Social media has become an integral part of today's web and allows users to share content and socialize. Understanding the factors that influence how users evolve over time - for example how their social network and their contents co-evolve is an issue of both theoretical and practical relevance. This paper sets out to study the temporal co-evolution of content and social networks on Twitter and bi-directional influences between them by using multilevel time series regression models. Our findings suggest that on Twitter social networks have a strong influence on content networks over time, and that social network properties, such as users' number of followers, strongly influence how active and informative users are. While our investigations are limited to one small dataset obtained from Twitter, our analysis opens up a path towards more systematic studies of network coevolution on platforms such as Twitter or Facebook. Our results are relevant for researchers and social media hosts interested in understanding how content-related and social activities of social media users evolve over time and which factors impact their co-evolution.
ries modeling [
In regression analysis variables often stem from different
levels. So called multilevel regression models are an appropriate
way to model such data. Hence, the measurement occasion
is the basic unit which is nested under an individual, the
cluster unit. In our dataset we have such a hierarchical
nested structure. For each day different properties are
measured repeatedly, but all of these values belong to different
individuals in our study. If we would apply a simple
autoregressive model to our data we would ignore the difference
between each user and would just calculate the so-called
fixed effects, because we can not assume that all
clusterspecific influences are included as covariates in the analysis
[
We chose Twitter as a platform for studying the co-evolution of communication content and social networks, since it is a popular micro-blogging service. We explore one random dataset in this work, which was crawled within a time period of 30 days. This random dataset consists of random users from the public timeline who do not have anything special in common.
To generate the random dataset, we randomly chose 1500 users from the public Twitter timeline who we used as seed users. We used the public timeline method from the Twitter API to sample users rather than using random user IDs since the timeline method is biased towards active Twitter users. To ensure that our random sample of seed users consists of active, English-speaking Twitter users, we further only kept users who mainly tweet in English, have at least 80 followers, 40 followees and 200 tweets. We also had to remove users from our dataset who deleted or protected their account during the 30 days of crawling. Hence, we ended up having 1.188 seed users for whom we were able to crawl their social network (i.e., their followers and followees) and their tweets and retweets. To identify retweets we used the flag provided by the official Twitter API and to extract URLs we used a regular expression. During a 30 day time period (from 15.03.2011 to 14.04.2011) we polled the data daily at about the same time.
The goal of our experiments is to study the co-evolution of social and content networks of Twitter users and influence patterns between them. In order to achieve this we firstly created a social and content network for each specific time point.
Social network: The social network is a one-mode directed network, where each vertex represents a user and the edges between these vertices represent the directed follow-relations between two users at a certain point in time. The constructed social network of seed users only reflects a sub-part of a greater network. Therefore it makes no sense to calculate and analyze specific network properties such as betweenness centrality or clustering coefficient, because these properties depend on the whole network and we only have data available for a certain sub-network.
Content network: The content network at each point in
time is a two-mode network, which connects users and tweets
via authoring-relations. From these user-tweet networks one
can extract specific tweet features, such as hashtags, links or
retweet information, and build, for example, a user-hashtag
network. It would also be possible to create further types of
content networks, such as hashtag co-occurrence networks
(see [
Overall, the social networks capture the social following relations between users, whereas our content networks account the tweets users publish. Finally, we can connect both networks via their user vertexes, since we know which user in the social network corresponds to which user in the content network and vice versa.
A further step towards our final results is the
normalization of our available data. This is done by subtracting the
time-overall mean and dividing the result by the time-overall
standard deviation. The fixed effects can now be analyzed
as the effect of one standard deviation of change in the
independent variable on the number of standard deviations
change in the dependent variable [
Based on the prepared data, the final model described in section 2 can be applied to identify potential influences between social and content network properties over time. Ta#MSM2012 Social Network -0.23 ble 1 describes each social and content network property used throughout our experiment. The properties are calculated for a corresponding social or content network at each time point t of the random Twitter dataset. The dependent variable of the model is always a property at time t and the independent variable are all properties at time t − 1 including the dependent variable at that time. Including the dependent variable in that step allows us to detect if a variable’s previous value influences it’s future value. Finally, the resulting statistical significant coefficients show a relationship between an independent variable at time t − 1 and a dependent variable at time t. Positive coefficients indicate that a high value of a property leads to an increase of another property, while negative coefficients indicate that a high value of a property leads to an decrease of another property. To reveal positive and negative influence relations between properties within and across different networks, we visualize them as graphical influence network.
