Authors:
Stefanie Haustein,
Timothy Bowman and Rodrigo Costas
Background:
‘Altmetrics’ have been introduced as a way to capture scientific output and impact beyond papers and citations based on traces on various social media platforms (Priem, Taraborelli, Groth, & Neylon, 2010), of which Twitter is believed to have a particular potential to reflect societal impact of research. The analysis and application of various altmetrics such as tweets to scientific papers, however, still lack adequate interpretative frameworks mainly because the processes behind the metrics are not yet fully understood. Currently each tweet is counted equally on platforms such as Altmetric.com or ImpactStory and studies tend to ignore user type and tweet content, although tweets have been shown to range from serious discussions to humour and self-promotion to automated mentions (Haustein et al., 2015).
Objective:
Communities of attention around scientific publications on Twitter are identified based on engagement and exposure of users. Engagement is measured as the degree to which the tweet text differs from the title of the tweeted paper. Exposure refers to the potential audience of the tweet as measured by the number of the user’s followers.
Methods:
Publications from 2012 covered by Web of Science were matched to tweets (until June 2014, excluding retweets) recorded by Altmetric.com via DOI resulting in 660,149 tweets, 237,222 tweeted papers, and 125,083 Twitter users. Engagement was calculated based on the dissimilarity between the tweet text (excluding user names and URLs) and the title of the tweeted document. User data (including the number of followers representing exposure) was collected from Altmetric.com and the Twitter API.
Four user categories were defined, classifying users into four quadrants A, B, C and D according to engagement and exposure values above and below the median of the whole dataset (Figure 1). Statistics based on the tweeting behaviour of users were calculated for each of the categories. The connections between 708 users with more than 100 publications based on co-mentions of the same papers were visualized in a network graph in Figure 1.
Results:
Users in the four categories differ according to tweeting behavior (Figure 1). Users in A have the highest mean tweets per day (based on all tweets) and those in D tweet more about scientific papers (typical for bots identified by Haustein et al. (2015)), while users in A and B discuss publications with slightly higher relative citation rates.
Future Work:
Categorizing users by engagement and exposure allows for more nuanced and meaningful indicators differentiating between types of tweets. Users of each category will be further analyzed including an analysis of Twitter account descriptions. We suspect science communicators to be prominent in category A, scientists in B, journal and publisher accounts in C, and bots in D.