Twitter data are collected via rtweet.
Authenticating with the Twitter API
The first step is to authorise access to the Twitter API. There are two ways of generating an access token. The “Dev Auth” approach involves authenticating using Twitter developer app API keys, which can be created after the user has been granted a Twitter developer account:
myDevKeys <- list(appName = "My App", apiKey = "xxxxxxxxxxxx", apiSecret = "xxxxxxxxxxxx",
accessToken = "xxxxxxxxxxxx", accessTokenSecret = "xxxxxxxxxxxx")
twitterAuth <- Authenticate("twitter", appName = myDevKeys$appName, apiKey = myDevKeys$apiKey,
apiSecret = myDevKeys$apiSecret, accessToken = myDevKeys$accessToken,
accessTokenSecret = myDevKeys$accessTokenSecret)The “Web Auth” approach undertakes Twitter authentication via user authorisation of an app that the user has access to (either because the user created the app, or else it was created by a third-party and access granted to the user). This will open a web browser, prompting the user to log into Twitter and authorise the app using Twitter app’s consumer key and secret. The Web Auth approach therefore does not require a Twitter developer account, as long as the user has a Twitter account and access to a Twitter app.
twitterAuth <- Authenticate("twitter", appName = "An App",
apiKey = "xxxxxxxxxxxx", apiSecret = "xxxxxxxxxxxx")In both cases, Twitter authentication creates an access token, which can be re-used by saving it to disk:
The following loads into the current session a previously-created access token:
Collecting Data from Twitter
The syntax for collecting Twitter data follow Twitter’s rules and filtering documentation. It is possible to collect tweets including particular terms (e.g. hashtags), and boolean searches (see standard search operators) are also possible. The collection may be filtered by, for example, type of Twitter activity (e.g. to include retweets only), number of collected tweets, language of tweet. As an example, the following collects 1000 recent tweets containing the ‘#auspol’ hashtag (a prominent hashtag for Australian politics), with retweets not being collected.
twitterData <- twitterAuth %>%
Collect(searchTerm = "#auspol", numTweets = 1000,
includeRetweets = FALSE, retryOnRateLimit = TRUE, writeToFile = TRUE,
verbose = TRUE)The Collect() function takes the following arguments (when used for collecting Twitter data): credential (object generated from Authenticate() with class name “twitter”(above we pass this via the pipe), searchTerm (character string that specifies a Twitter search term), searchType (character string indicating how to filter returned tweets with options ‘recent’, ‘mixed’ or ‘popular’; default type is ‘recent’), numTweets (numeric vector that specifies how many tweets to be collected; default is 100), includeRetweets (whether the search should filter out retweets or not; default is TRUE), retryOnRateLimit (whether to automatically pause collection when the Twitter API rate limit is reached, and then restart; default is FALSE.), writeToFile (whether to write the returned dataframe to disk as an .rds file; default is FALSE), and verbose (whether to output information about the data collection; default is FALSE).
The Collect() function returns a tibble (an enhanced dataframe which has features that can make working with the data easier). We can view the data we just collected (the following has been modified to anonymise the data):
> twitterData
# A tibble: 999 x 90
user_id status_id created_at screen_name text source
<chr> <chr> <dttm> <chr> <chr> <chr>
1 xxxxxx… xxxxxxxx… 2020-01-09 12:02:13 xxxx "htt… Twitt…
2 xxxxxx… xxxxxxxx… 2020-01-09 12:01:32 xxxxxxxxx "Fir… Twitt…
3 xxxxxx… xxxxxxxx… 2020-01-09 12:00:44 xxxxxxxxxxx "Ser… Twitt…
[snip]
… with 989 more rows, and 84 more variables: display_text_width <dbl>,
...Note that if you read the Twitter dataframe from disk, you will need to modify the class values for the object before you can use it with vosonSML. This can be done manually or by using the handy ImportData() function:
Creating Twitter Networks
It is currently possible to create four types of networks using Twitter data: (1) actor network; (2) activity network; (3) 2-mode network and (4) semantic network.
Actor Network
In the Twitter actor network, nodes are users who have either tweeted using the target search terms (#auspol in the above example) or else are mentioned or replied to in tweets featuring the search terms. Edges represent interactions between Twitter users, and an edge attribute indicates whether the interaction is a mention, reply, retweet, quoted retweet or self-loop. Self-loop edges are created in two situations: (1) a user authors a tweet and mentions or replies to themself; (2) a user authors a tweet containing the search term, but does not mention or reply to any other user in that tweet.
