library(dplyr) # %>%

library(DataExplorer) # EDA
library(ggplot2) # plot

library(stringr) # text process
library(tm) # text mining
library(SnowballC) # stemDocument

library(caret) # train
library(naivebayes) # Naive Bayes Classifier
library(ROCR) # ROC and AUC
RCopy

# read data
data = read.csv("spam.csv", fileEncoding = "UTF-8") %>%
  mutate(Category = factor(Category))
data$Category_num = ifelse(data$Category == "ham", 0, 1)
RCopy

# calculate the number of character, word and sentence
data$num_character = nchar(data$Message)
data$num_word = lapply(data$Message, function(text){
  length(unlist(strsplit(text, "\\s+")))
  }) %>% unlist()
data$num_sentence = lapply(data$Message, function(text){
  length(unlist(strsplit(text, "[.!?]+")))
  }) %>% unlist()
RCopy

# check missing value and duplicated
data$Message %>% is.na() %>% sum()
data$Message %>% duplicated() %>% sum()

# remove duplicated
data = data[! duplicated(data$Message), ]
RCopy

# proportion of ham and spam
data$Category %>% table() %>% prop.table()
RCopy

# correlation plot
plot_correlation(
  data[, c("Category_num", "num_character", "num_word", "num_sentence")], 
  theme_config = list(axis.text.x = element_text(angle = 45))
)
RCopy

plot_densities = function(indexs, bw = 1){
  for(index in indexs){
    density_ham = data[data$Category == "ham", index] %>% density(from = 0, bw = bw)
    density_ham_x = density_ham$x %>% range()
    density_ham_y = density_ham$y %>% range()
    
    density_spam = data[data$Category == "spam", index] %>% density(from = 0, bw = bw)
    density_spam_x = density_spam$x %>% range()
    density_spam_y = density_spam$y %>% range()
    
    plot(density_ham, 
         xlim = c(min(density_ham_x[1], density_spam_x[1]), 
                  max(density_ham_x[2], density_spam_x[2])), 
         ylim = c(min(density_ham_y[1], density_spam_y[1]), 
                  max(density_ham_y[2], density_spam_y[2])), 
         main = paste("Density plot of", index), xlab = index, 
         col = "green", )
    lines(density_spam, col = "red")
    legend("topright", legend = c("spam", "ham"), pch = 16, col = c("red", "green"))
  }
}

c("num_character", "num_word", "num_sentence") %>% plot_densities()
RCopy

# sampling
set.seed(0)
train_index = createDataPartition(data$Category, p = 0.7, list = F)
train = data[train_index, ]
test = data[-train_index, ]
RCopy

# transform message to corpus
corpus = Corpus(VectorSource(train$Message))

# processing
corpus = corpus %>% 
  tm_map(content_transformer(tolower)) %>% # text becomes lowercase
  tm_map(removePunctuation) %>% # remove punctuation
  tm_map(removeNumbers) %>% # remove numbers
  tm_map(stripWhitespace) %>% # strip white space
  tm_map(stemDocument) # stemming
  # tm_map(removeWords, stopwords("english")) # remove stop words

# £ to pound_sign, in to intransform, the following word are reserve word in modeling
custom_replace = function(x) {
  x = gsub("£", " num_pound ", x, fixed = TRUE)
  x = gsub("\\bin\\b", "num_in", x, ignore.case = TRUE)
  x = gsub("\\bfor\\b", "num_for", x, ignore.case = TRUE)
  return(x)
}

corpus = corpus %>% tm_map(content_transformer(custom_replace))
RCopy

# number of punctuation
punctuation = data.frame(num_punct = str_count(train$Message, "[[:punct:]]"), 
                         num_exclam = str_count(train$Message, "!"), 
                         num_period = str_count(train$Message, "\\.")) %>% 
  data.matrix()
RCopy

# document to term matrix
dtm = DocumentTermMatrix(corpus, control = list(wordLengths = c(1, Inf)))
dtm_matrix = as.matrix(dtm) 
dtm_matrix = punctuation %>% cbind(dtm_matrix) 
new_train = cbind(train, dtm_matrix)
RCopy

# check the different between ham and spam
non_zero_ham = apply(t(dtm_matrix[which(train$Category == "ham"), ]), 1, function(row) sum(row != 0)) / sum(train$Category == "ham")
non_zero_spam = apply(t(dtm_matrix[which(train$Category == "spam"), ]), 1, function(row) sum(row != 0)) / sum(train$Category == "spam")

diff = (non_zero_ham - non_zero_spam) %>% sort(decreasing = T)

head(diff, 10)
tail(diff, 10)
RCopy

# 處理 test data
test_corpus = Corpus(VectorSource(test$Message))

# 預處理
test_corpus = test_corpus %>% 
  tm_map(content_transformer(tolower)) %>% # 小寫字母
  tm_map(removePunctuation) %>% # 移除標點符號
  tm_map(removeNumbers) %>% # 移除數字
  tm_map(stripWhitespace) %>% # 移除空白
  tm_map(stemDocument) %>% # 詞幹提取 stemming
  tm_map(content_transformer(custom_replace))
  # tm_map(removeWords, stopwords("english")) # 移除無意義字
RCopy

