Analyzing Large-Scale Public Campaigns on Twitter

Abstract

Social media has become an important instrument for running various types of public campaigns and mobilizing people. Yet, the dynamics of public campaigns on social networking platforms still remain largely unexplored. In this paper, we present an in-depth analysis of over one hundred large-scale campaigns on social media platforms covering more than 6 years. In particular, we focus on campaigns related to climate change on Twitter, which promote online activism to encourage, educate, and motivate people to react to the various issues raised by climate change. We propose a generic framework to identify both the type of a given campaign as well as the various actions undertaken throughout its lifespan: official meetings, physical actions, calls for action, publications on climate related research, etc. We study whether the type of a campaign is correlated to the actions undertaken and how these actions influence the flow of the campaign. Leveraging more than one hundred different campaigns, we build a model capable of accurately predicting the presence of individual actions in tweets. Finally, we explore the influence of active users on the overall campaign flow.

Appendices

A Action Detection

Examples of tweet actions are shown in Table 1.

Table 1. Sample tweets for each type of action considered.

Results of the action detection based on Decision Tree Ensembles is shown in Table 2. It shows the precision and recall results for the four types of actions using 10-fold cross-validation.

Table 2. Precision, Recall and F1-score values for classification of different types of actions with different sets of features.

Below we show examples of manually created rules to preselect the tweets given specified action types.

Table 3. Examples of rules and number of tweets for each type of action.

B Unique Tweets Identification and Retweets Count

One of the main issues with the data collected from Topsy is that the tool does not provide information about retweets. Therefore, we had to create heuristics to make sure that we could properly identify all retweeted messages. Taking into account that all tweets returned by the tool are sorted by timestamp, we can easily figure out the origin of all the tweets using a simple regex pattern ((RT|MT) @author tweet_prefix). This approach has a number of limitations, however. It does not identify complex retweet structures, such as where a tweet text is cited using quotes. We found that such cases are quite rare on Twitter and amount for $\sim$0.5 % of all tweets.

In order to compute the complex retweet cases, we aggregated the tweets with at most 5 characters edit distance. Further, we discarded explicit retweets ((RT|MT) @author) and exact duplicates. However, certain retweets can be missing when a hashtag does not fit into the message due to the tweet length limit. To solve this problem, we leveraged the Topsy API, by returning and analyzing related tweets for each requested tweet in order to identify all further retweets. Finally, we note that we apply this process recursively—searching for retweets of retweets iteratively—in order to capture potentially complex retweet patterns. When no new retweets can be identified, we identify content that was not retweeted but duplicated. The practice of duplicating tweets gained traction on the platform as it can help promote topics into Twitter Trends. We consider a tweet to be a duplicate whenever at least 80 % of its contents exactly matches an original tweet excluding punctuation.