1 트윗 데이터셋 ¹

tidytext와 caret을 활용하여 자연어 텍스트 분류기를 제작하여 본다. 이를 위해서 Disasters on Social Media에 올라온 데이터를 바탕으로 예측모형 개발을 시작한다. 더불어 figure eight 웹사이트에 흥미로운 데이터셋도 많이 올라와 있다.

library(tidyverse)
library(tidytext)
library(caret)

raw_dat <- read_csv("data/socialmedia-disaster-tweets-DFE.csv")

raw_dat %>% 
  count(choose_one)

# A tibble: 3 x 2
  choose_one       n
  <chr>        <int>
1 Can't Decide    16
2 Not Relevant  6187
3 Relevant      4673

clean_df <- raw_dat %>%
  filter(choose_one != "Can't Decide") %>%
  mutate(id = `_unit_id`,
         disaster = choose_one == "Relevant",
         text = str_replace_all(text, " ?(f|ht)tp(s?)://(.*)[.][a-z]+", "")) %>%
  select(id, disaster, text)

clean_df %>%
  count(disaster) %>% 
  mutate(pcnt = n / sum(n))

# A tibble: 2 x 3
  disaster     n  pcnt
  <lgl>    <int> <dbl>
1 FALSE     6187 0.570
2 TRUE      4673 0.430

2 텍스트 데이터 전처리

count_df <- clean_df %>% 
  unnest_tokens(word, text) %>% 
  anti_join(stop_words, by = "word") %>%
  count(id, word, sort = TRUE)

words_10 <- count_df %>% 
  group_by(word) %>% 
  summarise(n = n()) %>% 
  filter(n >= 10) %>%
  select(word)

dtm_df <- count_df %>%
  right_join(words_10, by = "word") %>%
  bind_tf_idf(word, id, n) %>%
  cast_dtm(id, word, tf_idf)

meta <- tibble(id = as.numeric(dimnames(dtm_df)[[1]])) %>%
  left_join(clean_df[!duplicated(clean_df$id), ], by = "id")

3 caret 예측모형

caret 팩키지를 활용하여 훈련/시험 데이터로 분할시키고 나서 설명변수와 예측변수로 나누는 작업을 수행하고 3가지 모형 아키텍처에 적합시켜 정확도를 기준으로 가장 성능이 좋은 예측모형을 선정한다.

library(caret)
library(extrafont)
loadfonts()

# 2. 예측모형 -----
## 2.1. 훈련/시험 데이터 분할 ------

y_index <- createDataPartition(meta$disaster, times =1, p=0.7, list=FALSE)

train_df <- dtm_df[y_index, ]  %>% as.matrix() %>% as.data.frame()
test_df  <- dtm_df[-y_index, ] %>% as.matrix() %>% as.data.frame()

response_train <- meta$disaster[y_index]

## 2.2. 모형 개발/검증 데이터셋 준비 ------

trctrl <- trainControl(method = "none")

## 2.2. 모형 개발/검증 데이터셋 준비 ------

library(doSNOW)
library(tictoc)
# 실행시간
tic()

cl <- makeCluster(4, type = "SOCK")
registerDoSNOW(cl)

svm_mod <- train(x = train_df,
                 y = as.factor(response_train),
                 method = "svmLinearWeights2",
                 trControl = trctrl,
                 tuneGrid = data.frame(cost = 1, 
                                       Loss = 0, 
                                       weight = 1))


nb_mod <- train(x = train_df,
                y = as.factor(response_train),
                method = "naive_bayes",
                trControl = trctrl,
                tuneGrid = data.frame(laplace = 0,
                                      usekernel = FALSE,
                                      adjust = FALSE))

rf_mod <- train(x = train_df, 
                y = as.factor(response_train), 
                method = "ranger",
                trControl = trctrl,
                tuneGrid = data.frame(mtry = floor(sqrt(dim(train_df)[2])),
                                      splitrule = "gini",
                                      min.node.size = 1))

Growing trees.. Progress: 51%. Estimated remaining time: 30 seconds.

stopCluster(cl)
toc()

65.87 sec elapsed

# 4. 모형 비교평가-----
svm_pred <- predict(svm_mod, newdata = test_df)
svm_cm <- confusionMatrix(svm_pred, as.factor(meta[-y_index, ]$disaster))

nb_pred <- predict(nb_mod, newdata = test_df)
nb_cm <- confusionMatrix(nb_pred, as.factor(meta[-y_index, ]$disaster))

rf_pred <- predict(rf_mod, newdata = test_df)
rf_cm <- confusionMatrix(rf_pred, as.factor(meta[-y_index, ]$disaster))

mod_results <- rbind(svm_cm$overall, 
                     nb_cm$overall,
                     rf_cm$overall) %>%
  as.data.frame() %>%
  mutate(model = c("SVM", "Naive-Bayes", "Random forest"))

mod_results %>%
  ggplot(aes(model, Accuracy)) +
    geom_point() +
    scale_y_continuous(labels = scales::percent, limits=c(0,1)) +
    geom_hline(yintercept = mod_results$AccuracyNull[1], color = "red") +
    labs(x="", y="정확도") +
    theme_minimal(base_family = "NanumGothic")