# number of punctuation
test_punctuation = data.frame(num_punct = str_count(test$Message, "[[:punct:]]"), 
                              num_exclam = str_count(test$Message, "!"), 
                              num_period = str_count(test$Message, "\\.")) %>% 
  data.matrix()
RCopy

# document to term matrix
test_dtm = DocumentTermMatrix(test_corpus, control = list(dictionary = Terms(dtm)))
test_dtm_matrix = as.matrix(test_dtm)
test_dtm_matrix = test_punctuation %>% cbind(test_dtm_matrix) 
new_test = cbind(test, test_dtm_matrix)
RCopy

# over sample
oversample = upSample(x = new_train %>% select(-"Category"),
                      y = new_train$Category, 
                      yname = "Category")
RCopy

diff_term = 30

diff_name = diff %>% names() %>% head(diff_term) 
diff_name = diff %>% names() %>% tail(diff_term) %>% c(diff_name)

forumla = paste("Category ~ num_character + num_word + num_sentence + ", paste(diff_name, collapse = " + ")) %>% 
  as.formula()
RCopy

# train control
train_control = trainControl(method = "cv", number = 10)
RCopy

# random forest
set.seed(0)
rf = train(forumla, 
           data = new_train, 
           trControl = train_control, 
           method = "rf")
RCopy

# train performance
predictions = predict(rf, newdata = new_train)
confusionMatrix(new_train$Category, predictions)

# ROC/AUC
predicted_prob = predict(rf, newdata = new_train, type = "prob")
prediction = prediction(predicted_prob[, 2], new_train$Category)
  
# 計算 ROC / AUC
roc = performance(prediction, "tpr", "fpr")
auc = performance(prediction, "auc")
auc = auc@y.values[[1]]

# 繪製 ROC
plot(roc, main = "ROC", col = "blue", lwd = 2)
abline(a = 0, b = 1, lty = 2, col = "red")
legend("bottomright", legend = paste("AUC =", round(auc, 4)), 
       col = "blue", lty = 1, cex = 0.8)
RCopy

# test performance
predictions = predict(rf, newdata = new_test)
confusionMatrix(new_test$Category, predictions)

# ROC/AUC
predicted_prob = predict(rf, newdata = new_test, type = "prob")
prediction = prediction(predicted_prob[, 2], new_test$Category)
  
# 計算 ROC / AUC
roc = performance(prediction, "tpr", "fpr")
auc = performance(prediction, "auc")
auc = auc@y.values[[1]]

# 繪製 ROC
plot(roc, main = "ROC", col = "blue", lwd = 2)
abline(a = 0, b = 1, lty = 2, col = "red")
legend("bottomright", legend = paste("AUC =", round(auc, 4)), 
       col = "blue", lty = 1, cex = 0.8)
RCopy

nb = multinomial_naive_bayes(x = new_train %>% dplyr::select(-c("Category", "Message")), 
                             y = new_train$Category)
nb
RCopy

# train performance
train_nb_trans = new_train %>% 
  dplyr::select(-c("Category", "Message")) %>%
  as.matrix()
predictions = predict(nb, newdata = train_nb_trans)
confusionMatrix(new_train$Category, predictions)

# ROC/AUC
predicted_prob = predict(nb, newdata = train_nb_trans, type = "prob")
prediction = prediction(predicted_prob[, 2], new_train$Category)
  
# 計算 ROC / AUC
roc = performance(prediction, "tpr", "fpr")
auc = performance(prediction, "auc")
auc = auc@y.values[[1]]

# 繪製 ROC
plot(roc, main = "ROC", col = "blue", lwd = 2)
abline(a = 0, b = 1, lty = 2, col = "red")
legend("bottomright", legend = paste("AUC =", round(auc, 4)), 
       col = "blue", lty = 1, cex = 0.8)
RCopy

test_nb_trans = new_test %>% 
  dplyr::select(-c("Category", "Message")) %>%
  as.matrix()

# test performance
predictions = predict(nb, newdata = test_nb_trans)
confusionMatrix(new_test$Category, predictions)

# ROC/AUC
predicted_prob = predict(nb, newdata = test_nb_trans, type = "prob")
prediction = prediction(predicted_prob[, 2], new_test$Category)
  
# 計算 ROC / AUC
roc = performance(prediction, "tpr", "fpr")
auc = performance(prediction, "auc")
auc = auc@y.values[[1]]

# 繪製 ROC
plot(roc, main = "ROC", col = "blue", lwd = 2)
abline(a = 0, b = 1, lty = 2, col = "red")
legend("bottomright", legend = paste("AUC =", round(auc, 4)), 
       col = "blue", lty = 1, cex = 0.8)
RCopy