Social Community Evaluation With R: Mining for Twitter Clusters


That is the ultimate installment in a three-part sequence on Twitter cluster analyses utilizing R and Gephi. Half one analyzed heated on-line dialogue about famed Argentine footballer Lionel Messi; half two deepened the evaluation to higher determine principal actors and perceive subject unfold.

Politics are polarizing. Once we discover attention-grabbing communities with drastically totally different opinions, Twitter messages generated from inside these camps are likely to densely cluster round two teams of customers, with a slight connection between them. This kind of grouping and relationship is known as homophily: the tendency to work together with these much like us.

Within the earlier article on this sequence, we targeted on computational strategies primarily based on Twitter information units and have been in a position to generate informative visualizations by Gephi. Now we wish to use cluster evaluation to know the conclusions we are able to draw from these strategies and determine which social information features are most informative.

We’ll change the sort of information we analyze to spotlight this clustering, downloading United States’ political information from Could 10, 2020, by Could 20, 2020. We’ll use the identical Twitter information obtain course of we used within the first article on this sequence, altering the obtain standards to the then-president’s title slightly than “Messi.”

The next determine depicts the interplay graph of the political dialogue; as we did within the first article, we plotted this information with Gephi utilizing the ForceAtlas2 structure and coloured by the communities as detected by Louvain.

A non-identified binary data cluster interaction graph generated within Gephi
Information Cluster Interplay Graph

Let’s dive deeper into the obtainable information.

Who Are in These Clusters?

As we’ve mentioned all through this sequence, we are able to characterize clusters by their authorities, however Twitter offers us much more information that we are able to parse. For instance, the consumer’s description discipline, the place Twitter customers can present a quick autobiography. Utilizing a phrase cloud, we are able to uncover how customers describe themselves. This code generates two phrase clouds primarily based on the phrase frequency discovered throughout the information in every cluster’s descriptions and highlights how individuals’s self-descriptions are informative in an mixture manner:

# Load vital libraries

# First, determine the communities by Louvain = cluster_louvain(as.undirected(simplify(web)),decision=0.4)

# Subsequent, get the customers that conform to the 2 largest clusters
largestCommunities <- order(sizes(, lowering=TRUE)[1:4]
community1 <- names(which(membership( == largestCommunities[1]))
community2 <- names(which(membership( == largestCommunities[2]))

# Now, break up the tweets’ information frames by their communities
# (i.e., 'republicans' and 'democrats')

republicans = tweets.df[which(tweets.df$screen_name %in% community1),]
democrats = tweets.df[which(tweets.df$screen_name %in% community2),]

# Subsequent, provided that we now have one row per tweet and we wish to analyze customers, 
# let’s hold just one row by consumer
accounts_r = republicans[!duplicated(republicans[,c('screen_name')]),]
accounts_d = democrats[!duplicated(democrats[,c('screen_name')]),]

# Lastly, plot the phrase clouds of the consumer’s descriptions by cluster

## Generate the Republican phrase cloud
## First, convert descriptions to tm corpus
corpus <- Corpus(VectorSource(distinctive(accounts_r$description)))

### Take away English cease phrases
corpus <- tm_map(corpus, removeWords, stopwords("en"))

### Take away numbers as a result of they don't seem to be significant at this step
corpus <- tm_map(corpus, removeNumbers)

### Plot the phrase cloud displaying a most of 30 phrases
### Additionally, filter out phrases that seem solely as soon as
pal <- brewer.pal(8, "Dark2")
wordcloud(corpus, min.freq=2, max.phrases = 30, random.order = TRUE, col = pal)

## Generate the Democratic phrase cloud

corpus <- Corpus(VectorSource(distinctive(accounts_d$description))) 
corpus <- tm_map(corpus, removeWords, stopwords("en"))
pal <- brewer.pal(8, "Dark2")
wordcloud(corpus, min.freq=2, max.phrases = 30, random.order = TRUE, col = pal)

Information from earlier US elections reveals that voters are extremely segregated by geographical area. Let’s deepen our id evaluation and give attention to one other discipline: place_name, the sphere the place customers can present the place they dwell. This R code generates phrase clouds primarily based on this discipline:

