In dataknut/hashTagR: Gets, saves and processes tweets and hashtags

knitr::opts_chunk$set(echo = FALSE) # do not echo code
knitr::opts_chunk$set(warning = FALSE)
knitr::opts_chunk$set(message = FALSE)
knitr::opts_chunk$set(fig_caption = TRUE)
knitr::opts_chunk$set(fig_height = 6) # default, make it bigger to stretch vertical axis
knitr::opts_chunk$set(fig_width = 8) # full width
knitr::opts_chunk$set(tidy = TRUE) # tidy up code in case echo = TRUE

# Set start time ----
startTime <- proc.time()

# Libraries ----

# additional libs required by this code
reqLibs <- c("data.table",
             "ggplot2",  
             "plotly",
             "readr", 
             "reshape2",
             "stringr",
             "knitr", 
             "kableExtra"
             )

print(paste0("Loading the following libraries: ", reqLibs))
# Use Luke's function to require/install/load
myUtils::loadLibraries(reqLibs)

dPath <- "~/Data/twitter/"

TL;DR

Just a bit of dataknut fun woven around the day job.

You'll be wanting Section \@ref(cumulative) for the trending hashtags...

Terms of re-use

License

CC-BY unless otherwise noted.

Citation

• Anderson, B (r lubridate::year(today())) Twitter hashtag analysis: r params$hashtag. Downloaded from: r params$pubUrl.

Code

See:

• Code: https://github.com/dataknut/hashTagR/blob/master/analysis/birdOfTheYear2018.Rmd
• Code/report history
• Raise an issue here

Purpose

I've not been in NZ at this time of year before so r params$hashtag is a whole new cultural experience.

The idea is to extract and visualise tweets and re-tweets of r params$hashtag (see r params$explHashTag and the Forest & Bird voting site).

Why? Err.... Just. Because.

How it works

Code borrows extensively from https://github.com/mkearney/rtweet

The analysis used rtweet to ask the Twitter search API to extract 'all' tweets containing the r params$hashtag hashtags in the 'recent' twitterVerse.

It is therefore possible that not quite all tweets have been extracted although it seems likely that we have captured most recent human tweeting which was the main intention. Future work should instead use the Twitter streaming API.

# load from pre-collected ----
message("Load from pre-collected data and check for duplicates")
# https://github.com/dataknut/hashTagR/blob/master/dataProcessing/getBoTY.R
# for testing params$hashtag <- "#birdoftheyear OR #boty"
raw_twDT <- hashTagR::loadTweets(dPath, params$hashtag) # we like data.tables

# this data.table will have duplicate entries because:
# 1. we may have run the search mutiple times
# 2. the search results are dynamic - they can change if a tweet is liked, re-tweeted etc as this data is updated

# we don't want duplicates, we just want the most recent tweet record by time of creation and screen_names

rn <- nrow(raw_twDT)
twDT <- unique(raw_twDT, fromLast = TRUE, by = c("created_at", "screen_name") ) # drop duplicates
un <- nrow(twDT)
message("Returning ", myUtils::tidyNum(un),
        " tweets after dropping ",
        myUtils::tidyNum(rn - un),
        " duplicates.")

twDT <- twDT[, ba_obsDate := lubridate::date(created_at)]
twDT <- twDT[, ba_obsTime := hms::as.hms(created_at)] # this will auto-convert to local time

The data has:

• r tidyNum(nrow(twDT)) tweets (including r tidyNum(nrow(twDT[is_quote == "TRUE"])) quotes and r tidyNum(nrow(twDT[is_retweet == "TRUE"])) re-tweets)
• from r tidyNum(uniqueN(twDT$screen_name)) tweeters
• between r min(twDT$created_at) and r max(twDT$created_at) (UTC).

