Source-linked AI summary
The Bursty Dynamics of the Twitter Information Network
Seth A. Myers, Jure Leskovec
TL;DR
The paper asks how information sharing interacts with the evolution of social-network connections. It studies complete Twitter network dynamics and develops a diffusion-based burst model, finding that cascades can reorganize local follower networks and that the model can predict diffusion events leading to bursts.
Problem
The paper addresses the limited understanding of how information posting and sharing interact with the creation and deletion of social-network connections.
Method
The authors analyze the complete dynamics of a 13.1-million-user Twitter subgraph and model bursts using the similarity of users exposed through information diffusion.
Results
Information cascades often create sudden connection bursts that make follower networks more cohesive and interest-homogeneous, while real-world events can also change edge creations and deletions.
Takeaways & Limitations
The model can predict which information-diffusion events will trigger spikes in network dynamics and indicate their effects on local network properties.
Takeaways & Limitations
The authors identify further work on unfollow spikes, user compatibility, topic effects, and incorporating tweet content into the model.
Abstract
from arXiv · showhide
In online social media systems users are not only posting, consuming, and resharing content, but also creating new and destroying existing connections in the underlying social network. While each of these two types of dynamics has individually been studied in the past, much less is known about the connection between the two. How does user information posting and seeking behavior interact with the evolution of the underlying social network structure? Here, we study ways in which network structure reacts to users posting and sharing content. We examine the complete dynamics of the Twitter information network, where users post and reshare information while they also create and destroy connections. We find that the dynamics of network structure can be characterized by steady rates of change, interrupted by sudden bursts. Information diffusion in the form of cascades of post re-sharing often creates such sudden bursts of new connections, which significantly change users' local network structure. These bursts transform users' networks of followers to become structurally more cohesive as well as more homogenous in terms of follower interests. We also explore the effect of the information content on the dynamics of the network and find evidence that the appearance of new topics and real-world events can lead to significant changes in edge creations and deletions. Lastly, we develop a model that quantifies the dynamics of the network and the occurrence of these bursts as a function of the information spreading through the network. The model can successfully predict which information diffusion events will lead to bursts in network dynamics.
1. INTRODUCTION
The paper examines how information sharing interacts with Twitter follower-network evolution, using a large-scale dynamic subgraph to study, explain, and predict bursts of edge changes. It finds that diffusion cascades can reorganize local networks toward greater cohesion and follower-interest similarity, while real-world events also affect connections.
- Present work: Information causes bursts in network evolution: The study analyzes 13.1 million English-speaking users, covering 1.2 billion tweets, 112.3 million new connections, and 39.2 million deleted connections.The analysis follows the complete dynamics of a fixed Twitter-user subgraph.
- Bursts of edge creations and deletions: About 9% of Twitter connections change monthly amid steady edge flux, but large information cascades can trigger abrupt bursts in network dynamics.An average user with 100 followers gains 10% more followers and loses about 3% of existing followers per month.
- Information causes bursts in network evolution: Information cascades produce coordinated unfollow bursts and follow bursts, increasing follower similarity and the density of connections among followers.Follow bursts expose similar users to one another, while unfollow bursts remove less-similar followers; both processes increase local-network coherence.
- Information causes bursts in network evolution: The paper finds evidence that external events, exemplified by Occupy Wall Street, connect users with similar interests as event-related news diffuses.Users interested in the event appear to connect to one another to learn more about it.
- Modeling and predicting bursts: A model predicts whether information diffusion will produce a follower burst by comparing exposed potential followers’ similarity with that of users regularly exposed to the target’s posts.The model quantifies burst occurrence as a function of information diffusion and can predict effects on local network properties.
2. ANALYSIS OF BURSTS
Twitter exhibits steady background churn punctuated by information-related bursts that alter follower connections and local network composition. Retweet exposure is associated with new follows, while burst-driven gains and losses make followers more similar to the focal user.
- 2.2 Twitter graph is highly dynamic: 9% of Twitter edges change monthly, reflecting persistent background churn in follows and unfollows.The measured rates are 7% new edges and 2.3% removed edges in the observed subpopulation.
