# Packages ----------------------------------------------------------------

# install.packages("tidyverse")
library(tidyverse)

# install.packages("tidytext")
library(tidytext)

# install.packages("rsample")
library(rsample)

# install.packages("glmnet")
library(glmnet)

# install.packages("yardstick")
library(yardstick)

# Data preparation --------------------------------------------------------

imdb <- read_csv("C:/Users/Bootcha/Desktop/skola/TM/2019/3/IMDB_Dataset.csv")

imdb <- imdb %>% 
  mutate(id = row_number(),
         review = str_to_lower(review),
         review = str_replace_all(review, "<br />", ""),
         review = str_replace_all(review, "\"", "")) %>% 
  head(10000)


# Data exploration --------------------------------------------------------

tidy_imdb <- imdb %>%
  unnest_tokens(word, review) %>%
  group_by(word) %>%
  filter(n() > 10) %>%
  ungroup()

tidy_imdb

tidy_imdb %>% 
  count(sentiment, word, sort=TRUE) %>% 
  anti_join(get_stopwords()) %>% 
  group_by(sentiment) %>% 
  top_n(20) %>% 
  ungroup() %>% 
  ggplot(aes(x=reorder_within(word, n, sentiment), y=n, fill=sentiment)) +
  geom_col(show.legend=TRUE) +
  scale_x_reordered() +
  coord_flip() +
  facet_wrap(~sentiment, scales="free") +
  scale_y_continuous(expand=c(0, 0)) +
  labs(
    x=NULL,
    y="Word count",
    title="Most frequent words after removing stop words"
  )

# Model -------------------------------------------------------------------

imdb_split <- imdb %>%
  select(id) %>%
  initial_split()

train_data <- training(imdb_split)
test_data <- testing(imdb_split)

sparse_words <- tidy_imdb %>%
  count(id, word) %>%
  inner_join(train_data) %>%
  cast_sparse(id, word, n)

class(sparse_words)

dim(sparse_words)

word_rownames <- as.integer(rownames(sparse_words))

imdb_joined <- data_frame(id = word_rownames) %>%
  left_join(imdb %>% select(id, sentiment))

is_positive <- imdb_joined$sentiment == "positive"
model <- cv.glmnet(sparse_words, is_positive, family="binomial", keep=TRUE)
# model <- cv.glmnet(sparse_words, is_positive, family="binomial", keep=TRUE, type.measure="class")

plot(model)

plot(model$glmnet.fit, label = TRUE)

coefs <- model$glmnet.fit %>%
  tidy() %>%
  filter(lambda == model$lambda.1se)

coefs %>%
  group_by(estimate > 0) %>%
  top_n(10, abs(estimate)) %>%
  ungroup() %>%
  ggplot(aes(fct_reorder(term, estimate), estimate, fill=estimate > 0)) +
  geom_col(alpha=0.8, show.legend=FALSE) +
  coord_flip() +
  labs(
    x = NULL,
    title = "Coefficients that increase/decrease probability the most")

intercept <- coefs %>%
  filter(term=="(Intercept)") %>%
  pull(estimate)

classifications <- tidy_imdb %>%
  inner_join(test_data) %>%
  inner_join(coefs, by=c("word" = "term")) %>%
  group_by(id) %>%
  summarize(score=sum(estimate)) %>%
  mutate(probability=plogis(intercept+score))

classifications

comment_classes <- classifications %>%
  left_join(imdb %>% select(sentiment, id), by="id") %>%
  mutate(sentiment=as.factor(sentiment))

comment_classes %>%
  roc_curve(sentiment, probability) %>%
  ggplot(aes(x=1-specificity, y=sensitivity)) +
  geom_line(color="midnightblue", size=1.5) +
  geom_abline(lty=2, alpha=0.5, color="gray50", size=1.2) +
  labs(
    title = "ROC curve for text classification using regularized regression"
  )

comment_classes %>%
  roc_auc(sentiment, probability)

comment_classes %>%
  mutate(
    prediction = case_when(
      probability > 0.5 ~ "positive",
      TRUE ~ "negative"),
    prediction = as.factor(prediction)) %>%
  conf_mat(sentiment, prediction)

comment_classes %>%
  filter(probability > .8, sentiment == "negative") %>%
  sample_n(10) %>%
  inner_join(imdb %>% select(id, review)) %>%
  select(probability, review)
  
comment_classes %>%
  filter(probability < .3, sentiment == "positive") %>%
  sample_n(10) %>%
  inner_join(imdb %>% select(id, review)) %>%
  select(probability, review)