Analysis

Tweets and Tweeters over time

dataCap <- paste0("Source: Data collected from Twitter's REST API via rtweet",
                  "\nAll (re)tweets and quotes containing ", params$hashtag, 
                    " from ",
                      min(twDT$created_at),
                    " to ",
                    max(twDT$created_at)
                    )

twDT <- twDT[, ba_tweetType := "Tweet"]
twDT <- twDT[is_retweet == TRUE, ba_tweetType := "Re-tweet"]
twDT <- twDT[is_quote == TRUE, ba_tweetType := "Quote"]

plotDT <- twDT[, .(
                 nTweets = .N,
                 nTweeters = uniqueN(screen_name)
               ), keyby = .(ba_obsDate, ba_tweetType)]

  myPlot <- ggplot(plotDT, aes(x = ba_obsDate)) +
    geom_line(aes(y = nTweets, colour = "N tweets")) +
    geom_line(aes(y = nTweeters, colour = "N tweeters")) +
    facet_grid(ba_tweetType ~ .) +
    scale_x_date(date_breaks = "1 day", date_labels = "%a %d %b %Y") +
    theme(strip.text.y = element_text(size = 9, colour = "black", angle = 90)) +
    theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust = 0.5)) +
    theme(legend.position = "bottom") +
    theme(legend.title = element_blank()) +
    labs(caption = dataCap,
         x = "Date",
         y = "Count"
    )

myPlot

Figure \@ref(fig:allDaysChart) shows the number of tweets and tweeters in the data extract by day. The quotes, tweets and re-tweets have been separated.

If you are in New Zealand and you are wondering why there are no tweets today (r lubridate::today()) the answer is that twitter data (and these plots) are working in UTC and (y)our today() may not have started yet in UTC. Don't worry, all the tweets are here - it's just our old friend the timezone... :-)

Who's tweeting?

Next we'll try by screen name.

plotDT <- twDT[, 
                    .(
                      nTweets = .N
                    ), by = .(screen_name, ba_obsDate)]

tilePlot <- ggplot(plotDT, aes(x = ba_obsDate)) +
    geom_tile(aes(y = screen_name, fill = nTweets)) +
    theme(strip.text.y = element_text(size = 9, colour = "black", angle = 0)) +
    theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust = 0.5)) +
    #scale_x_reverse() + # fix reverse plotting of long
    scale_x_date(date_breaks = "1 day", date_labels = "%a %d %b %Y") +
    scale_fill_gradient(low="green", high = "red") +
    theme(legend.position = "bottom") +
    theme(legend.title = element_blank()) +
    labs(caption = dataCap,
         x = "Date",
         y = "Screen name"
    )
tilePlot

Figure \@ref(fig:screenNamesTile) is a really bad visualisation of all tweeters tweeting over time. Each row of pixels is a tweeter (the names are probably illegible) and a green dot indicates a few tweets in the given day while a red dot indicates a lot of tweets.

So let's re-do that for the top 50 tweeters so we can see their tweetStreaks (tm)...

Top tweeters:

allTweetersDT <- twDT[, .(nTweets = .N), by = screen_name][order(-nTweets)]

kableExtra::kable(caption = "Top 15 tweeters (all days)", 
                  head(allTweetersDT, 15)) %>%
  kable_styling()

And their tweetStreaks are shown in Figure \@ref(fig:screenNameTop50)...

myDataCap <- paste0(dataCap,
                    "\nScreen names in reverse alphabetical order"
                          )

matchDT <- head(allTweetersDT,50)
matchDT <- matchDT[, maxT := nTweets]
setkey(matchDT, screen_name)
setkey(twDT, screen_name)

tempDT <- merge(twDT, matchDT)

plotDT <- tempDT[matchDT, 
                    .(
                      nTweets = .N
                    ), keyby = .(maxT,screen_name,ba_obsDate)]

plotDT <- plotDT[order(plotDT$maxT,plotDT$screen_name)]

myPlot <- ggplot(plotDT, aes(x = ba_obsDate)) +
    geom_tile(aes(y = screen_name, fill = nTweets)) +
    theme(strip.text.y = element_text(size = 9, colour = "black", angle = 0)) +
    theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust = 0.5)) +
    scale_x_date(date_breaks = "1 day", date_labels = "%a %d %b %Y") +
    scale_fill_gradient(low="green", high = "red") +
    theme(legend.position = "bottom") +
    theme(legend.title = element_blank()) +
    labs(caption = myDataCap,
         x = "Date",
         y = "Screen name"
    )

myPlot 

Any twitterBots...?