# Convert place names to tm corpus corpus <- Corpus(VectorSource(accounts_d[!$place_name),]$place_name))

# Take away English cease phrases
corpus <- tm_map(corpus, removeWords, stopwords("en"))

# Take away numbers
corpus <- tm_map(corpus, removeNumbers)

# Plot
pal <- brewer.pal(8, "Dark2")
wordcloud(corpus, min.freq=2, max.phrases = 30, random.order = TRUE, col = pal)

## Do the identical for accounts_r

The RStudio-generated word clouds for each data cluster
Phrase Clouds

The names of some locations might seem in each phrase clouds as a result of voters in each events dwell in most places. However some states, like Texas, Colorado, Oklahoma, and Indiana, strongly signify the Republican occasion whereas some cities, like New York, San Francisco, and Philadelphia, strongly correlate with the Democratic occasion.


Let’s discover one other aspect of the information, specializing in consumer habits and analyzing the distribution of when accounts inside every cluster have been created. If there isn’t any correlation between the creation date and the cluster, we are going to see a uniform distribution of customers for every day.

Let’s plot a histogram of the distribution:

# First we have to format the account date discipline to be successfully learn as Date
## Notice that we're utilizing the accounts_r and accounts_d information body, it's because we wish to give attention to distinctive customers and don’t distort the plot by the variety of tweets that every consumer has submitted

accounts_r$date_account <- as.Date(format(as.POSIXct(accounts_r$account_created_at,format="%Y-%m-%d %H:%M:%S"),format="%Y-%m-%d"))

# Now we plot the histogram
ggplot(accounts_r, aes(date_account)) + geom_histogram(stat="depend")+scale_x_date(date_breaks = "1 yr", date_labels = "%b %Y") 

## Do the identical for accounts_d

A histogram generated with RStudio showing the number of Republican users created for each date within the data set
Variety of Republican Customers Created by Date

A histogram generated with RStudio showing the number of Democrat users created for each date within the data set
Variety of Democratic Customers Created by Date

We see that Republican and Democratic customers are usually not distributed uniformly. In each instances, the variety of new consumer accounts peaked in January 2009 and January 2017, each months when inaugurations occurred following presidential elections within the Novembers of the earlier years. May or not it’s that the proximity to these occasions generates a rise in political dedication? That may make sense, provided that we’re analyzing political tweets.

Additionally attention-grabbing to notice: The largest peak throughout the Republican information happens after the center of 2019, reaching its highest worth in early 2020. May this variation in habits be associated to digital habits introduced on by the pandemic?

The information for the Democrats additionally had a spike throughout this era however with a decrease worth. Possibly Republican supporters exhibited the next peak as a result of they’d stronger opinions about COVID lockdowns? We’d must rely extra on political information, theories, and findings to develop higher hypotheses, however regardless, there are attention-grabbing information developments we are able to analyze from a political perspective.

One other approach to evaluate behaviors is to research how customers retweet and reply. When customers retweet, they unfold a message; nonetheless, once they reply, they contribute to a selected dialog or debate. Usually, the variety of replies correlates to a tweet’s diploma of divisiveness, unpopularity, or controversy; a consumer who favorites a tweet signifies settlement with the sentiment. Let’s study the ratio measure between the favorites and replies of a tweet.

Based mostly on homophily, we might anticipate customers to retweet customers from the identical neighborhood. We are able to confirm this with R:

# Get customers who've been retweeted by either side
rt_d = democrats[which(!$retweet_screen_name)),]
rt_r = republicans[which(!$retweet_screen_name)),]

# Retweets from democrats to republicans
rt_d_unique = rt_d[!duplicated(rt_d[,c('retweet_screen_name')]),]
rt_dem_to_rep = dim(rt_d_unique[which(rt_d_unique$retweet_screen_name %in% unique(republicans$screen_name)),])[1]/dim(rt_d_unique)[1]