Our results reveal interesting influence patterns between social networks and content networks. The influence network in figure 1 shows the correlations detected in the multilevel regression analysis via arrows that point out influences between a property at time t and another property at time t + 1.
The influence network reveals significant influences of social properties on content network properties. The strongest positive effects can be observed between the number of followers of a user and the content network properties - i.e., users’ number of followers positively influences their link ratio, their retweeted ratio and their number of tweets. This indicates that users start providing more tweets and also Social Content Content Content Content more links in their tweets if their number of followers increases. Not surprisingly, users’ tweets are also more likely to get retweeted if their number of followers increases, because more users are potentially reading their tweets. Further, figure 1 shows that the number of followees of the social network has positive and negative influences on the content network in our random dataset. While the positive effects point to the link and hashtag ratio, the negative effects point to the number of tweets and the retweeted ratio. This suggests that users who start following other users also start using more hashtags and links. One possible explanation for this is that users get influenced by the links and #MSM2012 hashtags used by the users they follow and might therefore use them more often in their own tweets. The negative effect of the number of followees on the number of tweets and the retweeted ratio suggests that users who start following many other users start behaving more like passive readers rather than active content providers.
Another observation of our experiment is that all properties influence themselves positively, which indicates that users who are active one day, tend to be even more active the next day. This indicates for example, that users who attract new followers one day tend to attract more new followers the day after.
The main contributions of this paper are the following: (i) We applied multilevel time series regression models to one selected Twitter dataset consisting of social and content network data and (ii) we explored influence patterns between social and content networks on Twitter. In our experiments we studied how the properties of social and content networks co-evolve over time. We showed that the adopted approach allows answering interesting questions about how users’ behavior on Twitter evolves over time and the factors that impact this evolution. While our results are limited to the dataset used, our work illuminates a path towards studying complex dynamics of network evolution on systems such as Twitter. Our analyses may also facilitate social media hosts to promote certain features of the platform and steer users and their behavior. For example, one can see from our analysis that usage of content features, such as hashtags and links, is highly influenced by social network properties such as the number of followers of a user. Therefore, social media hosts could try to encourage users to use more content features by introducing new measures such as a friend recommender techniques which might impact the social network of users. However, further work is warranted to study these ideas.
Overall, our findings on one small Twitter dataset suggest
that there are manifold sources of influence between social
and content network properties. Our results indicate that
users’ behavior and the co-evolution of content and social
networks on Twitter is driven by social factors rather than
content factors. Previous research by Anagnostopoulos et
al. [
Our experiments suggest that the number of followers powerfully influences properties of the content network. One interpreation for that is that the number of followers is a very important motivation for Twitter users to add more content and use more content features like hashtags, URLs or retweets. However, the number of users a user is following can also have a negative influence on content network properties as one can see from figure 1. Our results suggests that an increase of a user’s followees (i.e., the number of users he/she follows) implies that the user starts tweeting less and that his/her tweets get less frequently retweeted. Further, our findings show that all properties influence themselves positively. This does not mean that the values of all properties always increase over time, but that they tend to increase depending on how much they increased the day before. For example, a Twitter user who started posting more links at day t, is likely to post even more links at day t + 1 or a user who gain new followers at day t is likely to gain even more new followers at day t + 1.
To summarize, our work highlights the existence of
interesting influence relationships between content and social
networks on Twitter, and shows that multilevel time series
regression analysis can be used to reveal such relationships and
to study how they evolve over time. Based on the techniques
developed by Wang and Groth [
This work is in part funded by the FWF Austrian Science Fund Grant I677 and the Know-Center Graz. Claudia Wagner is a recipient of a DOC-fForte fellowship of the Austrian Academy of Science.