Which birds are mentioned the most (by hashtag)

This is very quick and dirty but... to calculate this we have to do a bit of string processing first.

htDT <- twDT[!is.na(hashtags), .(hashtags, ba_obsDate, ba_tweetType)] # remove any tweets without hashtags. How can there be no hashtags when we searched on hashtags?

# now string split them
# https://stackoverflow.com/questions/33200179/dynamically-assign-number-of-splits-in-data-table-tstrsplit
splits <- max(lengths(strsplit(htDT$hashtags, "|", , fixed=T)))
htDT <- htDT[, paste0("ht", 1:splits) := tstrsplit(hashtags, "|", fixed=T)]
# reshape the list
htLongDT <- reshape2::melt(htDT, id=c("hashtags","ba_obsDate", "ba_tweetType"))
# remove NA
htLongDT <- htLongDT[!is.na(value)]
message("We have ", nrow(htLongDT), " hashtags.")
message("That's about ~ ", round(nrow(htLongDT)/nrow(twDT),2), " hashtags per tweet...")
# process in steps to be clear
htLongDT <- htLongDT[, htOrig := value]

This is how I have tidied the hashtags (make other suggestions here):

# First we make everything lower case
htLongDT <- htLongDT[, htLower := tolower(htOrig)] # lower case

# Next we remove the macrons so that TeamKakī == TeamKaki & takahe == takahē etc 
# h/t: https://twitter.com/Thoughtfulnz/status/1046685305569345536
htLongDT <- htLongDT[, htClean := stringr::str_replace_all(htLower,"[āēīōū]",myUtils::deMacron)]

# Now remove 'team' from a string so that e.g. teamkaki == kaki
htLongDT <- htLongDT[, htClean := gsub("team", "",htClean)]

# Now remove variants on 'vote'
htLongDT <- htLongDT[, htClean := gsub("vote4", "",htClean)]
htLongDT <- htLongDT[, htClean := gsub("vote", "",htClean)]

Table \@ref(tab:tweetTable) shows the total count of each #hashtag by (re)tweet type. With thanks to David Hood for code to help make sure that kakī == kaki (etc).

kableExtra::kable(caption = "Top 20 hashtags in tweets containing #birdoftheyear or #boty",
                  head(htLongDT[, .(count = .N), by = .(hashTag = htClean, type = ba_tweetType)][order(-count)],20)) %>%
  kable_styling()

threshold <- 10

Figure \@ref(fig:plotHashtags) plots the daily occurence of these hashtags after removing variants of #birdOfTheYear and #boty and selecting only those which have more than r threshold mentions on any day. For clarity tweets and re-tweets are aggregated. See Section \@ref(problems) for the problems with this #hashTag counting approach.

# remove any which are _only_ birdoftheyear
testDT <- htLongDT[!(grepl("birdoftheyear",htClean,ignore.case = TRUE) | 
                grepl("boty",htClean,ignore.case = TRUE) )]

htTabDT <- testDT[, .(count = .N), keyby = .(ba_obsDate, htClean, ba_tweetType)]

ggplot2::ggplot(htTabDT[count > threshold], aes(x = ba_obsDate, y = htClean, fill = count)) +
  geom_tile() + 
  theme(strip.text.y = element_text(size = 9, colour = "black", angle = 0)) +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust = 0.5)) +
  scale_x_date(date_breaks = "1 day", date_labels = "%a %d %b %Y") +
  scale_fill_gradient(low="green", high = "red") +
 # facet_grid(ba_tweetType ~ .) +
  theme(legend.position = "bottom") +
  theme(legend.title = element_blank()) +
  labs(caption = paste0(dataCap,"\nReverse alphabetical order"),
       x = "Date",
       y = "Unique hashtags"
  )

