train, method = "svmLinear", trControl=trctrl, preProcess = c("center", "scale"), tuneLength=20)
#alc_fit


alc_fit2 <- train(as.factor(alc) ~ sex + famsize * famrel + Pstatus * famrel + studytime * absences + health + absences * goout, data = data_train, method = "svmLinear", trControl=trctrl, preProcess = c("center", "scale"), tuneLength=20)
alc_fit2



```


```{r}
data_test$svm1<-predict(alc_fit2, newdata=data_test)
#data_test$svm1<-as.integer(ifelse(data_test$svm1 >= 0.5, 1, 0))
confusionMatrix(as.factor(data_test$alc), as.factor(data_test$svm1))
```
The base accuracy and kappa are 65.59% and 25.88%. We will try to increase that by ensembling models

```{r}
library(xgboost)
fit2 <- train(as.factor(alc) ~., data = data_train, metric="Accuracy", method = "xgbLinear", trControl=trctrl, preProcess = c("center", "scale"), tuneLength=1)
data_test$fit2<-predict(fit2, newdata=data_test)
confusionMatrix(as.factor(data_test$alc), as.factor(data_test$fit2))
```
```{r}
fit3 <- train(as.factor(alc) ~., data = data_train, metric="Accuracy", method = "xgbTree", trControl=trctrl, preProcess = c("center", "scale"), tuneLength=1)
data_test$fit3<-predict(fit3, newdata=data_test)
confusionMatrix(as.factor(data_test$alc), as.factor(data_test$fit3))
```

```{r}

fit4 <- train(as.factor(alc) ~., data = data_train, metric="Accuracy", method = "xgbLinear", trControl=trctrl, preProcess = c("center", "scale"), tuneLength=1, tuneGrid=data.frame(nrounds=45, lambda=0, alpha=0, eta=0.5))
data_test$fit4<-predict(fit4, newdata=data_test)
confusionMatrix(as.factor(data_test$alc), as.factor(data_test$fit4))

```




```{r}
data_test$ensemble <- as.numeric(as.character(data_test$svm1)) + as.numeric(as.character(data_test$fit2)) + as.numeric(as.character(data_test$fit3)) + as.numeric(as.character(data_test$fit4)) * 2

data_test$ensemble_predicted <- ifelse(data_test$ensemble >= 3, 1, 0)
data_test$ensemble_predicted <- as.factor(data_test$ensemble_predicted)

confusionMatrix(as.factor(data_test$alc), data_test$ensemble_predicted)
```