Source-linked AI summary
Trends in Social Media : Persistence and Decay
Sitaram Asur, Bernardo A. Huberman, Gabor Szabo, Chunyan Wang
TL;DR
The paper asks what causes some social-media topics to become popular trends and persist despite competition for attention. It develops a theoretical account and empirically studies Twitter trends, finding that content resonance and retweeting matter more than posting frequency or follower counts. Long-trending content largely originates in traditional media and is amplified through repeated retweets.
Problem
The paper addresses the unclear causes of trend formation and persistence in a social-media environment where content competes for limited user attention.
Method
The paper derives a stochastic model of trend growth and analyzes Twitter trending topics, their tweet distributions, persistence, and user impacts.
Results
Content resonance and retweets are more important for trends than users’ posting rates or follower counts, while long-trending content largely comes from traditional media.
Takeaways & Limitations
Twitter acts as a selective amplifier in which resonating traditional-media content is propagated through chains of retweets to generate trends.
Takeaways & Limitations
The persistence analysis assumes trends stop when relative tweet growth falls below a threshold and treats long-trending topics after novelty decay as governed mainly by random variation.
Abstract
from arXiv · showhide
Social media generates a prodigious wealth of real-time content at an incessant rate. From all the content that people create and share, only a few topics manage to attract enough attention to rise to the top and become temporal trends which are displayed to users. The question of what factors cause the formation and persistence of trends is an important one that has not been answered yet. In this paper, we conduct an intensive study of trending topics on Twitter and provide a theoretical basis for the formation, persistence and decay of trends. We also demonstrate empirically how factors such as user activity and number of followers do not contribute strongly to trend creation and its propagation. In fact, we find that the resonance of the content with the users of the social network plays a major role in causing trends.
1 Introduction
The paper frames trending topics as attention winners in a crowded social-media environment, while noting that their formation and persistence remain unclear. It studies Twitter trends to identify the factors behind their creation and evolution.
- Social media produces vast quantities of continuously shared content competing for users’ limited attention.
- Some topics attract disproportionate attention, becoming visible trends that can shape collective awareness and public agendas.
- Novelty, competition from newer content, and content resonance with followers are identified as factors associated with trend decay or propagation.
- The study analyzes Twitter trending topics, treating them as community-driven indicators of crowdsourced popularity.
- Long trends are linked more strongly to resonating content and retweet chains than to posting frequency or follower counts, with traditional media as a major source.
2 Related work
Prior Twitter research examined social connections, advertising, URL propagation, temporal hashtag patterns, and influence. These studies provide context for analyzing how trends spread and persist.
- Earlier work studied Twitter’s social interactions, word-of-mouth advertising, and prediction of which users would share particular URLs.
- Other research identified recurring temporal patterns in hashtag activity and found that influential bloggers were not necessarily the most active.
3 Twitter
Twitter is a large directed microblogging network where users publish short updates to followers. Retweets allow posts to propagate through the community, supporting Twitter’s role in news dissemination and commercial communication.
- Twitter had close to 200 million users and operated as a directed network in which users selected whom to follow.
- Tweets were short status updates containing personal information, news, or links, displayed on authors’ profiles and to their followers.
- Retweets forwarded posts from one user to another and propagated interesting content and links through Twitter.
- News organizations used Twitter to disseminate updates that the community filtered and commented on, while businesses used it for advertising and information sharing.
4 Twitter Trends Data
The dataset was built by repeatedly collecting Twitter’s algorithmically selected trending topics and then retrieving matching tweets in 20-minute windows. Reappearing topics were split into separate trend sequences.
- Twitter displayed two- to three-word trending expressions selected by a proprietary algorithm and updated every few minutes.
- The researchers queried trending topics every 20 minutes, then collected matching tweets, authors, text, and timestamps through the Search API.
- 20-minute intervals balanced discovering new trends with capturing all tweets under Twitter’s 1500-tweet search-query limit.
- Topics that became trends again after stopping were treated as separate sequences because trend spreading was expected to have a time-dependent rise and decline.
- The procedure divided 3468 originally collected trend titles into 6084 individual trend sequences.
5 Distribution of tweets
Trending-topic tweet counts exhibit log-normal behavior consistent with multiplicative growth under novelty decay. The decay factor follows an inverse-time power law, while most topics lose prominence quickly.
- Distribution of tweets: Cumulative tweet-count ratios across time frames are approximately log-normally distributed, especially for later frames.The logarithmically rescaled distributions are close to normal, except for 10–15 unusually frequent high-end outliers.
- Distribution of tweets: The multiplicative model treats each period’s tweet count as the previous count multiplied by noisy growth and a decreasing novelty factor.Past tweets proxy the users aware of a topic, while γ(t) accounts for novelty decay.
- Distribution of tweets: The model’s logarithm becomes a sum of random variables, producing an approximately normal distribution through the central limit theorem.The result remains normal when discounting factors do not decline too rapidly.
