Can You Guess These Presidential Candidates Based On Their Most Popular Talking Points?

Data

We gathered data on tweets from the campaign accounts of 23 Democrats running for president, for the previous 90 days or since their campaign accouncement, if more recent, using using the rtweet R package to access the Twitter API. (To run this code, you will need to authorize for your own Twitter API access tokens/)

# required packages
library(rtweet)
library(readr)
library(dplyr)
library(lubridate)
library(DT)

# Set default timezone for session to UTC
Sys.setenv(TZ = "UTC")

# load campaign account data
candidates <- read_csv("campaign_accounts.csv")

# get tweets
tweets <- tibble()

for (s in candidates$screen_name) {
  tmp <- get_timeline(s, n = 3200)
  tweets <- bind_rows(tweets, tmp)
}

tweets <- inner_join(tweets, candidates) %>%
  mutate(est_timestamp = with_tz(created_at, "America/New_York"),
         est_date = date(est_timestamp)) %>%
  filter(est_date >= announced & est_date >= today(tzone = "America/New_York")-90)

rm(tmp, s)

# summary data
summary <- tweets %>%
  group_by(screen_name, announced) %>%
  summarize(first = min(est_date),
            tweets = n())

datatable(summary)

Setting up for text analysis

# load required packages
library(tidytext)
library(tidyr)
library(stringr)
library(wordcloud)

# palettes for wordclouds
dem_pal <- c("#47B5FF", "#1482EE", "#004FBB")

# regex for parsing tweets using the tidytext package
replace_reg <- "https?://[^\\s]+|&amp;|&lt;|&gt;|\\bRT\\b"

Tokenize tweets to bigrams

We tokenized original tweets only, excluding retweets, filtering for common “stop words” plus a custom set of words including the names of candidates or their relatives, distinctive home towns, and hashtags, and in the case of Tulsi Gabbard, the word “aloha.”

# tokenize into bigrams
bigrams <- tweets %>%
  filter(is_retweet == FALSE) %>%
  select(screen_name,text,full_name) %>%
  mutate(text = str_replace_all(text, replace_reg, "")) %>%
  unnest_tokens(bigram, text, token = "ngrams", n = 2) %>%
  separate(bigram, into = c("first","second"), sep = " ", remove = FALSE) %>%
  anti_join(stop_words, by = c("first" = "word")) %>%
  anti_join(stop_words, by = c("second" = "word")) %>%
  filter(str_detect(first, "[a-z]") &
         str_detect(second, "[a-z]"))

# filter for custom stop words
bigrams <- bigrams %>%
  filter(!grepl("bennet|biden|booker|buttigieg|pete|bullock|castro|julian|deblasio|blasio|delaney|gabbard|aloha|tulsi|gillibrand|harris|hickenlooper|inslee|klobuchar|messam|wayne|moulton|o'rourke|orourke|beto|ryan|tim|sanders|swalwell|warren|williamson|yang|chirlane|caroline|san antonio|south bend|el paso",bigram))

Bigram clouds for each candidate

bigram_count <- bigrams %>%
  group_by(bigram,full_name) %>%
  count() %>%
  ungroup()
  
for (p in unique(bigram_count$full_name)) {
  tmp <- bigram_count %>%
    filter(full_name == p)
  png(paste0("wordclouds/",p,".png"), width = 600, height = 500)
  try(wordcloud(tmp$bigram,
                tmp$n,
                family = "BasierSquare-SemiBold",
                max.words = 50, 
                random.order = FALSE,
                colors = dem_pal,
                rot.per = 0))
  dev.off()
}