- 2.3 Information diffusion and follows/unfollows: 21% of new follows come from users who recently saw a retweet by the user they began following.Retweet exposure can reveal the originator to followers of an intermediate user.
- 2.3 Information diffusion and follows/unfollows: Retweets and new followers are related even after conditioning on indegree, whereas about 10 monthly tweets minimizes follower loss for users with degree 1000-2000.Excessive and insufficient tweeting are both associated with more unfollows.
- 2.4 Detecting bursts in the flow of followers: Retweet bursts can be followed by large follower bursts, but some retweet bursts produce negligible follower increases and background churn persists without user activity.The temporal patterns therefore include both coupled and uncoupled activity changes.
- 2.4 Detecting bursts in the flow of followers: Burst detection removes daily periodicity using locally weighted regression based on observations separated by 24 hours and exponentially decaying weights.The method addresses noise and regular daily activity cycles in arrival rates.
- 2.5 User’s ego-network during a burst: Follower tweet similarity increases 25.5% faster during retweet-follow bursts, because newly acquired followers are 76.6% more similar than unexposed new followers and 109.5% more similar than existing followers.Tweet-unfollow bursts accelerate similarity growth by 11%, partly because less-similar followers leave.
3. MODELING FOLLOWER BURSTS
The model predicts follow bursts by identifying compatible 2-hop users who encounter a source through retweet exposure for the first time. It outperforms baselines, with retweet exposure and follower-interest similarity providing the strongest signals.
- Model goal: The model predicts which retweet bursts will cause spikes of new followers without considering tweet content.It is designed to identify bursts that create new follower connections.
- Burst mechanism: New follows are predominantly triadic: the average directed shortest path before a new edge forms is 2.036 ± 0.007.The model therefore focuses on transitions from the 2-hop neighborhood to the 1-hop neighborhood.
- Burst mechanism: A retweet-follow burst occurs when compatible 2-hop users discover a source through a retweet cascade and subsequently follow it.These users are connected through current followers who rarely retweet the source, so they have not previously been exposed.
- Similarity model: Follower tweet similarity increases the probability of a new follow exponentially, and more-similar potential followers are almost an order of magnitude more likely to follow.Yij normalizes a candidate’s similarity relative to the source’s current follower-similarity distribution.
- Prediction results: Burst likelihood depends more on newly exposed compatible followers than on a user’s current visibility or raw retweet count.Users with many already-exposed compatible followers are less primed for bursts, whereas exposure through users with large followings is especially informative.
- Prediction results: 0.519 AUC versus 0.382 for the best baseline: the model substantially outperforms alternatives in predicting follow bursts.Retweet exposure is the strongest baseline, while pre-burst follower count performs only marginally better than random.
4. RELATED WORK
Related work has studied network link creation, deletion, and information diffusion largely as separate processes. More recent studies examine their interaction, including information diffusion’s contribution to new edges and temporal correlations between network and information dynamics.
- Network dynamics: Prior work models network evolution and predicts both the addition and deletion of specific user connections.These studies identify features useful for forecasting local edge changes.
- Information diffusion: Related studies connect information diffusion with network dynamics through null models and autoregressive analyses.One study estimates that information diffusion motivates about 12% of newly formed edges.
- Research gap: The paper distinguishes its focus by examining how information diffusion interacts with sudden interruptions of otherwise steady follow and unfollow arrivals.This framing links temporal information flow to bursts in network evolution.
5. CONCLUSION
The paper links bursts in network evolution to large information-diffusion events and develops a model to predict when diffusion triggers graph-dynamics spikes. It identifies content analysis, unfollow spikes, and user compatibility as directions for further study.
- Large information-diffusion events cause bursts of edge creations or deletions in network evolution.These bursts appear as sudden influxes of network changes.
- The developed model predicts with high accuracy whether a diffusion event will trigger a spike in graph dynamics.It also provides insight into what causes spikes.
- Future work should examine unfollow spikes and the user-compatibility factors associated with bursts of new followers.
- Further research could analyze how content and different topics affect network dynamics.
- Extending the model to incorporate each tweet's content remains an open direction.