- Distribution of tweets: γ(t) is estimated by averaging fractions between consecutive tweet counts, using the unit expected value of the noise term.The decay factor is measured from the observed tweet-count sequence.
- Distribution of tweets: γ(t) follows a power-law decay with exponent −1, giving γ(t) ∼ 1/t.The corresponding regression fit has R2 = 0.98.
6 The growth of tweets over time
Because γ(t) ∼ 1/t, the multiplicative model predicts approximately linear cumulative tweet growth. Empirical topic trajectories are mostly linear after a brief initial sublinear phase, especially for longer-lasting trends.
- The growth of tweets over time: γ(t) ∼ 1/t yields Nq(t) ≈ Nq(0)t, so cumulative tweet counts increase linearly over the relevant time range.The paper states that no other decay form produces linear growth under the model’s approximations.
- The growth of tweets over time: Randomly selected topics show approximate initial linear growth in cumulative tweets during their first 48 hours.The trajectories are normalized to 1 for comparison across topics with different popularity levels.
- The growth of tweets over time: Across all topics, growth is slightly sublinear immediately after trending begins and then becomes mostly linear.The second-derivative data are distributed around zero after the initial regime.
- The growth of tweets over time: Topics trending for more than 4 hours exhibit an even stronger linear trend than the full topic set.Short-lived topics may show concave curvature because they lose popularity quickly and are removed from the trends.
- The growth of tweets over time: The authors suggest that visibility on Twitter’s trends site creates a constant probability that visitors will tweet about a highlighted topic.This proposed mechanism contrasts with informal social-network diffusion, which is expected to accelerate and then decay.
7 Persistence of Trends
Trending topics often persist through multiple bursts, but individual sequences are usually short. The paper links persistence to changing authors and models long-trend lifetimes with geometric or exponential decay.
- Topic recurrence: 34% of topics appear in more than one sequence, meaning they stop trending before beginning to trend again.The paper suggests time-zone differences and changing user activity as possible reasons for recurrence.
- Sequence lengths: Most topic sequences are short, while a few last much longer, following a power-law distribution.Competition for users’ limited attention may shorten most appearances, while recurrent topics can accumulate additional trending time.
- Relation to authors and activity: The correlation between unique authors and trend duration is 0.80, indicating that topics with more authors tend to remain trends longer.The authors interpret this pattern as evidence that propagation through the network contributes to trend persistence.
- Relation to authors and activity: Active-ratio quickly saturates and varies little with time, while the correlation between tweet count and authors is 0.83.This suggests that authors change as a topic propagates rather than a fixed group continually producing the trend.
- Persistence of long trending topics: A trend lifetime follows a geometric distribution under independent relative-growth changes, becoming exponential in the continuum case.The model assumes trending stops when relative tweet growth falls below a threshold θ after novelty decay has subsided.
- Persistence of long trending topics: The exponential fit to trending-duration density achieves R2 = 0.9112 for topics lasting more than 10 timestamps.The comparison uses only long-trending topics and is presented alongside the geometric-distribution prediction with p=0.12.
8 Trend-setters
Trend initiation is concentrated among a small set of prolific authors, but activity and audience size are weak predictors of how many trends they create. Persistence instead depends more on broad, sustained propagation than on domination by one author.
- Sources: A few authors contribute to many trending topics, following a power-law distribution among each topic’s first 100 authors.The analysis focuses on authors who contributed to at least five trending topics.
- Sources: 26.38 tweets per day was the mean tweet-rate, but its correlation with contributed trending topics was only 0.22.The weak relationship indicates that high activity alone does not strongly determine whether topics become trends.
- Propagators: 31% of trending-topic tweets were retweets, and retweet count correlated 0.96 with trend duration.Topics remain interesting as long as users continue retweeting them.
- Propagators: A negative −0.19 correlation between domination ratio and trend duration shows that topics centered on one author typically do not last long.Long-lasting trends require many people to contribute actively.
- Propagators: Influential authors retweeted across at least 50 trending topics included major news sources such as CNN, The New York Times, and ESPN.The analysis uses each author’s retweet-to-topic ratio to identify repeatedly influential contributors.
9 Conclusions
The paper models and measures how Twitter trends grow, persist, and decay. It finds that user attributes are weak drivers, while retweets and content—often traditional news amplified on Twitter—are more closely associated with trends.
- Conclusions: A stochastic growth model produces a log-normal distribution of trending-topic activity, matching the paper’s empirical results.Most topics trend briefly, while long-trending topics have geometrically distributed persistence.
- Conclusions: Follower counts and tweet-rates are not the attributes that cause trends; retweets by other users are more important.The paper relates retweeting more to the content shared than to the originating users’ attributes.
- Conclusions: Content from traditional media sources was often amplified through repeated Twitter retweets to generate trends.This finding connects the source of content with its network propagation.