# This R script will run on our backend. You can write arbitrary code here!

# Many standard libraries are already installed, such as randomForest
library('stringr')
library('rpart')
library('ggplot2')
library('dplyr')
library('randomForest')
library('caret')
library('purrr')
library('mice')
library('tidyr')

# The train and test data is stored in the ../input directory
train_raw <- read.csv("../input/train.csv", stringsAsFactors = FALSE)
test_raw  <- read.csv("../input/test.csv", stringsAsFactors = FALSE)
test_raw$Survived <- NA

# Stack the training and testing dataset to handle missing values
all <- rbind(train_raw, test_raw)
all[all == ''] <- NA

# Set factor variables
all$Pclass <- as.factor(all$Pclass)

all$Sex <- as.factor(all$Sex)


# Compute the family size
Family_size <- all$SibSp + all$Parch + 1
table(Family_size)
all$Family[Family_size==1] <- 'Single'
all$Family[Family_size>1 & Family_size<=4] <- 'Small'
all$Family[Family_size>=5] <- 'Big'



# Create a title variable
all$Title <- gsub(pattern = '.*,\\s|\\..*','',all$Name)
names(table(all$Title)[table(all$Title) < 20])
#Male Rare title: 
all$Title[all$Title %in% 
               c('Capt', 'Don', 'Major', 'Sir', 'Col', 'Jonkheer', 'Rev', 'Dr')] <- 'Mr'

#Female Rare title:
all$Title[all$Title %in% c('Mlle', 'Mme', 'Ms')] <- 'Miss'
all$Title[all$Title %in% c('the Countess', 'Lady', 'Dona')] <- 'Mrs'



# Make character variables factors
fac_var <- c('PassengerId', 'Name', 'Ticket', 'Cabin', 'Embarked', 'Title', 'Family',
             'Survived')
all[,fac_var] <- all[,fac_var] %>%
     map_df(as.factor)



# Find which variables have missing value
apply(all, 2, function(x) {sum(is.na(x))}) # Age, Fare, Cabin, Embarked


#impute using mice()
all_rf <- all[,-c(1,2,4,7,8,9,11)] %>%
     mice(method = 'rf') %>%
     mice::complete()

all_cart <- all[,-c(1,2,4,7,8,9,11)] %>%
     mice(method = 'cart') %>%
     mice::complete()

# visulization of age
age_orig <- all$Age
age_rf <- all_rf$Age
age_cart <- all_cart$Age
age1 <- as.data.frame(cbind(age_orig,age_cart,age_rf))
names(age1) <- c('Original', 'CART', 'RF')
age <- gather(age1, key = origin, value = age, Original, CART, RF)
ggplot(age,aes(x = age, y = ..density.., fill = origin)) + 
     stat_bin(alpha = 0.5, position = position_dodge(1))
# From the histogram, we find random forest has a slightly better result than 
# cart, so we use the RF imputation

all[,c('Age', 'Fare', 'Embarked')] <- all_rf[,c('Age', 'Fare', 'Embarked')]



# Split the data back to training and testing dataset
train <- all[1:nrow(train_raw),]
test <- all[-(1:nrow(train_raw)),]

set.seed(16)

# Model 1.Random Forest
control_rf <- trainControl(method = 'cv', number = 10, verboseIter = TRUE, 
                           savePredictions = 'final' )
model_rf <- train(Survived ~ Pclass + Sex + Age + Family + Embarked + Title,
                  train, method = 'rf', trControl = control_rf)


# train_rf_pred <- predict.train(model_rf,train[,-2])
train_rf_pred <- model_rf$pred[,c('pred', 'obs')]

# Assess the training result
class_assess <- function(x) {
     # train$Pred <- x
     pred_table <- table(x$obs,x$pred)
     tp <- pred_table[1,1]
     fp <- pred_table[1,2]
     fn <- pred_table[2,1]
     tn <- pred_table[2,2]
     Accuracy <- sum(diag(pred_table)) / sum(pred_table)
     Precision <- tp / (tp + fp)
     Recall <- tp / (tp + fn)
     F1_Score <- 2 * Precision * Recall / (Precision + Recall)
     return(data.frame(Accuracy = Accuracy, Precision = Precision,
                       Recall = Recall, F1_Score = F1_Score))
}

RF <- class_assess(train_rf_pred)





# Model 2.Support Vector Machine
control_svm <- trainControl(method = 'cv', number = 10, verboseIter = TRUE, 
                            savePredictions = 'final' )
model_svm <- train(Survived ~ Pclass + Sex + Age + Family + Embarked + Title,
                   train, method = 'svmLinear', trControl = control_svm)

train_svm_pred <- model_svm$pred[,c('pred', 'obs')]
SVM <- class_assess(train_svm_pred)




# Model 3.K Nearest Neighbor
control_knn <- trainControl(method = 'cv', number = 10, verboseIter = TRUE, 
                            savePredictions = 'final' )
model_knn <- train(Survived ~ Pclass + Sex + Age + Family + Embarked + Title,
                   train, method = 'kknn', trControl = control_knn)

train_knn_pred <- model_knn$pred[,c('pred', 'obs')]
KNN <- class_assess(train_knn_pred)


# Model 4.Neural Network
control_nnet <- trainControl(method = 'cv', number = 10, verboseIter = TRUE, 
                             savePredictions = 'final' )
model_nnet <- train(Survived ~ Pclass + Sex + Age + Family + Embarked + Title,
                    train, method = 'nnet', trControl = control_nnet)

train_nnet_pred <- model_nnet$pred[,c('pred', 'obs')]
NNET <- class_assess(train_nnet_pred)

comparison <- rbind(RF = RF,SVM = SVM,KNN = KNN,NNET = NNET)
comparison #Neural Network has the best result


# Making prediction based on the best model
test_pred <- predict.train(model_nnet,test)
titanic_prediction <- data.frame(PassengerId = test$PassengerId, 
                                 Survived = test_pred)
write.csv(titanic_prediction, file = "Titanic_prediction.csv",row.names=FALSE)