So who's gonna win? {#cumulative}

No idea. Your vote is as good as mine :-)

There are a lot of problems with this approach (see Section \@ref(problems)) but if the hashtags have any predictive value at all then Figure \@ref(fig:cumulativeHtCounts) should be an indicator of the direction of travel (watch for lines of apparently dis-similar hashtags where the macron fix has failed) and Figure \@ref(fig:htCountsTodate) shows the totals to date.

Figure \@ref(fig:cumulativeHtCounts) uses plotly to avoid having to render a large legend - just hover over the lines to see who is who...

plotDT <- htTabDT[, .(count = sum(count)), keyby = .(ba_obsDate,htClean)]
plotDT <- plotDT[, cumCount := cumsum(count), by = htClean]
plotDT <- plotDT[, maxCount := max(cumCount), by = htClean] # so we can filter

linePlot <- ggplot2::ggplot(plotDT[maxCount > 30], aes(x = ba_obsDate, y = cumCount, colour = htClean)) +
  geom_line() + 
  theme(strip.text.y = element_text(size = 9, colour = "black", angle = 0)) +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust = 0.5)) +
  scale_x_date(date_breaks = "1 day", date_labels = "%a %d %b %Y") +
  theme(legend.position = "none") +
  theme(legend.title = element_blank()) +
  labs(caption = paste0(dataCap),
       x = "Date",
       y = "Cumulative count per day"
  )

plotly::ggplotly(linePlot)

plotDT <- htTabDT[, .(count = sum(count)), keyby = .(htClean)]

plotDT <- plotDT[, htCleanOr := reorder(htClean, -count)]

ggplot2::ggplot(plotDT[count > 30], aes(x = htCleanOr, y = count, fill = htCleanOr)) +
  geom_col() + 
  theme(strip.text.y = element_text(size = 9, colour = "black", angle = 0)) +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust = 0.5)) +
  theme(legend.position = "none") +
  theme(legend.title = element_blank()) +
  labs(caption = paste0(dataCap,"\nOnly count > 30 shown"),
       x = "hashtag",
       y = "Count to date"
  )

Problems {#problems}

Loads of them. But primarily:

• hashTags as indicators:
  • tweets mentioning birds without using a #<birdName> hashtag (like this rather popular one) will not show up. #YMMV;
  • there is no sentiment analysis so tweets saying "vote #takahe and the gannet gets it" look the same as "vote #takehe or the gannet gets it";
• removing the macrons: the tables and plots display incorrect Māori names. Unfortunately I have not found a way to put them back in the right words (and only the right words).

About

t <- proc.time() - startTime

elapsed <- t[[3]]

Analysis completed in r elapsed seconds ( r round(elapsed/60,2) minutes) using knitr in RStudio with r R.version.string running on r R.version$platform.

A special mention must go to https://github.com/mkearney/rtweet [@rtweet-package] for the twitter API interaction functions.

Other R packages used:

• base R - for the basics [@baseR]
• data.table - for fast (big) data handling [@data.table]
• ggplot2 - for slick graphs [@ggplot2]
• plotly - for interactive graphs [@plotly]
• readr - for nice data loading [@readr]
• reshape2 - wide to long [@reshape2]
• rtweet - twitter API search [@rtweet-package]
• stringr - processing strings [@stringr]
• knitr, bookdown & kableExtra - to create this document [@knitr,@bookdown, @kableExtra]

References

dataknut/hashTagR documentation built on Sept. 20, 2023, 7:24 a.m.