# Retweets from democrats to democrats

rt_dem_to_dem = dim(rt_d_unique[which(rt_d_unique$retweet_screen_name %in% unique(democrats$screen_name)),])[1]/dim(rt_d_unique)[1]

# The rest
relaxation = 1 - rt_dem_to_dem - rt_dem_to_rep

# Create a dataframe to make the plot
information <- information.body(
  class=c( "Democrats","Republicans","Others"),
# Compute percentages
information$fraction <- information$depend / sum(information$depend)

# Compute the cumulative percentages (high of every rectangle)
information$ymax <- cumsum(information$fraction)

# Compute the underside of every rectangle
information$ymin <- c(0, head(information$ymax, n=-1))

# Compute label place
information$labelPosition <- (information$ymax + information$ymin) / 2

# Compute a very good label
information$label <- paste0(information$class, "n ", information$depend)

# Make the plot

ggplot(information, aes(ymax=ymax, ymin=ymin, xmax=4, xmin=3, fill=c('purple','blue','inexperienced'))) +
  geom_rect() +
  geom_text( x=1, aes(y=labelPosition, label=label, colour=c('purple','blue','inexperienced')), dimension=6) + # x right here controls label place (interior / outer)

  coord_polar(theta="y") +
  xlim(c(-1, 4)) +
  theme_void() +
  theme( = "none")

# Do the identical for rt_r

Two ring graphs showing which user types retweet tweets from each cluster. Looking at Republican retweets, 76.3% are from other Republicans and 1.3% are from Democrats, while 22.4% are from nonclustered users. When looking at Democratic retweets, 75.3% are from other Democrats and 2.4% are from Republicans, while 22.3% are from nonclustered users.
Consumer Sort Retweet Distribution

As anticipated, Republicans are likely to retweet different Republicans and the identical is true for Democrats. Let’s see how occasion affiliation applies to tweet replies.

Two ring graphs showing which user types reply to tweets from each cluster. Looking at replies to Republican tweets, 36.5% are from Republicans and 16.2% are from Democrats, while 47.3% are from nonclustered users. When looking at replies to Democratic tweets, 28% are from Democrats and 20.6% are from Republicans, while 51.5% are from nonclustered users.
Consumer Sort Tweet Reply Distribution

A really totally different sample emerges right here. Whereas customers are likely to reply extra usually to the tweets of people that share their occasion affiliation, they’re nonetheless more likely to retweet them. Additionally, it seems that individuals who don’t fall throughout the two principal clusters are likely to favor to answer.

Through the use of the subject modeling approach specified by half two of this sequence, we are able to predict what sort of conversations customers will select to interact in with individuals of their similar cluster and with individuals of the other cluster.

The next desk particulars the 2 most necessary matters mentioned in every sort of interplay:

Democrats to Democrats Democrats to Republicans Republicans to Democrats Republicans to Republicans
Subject 1 Subject 2 Subject 1 Subject 2 Subject 1 Subject 2 Subject 1 Subject 2
pretend individuals trump individuals information biden individuals china
putin covid information trump pretend obama cash information
election virus pretend useless cnn obamagate nation individuals
cash taking lies individuals learn joe open media
trump useless fox deaths fake_news proof again pretend

It seems that pretend information was a sizzling subject when customers in our information set replied. No matter a consumer’s occasion affiliation, once they replied to individuals from the opposite occasion, they talked about information channels sometimes favored by individuals of their explicit occasion. Secondly, when Democrats replied to different Democrats, they tended to speak about Putin, pretend elections, and COVID, whereas Republicans targeted on stopping the lockdown and faux information from China.

Polarization Occurs

Polarization is a typical sample in social media, taking place everywhere in the world, not simply within the US. We now have seen how we are able to analyze neighborhood id and habits in a polarized situation. With these instruments, anybody can reproduce cluster evaluation on a knowledge set of their curiosity to see what patterns emerge. The patterns and outcomes from these analyses can each educate and assist generate additional exploration.

Additionally in This Sequence:

Additional Studying on the Toptal Engineering Weblog:


Leave a Reply

Your email address will not be published. Required fields are marked *