actorNetwork <- twitterData %>% Create("actor", writeToFile = TRUE, verbose = TRUE)
actorGraph <- actorNetwork %>% Graph(writeToFile = TRUE)Create("actor") returns a named list containing two dataframes named “nodes” and “edges” (the following has been modified to preserve anonymity):
> actorNetwork
$edges
# A tibble: 1,725 x 5
from to edge_type timestamp status_id
<fct> <fct> <fct> <fct> <fct>
1 xxxxxxxx xxxxxxxx quote 2020-01-09 12:00… xxxxxxxxxxxxxxx…
2 xxxxxxxx xxxxxxxxx quote 2020-01-09 09:37… xxxxxxxxxxxxxxx…
[snip]
# … with 1,715 more rows
$nodes
# A tibble: 1,158 x 2
user_id screen_name
<fct> <fct>
1 xxxxxxxx xxxx
2 xxxxxxxx xxxxxxxxx
[snip]
# … with 1,148 more rows
attr(,"class")
[1] "list" "network" "actor" "twitter"This list is then passed to Graph(), which returns an igraph graph object, and in the above example, the writeToFile parameter is used to write the graph to file in GraphML format. The following shows a summary of the graph:
> actorGraph
IGRAPH bc177a6 DN-- 1158 1725 --
+ attr: type (g/c), name (v/c), screen_name (v/c), label (v/c),
| edge_type (e/c), timestamp (e/c), status_id (e/c)
+ edges from bc177a6 (vertex names):
[1] xxxxxxxx ->xxxxxxxx
[2] xxxxxxxx ->xxxxxxxxx
[snip]
+ ... omitted several edgesThe Twitter actor network contains a graph attribute type which is set to “twitter” (this attribute is required for VOSON Dashboard). The following node attributes are collected from the Twitter profile data: name (Twitter ID), screen_name (Twitter handle or screen name) and label (a concatenation of the ID and screen name). The edge attributes are: edge_type (whether the edge is a mention, reply, retweet, quoted retweet or self-loop), timestamp (when the tweet that led to the creation of the edge was authored) and status_id (the Twitter ID for the tweet).
The example actor network contains 1158 nodes and 1725 edges. The following code uses igraph functions to: (1) remove all edges other than reply edges; (2) construct a subnetwork consisting of the giant component (the largest set of connected nodes); (3) plot this network.
library(igraph)
g2 <- delete.edges(actorGraph, which(E(actorGraph)$edge_type!="reply"))
cc <- clusters(g2)
g2 <- induced_subgraph(g2,which(cc$membership == which.max(cc$csize)))
png("twitter_actor_reply_gc.png", width=600, height=600)
plot(g2, vertex.label="", vertex.size=4, edge.arrow.size=0.5)
dev.off()
Twitter actor network - replies only - giant component
It is often useful to have the tweet text content in the network. This can be achieved by using the status_id edge attribute to pull the tweet text content from the dataframe returned by Collect(), and store it as an edge attribute. vosonSML makes this easy with the AddText() function, with the following example creating a graph with an edge attribute vosonTxt_tweet which stores the tweet text content.
Now we have tweet text cotent stored as an edge attribute, we can use it for text analysis or creating new node attributes. As an example, the following creates a new node attribute tweetedBushfires which has the value “yes” if the Twitter user authored at least one tweet containing the word “bushfire” and “no” otherwise.
#vector of igraph vertices - the users who tweeted with "bushfire"
ind <- tail_of(actorGraphWithText,grep("bushfire",tolower(E(actorGraphWithText)$vosonTxt_tweet)))
V(actorGraphWithText)$tweetedBushfires <- "no"
V(actorGraphWithText)$tweetedBushfires[ind] <- "yes"The following creates and plots the giant component in the reply network, with red nodes indicating those users who tweeted using the term “bushfire”.
g3 <- delete.edges(actorGraphWithText, which(E(actorGraphWithText)$edge_type!="reply"))
cc <- clusters(g3)
g3 <- induced_subgraph(g3,which(cc$membership == which.max(cc$csize)))
V(g3)$color <- ifelse(V(g3)$tweetedBushfires=="yes","red","grey")
png("twitter_actor_reply_gc_bushfires.png", width=600, height=600)
plot(g3, vertex.label="", vertex.size=4, edge.arrow.size=0.5)
dev.off()
Twitter actor network - replies only - giant component (red nodes indicate who tweeted at least once about bushfires)
The igraph graph object can then be saved to disk as a GraphML file using the igraph function write.graph, and later imported into network analysis software such as igraph, VOSON Dashboard and Gephi:
Finally, the AddUserData() function can be used to create additional node attributes from the Twitter profile, for example, number of followers and number of tweets authored by the user.
