Competition among memes in a world with limited attention

Lada Weng; Alessandro Flammini; Alessandro Vespignani; Filippo Menczer

doi:10.1038/srep00335

In brief

This paper models how information spreads through social networks when users have limited attention. The key finding: the combination of social network structure and finite user attention alone can explain massive heterogeneity in meme popularity and persistence—without assuming differences in meme quality or user influence. The authors validate the model against 120 million Twitter retweets from 2010–2011, showing that a few memes go viral while most fade, driven by competition for a fixed attentional budget per user.

Summary figure for Competition among memes in a world with limited attention

Contributions

First explicit agent-based model of meme competition for limited attention on social networks, showing it accounts for observed long-tailed distributions without exogenous factors
Empirical validation that user attention breadth remains constant despite variation in system-wide information diversity (Shannon entropy analysis of 120M retweets)
Demonstration that user interests (memory of past tweets) predict future retweet behavior (ρ = 0.98)
Ablation studies showing both social network heterogeneity and competition for attention are necessary to reproduce empirical distributions; random networks or weak competition fail

Method

The authors build a discrete-time agent-based model on a directed social network (follower graph). Each agent maintains a bounded screen (displaying received posts) and memory (recording posted memes), both with finite capacity—posts disappear after one time unit. At each step, a randomly selected agent either posts a new meme with probability $p_n$ or scans the screen. Each on-screen meme attracts attention with probability $p_r$; the agent then retweets it (probability $1 - p_m$) or triggers a meme from memory (probability $p_m$). All parameters are estimated directly from Twitter data: $p_n = 0.45$, $p_r$ tuned per simulation, $p_m = 0.4$.

The model is simulated on a sampled Twitter follower network (10^5 nodes, ~3×10^6 edges) and also on random Erdős–Rényi networks. Meme lifetime, popularity, user activity, and user attention breadth are measured and compared to empirical distributions from the full Twitter dataset.

Results

Model fit: Simulations on the real social network closely match empirical distributions of meme lifetime and popularity, user activity, and breadth of attention (entropy). Random networks produce far narrower distributions.
Role of network structure: The heterogeneous degree distribution of real social networks—hub users with many followers—explains why some memes achieve high popularity. Random networks, with Poisson degree distributions, fail to reproduce observed extremes.
Role of limited attention: Varying the time window $t_w$ in which posts are retained shows strong competition ($t_w < 1$) suppresses long-lived memes; weak competition ($t_w > 1$) fails to generate extremely popular memes. Standard competition ($t_w = 1$) reproduces all regimes.
Role of user memory: Removing memory (setting $p_m = 0$) eliminates the concentration of users with narrow topical focus observed in data, suggesting endogenous interests are critical.

Notes

Strengths:

Elegant parsimony: no exogenous factors (intrinsic meme quality, external media, user expertise differences) yet explains the data
Large-scale empirical validation (120M retweets, 12.5M users) with transparent parameter estimation
Rigorous ablations isolate the effects of network structure and attention limits separately
Clear operationalization of memes (hashtags) and attention (screen/memory capacity) in Twitter context

Weaknesses:

Model assumes frozen network and discrete time; does not capture network growth or continuous activity patterns observed in real platforms
Hashtag-based proxy for memes may conflate distinct topics under single tags and misses cross-platform propagation
Does not model explicit user influence (some spreaders may be more trusted); treats all users symmetrically
Limited to 2010–2011 Twitter data; platform dynamics and user behavior have shifted significantly since
Does not distinguish true information from false; implications for misinformation vs. authentic information diffusion remain implicit

Future directions:

Extend to dynamic networks and continuous-time models
Incorporate user credibility or topic-specific salience
Test on contemporary platforms (TikTok, Instagram) with different attention dynamics
Apply to false news / rumor propagation explicitly, measuring whether false claims face different attention curves than true claims