Note that by default, vosonSML will only collect Twitter profile data for those users who authored tweets that have been collected. That is, in the above example, profile data will not be collected (by default) for a user who was mentioned in a tweet that contained #auspol, but did not author such a tweet. The lookupUsers argument can be used to make additional Twitter API calls and collect the profile data for mentioned users whose profile data is absent, so that their additional profile attributes can also be added:
Activity Network
In the Twitter activity network, nodes represent tweets and edge types are: replies, retweets and quoted retweets.
activityNetwork <- twitterData %>% Create("activity") %>% AddText(twitterData)
activityGraph <- activityNetwork %>% Graph(writeToFile = TRUE)Create("activity") returns a named list containing two dataframes named “nodes” and “edges” (the following has been modified to preserve anonymity):
> activityNetwork
$nodes
# A tibble: 1,408 x 5
status_id user_id screen_name created_at vosonTxt_tweet
<chr> <chr> <chr> <chr> <chr>
1 xxxxxxxxxxx… xxxxxxxx xxxx 2020-01-09 … "xxxxxxxxxxxxxxxxxxxxxxxxxx…
2 xxxxxxxxxxx… xxxxxxxx xxxxxxxxx 2020-01-09 … "xxxxxxxxxxxxxxxxxxxxxxxxxx…
[snip]
# … with 1,398 more rows
$edges
# A tibble: 662 x 3
from to edge_type
<chr> <chr> <chr>
1 xxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxx quote
2 xxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxx quote
[snip]
# … with 652 more rows
attr(,"class")
[1] "list" "network" "activity" "twitter" "voson_text"Note that in the above, AddText() was used to add the comment text data to the network dataframe, stored as a node attribute. This list is then passed to Graph(), which returns an igraph graph object (the following has been anonymised):
> activityGraph
IGRAPH e60c486 DN-- 1408 662 --
+ attr: type (g/c), name (v/c), user_id (v/c), screen_name (v/c),
| created_at (v/c), vosonTxt_tweet (v/c), label (v/c), edge_type (e/c)
+ edges from e60c486 (vertex names):
[1] xxxx->xxxx
[2] xxxx->xxxx
[snip]
+ ... omitted several edgesThe Twitter activity network contains a graph attribute type (set to “twitter”). The node attributes are: name (Twitter ID for the tweet), user_id (Twitter ID for the user who authored the tweet), screen_name (Twitter handle or screen name of the user who authored the tweet), created_at (timestamp when the tweet was authored), vosonTxt_tweet (text content of the tweet) and label (a concatenation of name and user_id). The edge attribute is edge_type which can have the value ‘reply’, ‘retweet’ or ‘quote’.
The example Twitter activity network contains 1408 nodes and 662 edges. The following is a visualisation of the network, where nodes are tweets and tweets mentioning bushfires are indicated in red.
cc <- clusters(activityGraph)
g2 <- induced_subgraph(activityGraph,which(cc$membership %in% which(cc$csize>5)))
ind <- grep("bushfire",tolower(V(g2)$vosonTxt_tweet))
V(g2)$color <- "grey"
V(g2)$color[ind] <- "red"
png("twitter_activity.png", width=600, height=600)
plot(g2, vertex.label="", vertex.size=4, edge.arrow.size=0.5)
dev.off()
Twitter activity network (red indicates tweets mentioning bushfires)
It should be noted that a limitation of the Twitter API is that retweet chains are not provided. That is: if user i tweeted an original tweet, and then user j retweeted this tweet, and user k retweeted j’s retweet, the activity network will show edges connecting the two retweets to the original tweet.
2-mode Network
In the Twitter 2-mode network, the two types of nodes are actors (Twitter users) and hashtags. There is an edge from user i to hashtag j if user i authored a tweet containing hashtag j.
# additional required packages
install.packages("tidytext")
twomodeNetwork <- twitterData %>% Create("twomode", removeTermsOrHashtags = c("#auspol"))
twomodeGraph <- twomodeNetwork %>% Graph(writeToFile = TRUE)Create("twomode") returns a named list containing two dataframes named “nodes” and “edges” (the following has been modified to preserve anonymity). Note that in this example, the removeTermsOrHashtags argument was used to exclude ‘#auspol’, since by construction all tweets contained this hashtag.
> twomodeNetwork
$nodes
# A tibble: 1,146 x 2
entity_id display_name
<chr> <chr>
1 xxxxxxxx xxxx
2 xxxxxxxx xxxxxxxxx
3 #auspol2020 #auspol2020
4 #australianbushfiredisaster #australianbushfiredisaster
[snip]
# … with 1,136 more rows
$edges
# A tibble: 1,675 x 5
from to edge_type timestamp status_id
<fct> <fct> <fct> <fct> <fct>
1 xxxxxxxx #auspol2020 hashtag 2020-01-09 12:0… xxxxxxxxxxxxxxx…
2 xxxxxxxx #australianbushfiredis… hashtag 2020-01-09 12:0… xxxxxxxxxxxxxxx…
[snip]
# … with 1,665 more rows
attr(,"class")
[1] "list" "network" "twomode" "twitter"This list is then passed to Graph(), which returns an igraph graph object (this has been anonymised):
> twomodeGraph
IGRAPH 68bd240 DN-- 1146 1675 --
+ attr: type (g/c), name (v/c), display_name (v/c), label (v/c),
| edge_type (e/c), timestamp (e/c), status_id (e/c)
+ edges from 68bd240 (vertex names):
[1] xxxx -> #auspol2020
[2] xxxx -> #australianbushfiredisaster
[snip]
+ ... omitted several edgesThe Twitter 2-model network has a graph attribute type (set to “twitter”). The node attributes are: name (hashtag or Twitter user ID), display_name (hashtag or Twitter handle or screen name), label (for users, a concatenation of name and display_name, while for hashtags it is name). The edge attributes are: edge_type (‘hashtag’), timestamp (timestamp of the tweet that led to the edge), status_id (Twitter ID of the tweet that led to the edge).
ind <- order(degree(twomodeGraph, mode="in"), decreasing=TRUE)[1:5] # top-5 hashtags
# users who tweet with these hashtags
ind2 <- unlist(lapply(ind, function(x){neighbors(twomodeGraph, x, mode="in")}))
g2 <- induced_subgraph(twomodeGraph, c(ind, as.numeric(ind2)))
V(g2)$color <- "grey"
V(g2)$color[which(degree(g2, mode="in")>0)] <- "blue"
V(g2)$label2 <- ifelse(degree(g2, mode="in")>0, V(g2)$label, "")
png("twitter_twomode.png", width=600, height=600)
plot(g2, vertex.label=V(g2)$label2, vertex.size=4, edge.arrow.size=0.5, vertex.label.cex=1.8,
vertex.label.color="red")
dev.off()
Twitter 2-mode network - top-5 hashtags, and the users who tweeted them
Semantic Network
In the Twitter semantic network, nodes represent entities extracted from the tweet text: common words, hashtags and usernames. Edges reflect co-occurrence i.e. there is an edge between entities i and j if they both occurred in the same tweet.
# additional required packages
install.packages(c("tidyr", "tidytext", "stopwords"))
semanticNetwork <- twitterData %>% Create("semantic",
removeTermsOrHashtags = c("#auspol", "auspol", "australia"),
termFreq = 5)
semanticGraph <- semanticNetwork %>% Graph(writeToFile = TRUE, directed = FALSE)Create("semantic") returns a named list containing two dataframes named “nodes” and “edges”:
> semanticNetwork
$nodes
# A tibble: 799 x 1
value
<fct>
1 just
2 one
3 fire
4 going
5 still
6 hard
7 trying
8 since
9 try
10 sick
# … with 789 more rows
$edges
# A tibble: 10,990 x 3
from to weight
<fct> <fct> <int>
1 #auspol2020 #australianbushfiredisaster 2
2 #auspol2020 government 2
3 #auspol2020 fire 4
4 #auspol2020 australian 2
5 #auspol2020 bushfire 2
6 #auspol2020 fires 4
7 #auspol2020 #australiafires 1
8 #auspol2020 #australianbushfiresdisaster 1
9 #auspol2020 #australia 4
10 #auspol2020 bushfires 2
# … with 10,980 more rows
attr(,"class")
[1] "list" "network" "semantic" "twitter" The removeTermsOrHashtags argument is used to exclude terms that we can expect to occur frequently (e.g. because of the parameters used for the data collection). The termFreq parameter is used to limit the network to the top 5 percent most frequently occurring entities.
This list is then passed to Graph(), which returns an igraph graph object:
> semanticGraph
IGRAPH cb8c381 UNW- 799 10990 --
+ attr: type (g/c), name (v/c), label (v/c), weight (e/n)
+ edges from cb8c381 (vertex names):
[1] #australianbushfiredisaster --#auspol2020
[2] government --#auspol2020
[3] fire --#auspol2020
[4] australian --#auspol2020
[5] bushfire --#auspol2020
[6] fires --#auspol2020
[7] #australiafires --#auspol2020
[8] #australianbushfiresdisaster--#auspol2020
+ ... omitted several edgesThe Twitter semantic network node contains a graph attribute type (set to “twitter”). The node attributes are: name (the entity i.e. hashtag or word), label (same as name). The edge attribute is weight (how many tweets the two entities co-occurred in).
The example Twitter semantic network has 799 nodes and 10990 edges. The following produces a visualisation of subnetwork of terms that contain the string “bushfire”:
ind <- grep("bushfire",tolower(V(semanticGraph)$name))
g2 <- induced_subgraph(semanticGraph,ind)
png("twitter_semantic.png", width=600, height=600)
plot(g2, layout=layout_with_lgl(g2), vertex.shape="none", vertex.size=4, edge.width=1+log(E(g2)$weight))
dev.off()
Twitter semantic network - entities containing the word “bushfire” (edge thickness proportional to edge weight)
