Source-linked AI summary
Dynamical Classes of Collective Attention in Twitter
Janette Lehmann, Bruno Gonçalves, José J. Ramasco, Ciro Cattuto
TL;DR
The paper asks how spikes of collective attention in Twitter hashtags vary over time and relate to content and propagation. It analyzes hashtag activity profiles, semantic content, and social-network diffusion, finding discrete dynamical classes and a minor role for epidemic spreading, with popularity mostly driven by exogenous factors.
Problem
The paper investigates how temporal hashtag popularity peaks reflect collective attention and how exogenous and endogenous processes contribute to them.
Method
The study detects hashtag activity peaks, groups their temporal profiles into dynamical classes, grounds associated tweets in WordNet concepts, and tracks hashtag propagation through Twitter’s social network.
Results
The evolution of hashtag popularity defines discrete classes linked to event semantics, while epidemic spreading plays a minor role and popularity is mostly driven by exogenous factors.
Takeaways & Limitations
Temporal dynamics, content type, and external information injection together characterize collective-attention peaks more than epidemic propagation alone.
Takeaways & Limitations
The proposed parameters lack predictive power because they require a record of past activity, and epidemic-parameter estimates depend on network sampling.
Abstract
from arXiv · showhide
Micro-blogging systems such as Twitter expose digital traces of social discourse with an unprecedented degree of resolution of individual behaviors. They offer an opportunity to investigate how a large-scale social system responds to exogenous or endogenous stimuli, and to disentangle the temporal, spatial and topical aspects of users' activity. Here we focus on spikes of collective attention in Twitter, and specifically on peaks in the popularity of hashtags. Users employ hashtags as a form of social annotation, to define a shared context for a specific event, topic, or meme. We analyze a large-scale record of Twitter activity and find that the evolution of hastag popularity over time defines discrete classes of hashtags. We link these dynamical classes to the events the hashtags represent and use text mining techniques to provide a semantic characterization of the hastag classes. Moreover, we track the propagation of hashtags in the Twitter social network and find that epidemic spreading plays a minor role in hastag popularity, which is mostly driven by exogenous factors.
1. INTRODUCTION
Online popularity peaks create scientific challenges about the mechanisms governing collective attention, while semantic analysis can connect activity profiles to content. This paper uses Twitter hashtags to study those temporal and topical relationships.
- Motivation: Popularity peaks occur across online pages, videos, trending topics, and news stories, often leaving characteristic activity profiles.These attention shifts matter both for online-content monetization and for understanding social-system dynamics.
- Motivation: Scientific analysis must explain the mechanisms governing sudden concentration and subsequent movement of public attention.
- Motivation: Semantic analysis and natural-language processing can relate popular items’ features to their observed activity profiles.
- Research setting: Twitter provides publicly accessible message data, short texts suitable for automated processing, and real-time discussion of topics of popular interest.The platform combines characteristics of communication media and online social networks.
- Research approach: The paper identifies hashtag dynamical classes, relates them to tweet semantics and social-network spreading, and summarizes implications for information propagation.Its sections cover related work, data and methods, dynamical classes, semantic relations, network spreading, and applications.
2. RELATED WORK
Prior work examined Twitter’s network, messaging, propagation, credibility, and temporal popularity patterns. Existing research also considered hashtag grouping, semantic persistence, event classification, and the balance between endogenous and exogenous attention drivers.
- Twitter research: Twitter research has addressed network topology, message relationships, information propagation, credibility, and population mood.
- Hashtag classification: Previous studies discussed classifying popular trends or hashtags into groups and examining how semantic differences affect hashtag persistence.
- Temporal dynamics: Popularity-peak shapes have been used to classify events, while power-law-like growth and decline exponents were proposed as possible universality-class indicators.
- Exogenous and endogenous drivers: Distinct popularity classes are thought to reflect both endogenous social-network propagation and exogenous injection through mass media.
3. DATA
The study analyzes a large Twitter dataset at the scale of daily hashtag activity, detects isolated popularity peaks, and connects temporal profiles with semantic content. It focuses on popular hashtags and uses WordNet-based processing to represent tweet concepts.
- Dataset: 130 million tweets from November 20, 2008 to May 27, 2009 cover about 6.1 million unique user accounts.
- Dataset: A directed follower network was constructed from follower and friend queries covering 2.7 million users with available neighbor information.The queries involved 3.5 million users; private profiles account for the discrepancy.
- Hashtag selection: Filtering hashtags used by at least 500 distinct users produced about 1.7 million tweets containing 402 popular hashtags.
- Temporal scope: Daily activity is analyzed because the chosen time scale captures events meaningful over days while excluding minute-level, circadian, and multi-week dynamics.
- Temporal scope: Hashtag profiles at the daily scale typically show continuous activity, periodic activity, or an isolated peak.The analysis concentrates on hashtags with activity concentrated around an isolated popularity peak.
- Activity peak detection: Peak detection compares each day’s tweet count with a 61-day median baseline, regularized by n_min = 10, and marks peaks when p(i0) > 10.The method retains the highest peak per hashtag, ignores peaks separated by less than one week, and offsets selected peaks to day 0.
- Semantic grounding: Semantic grounding removes mentions, hashtags, URLs, and stop words, then applies stemming, lemmatization, WordNet lookup, and language filtering.The procedure identifies about 18,000 distinct concepts associated with the hashtags under study.
4. CLASSES OF POPULAR HASHTAGS
The paper classifies hashtag popularity peaks using coarse-grained activity before, during, and after the peak, yielding four stable temporal classes linked to event semantics. These classes distinguish anticipatory, reactive, symmetric, and single-day collective attention patterns.
- 4.1 Identifying Classes: Four clusters emerge by fitting a Gaussian mixture model to the before-peak and after-peak activity fractions.The number of clusters is supported by Bayesian Information Criterion and 10-fold cross-validation.
- 4.1 Identifying Classes: 77% of hashtags have classification accuracy below 5%, while only 6% exceed 20%.
- 4.1 Identifying Classes: The four clusters represent activity concentrated before and during the peak, during and after it, symmetrically around it, or almost entirely on the peak day.
- 4.1 Identifying Classes: These temporal classes correspond to distinct event patterns, including anticipated events, unexpected events, movie releases, and transient speeches or service failures.
- 4.2 Social Semantics of Classes: Before-peak activity is associated with social events and time periods, whereas after-peak activity includes unexpected events and marketing concepts such as “free” and “evidence”.
- 4.2 Social Semantics of Classes: Peak-day activity corresponds to short-lived sport and media events, while class-level content selectivity may support tagging based on popularity dynamics.
5. INFORMATION SPREADING
Hashtag classes exhibit distinct information-spreading patterns, with retweeting, seeding, follower adoption, and user persistence varying across activity profiles. Exogenous injection appears important for post-peak activity, while follower infectiousness is broadly similar across classes.
- Retweeting: Retweet fractions are higher for hashtags symmetric around the peak or concentrated on the peak day, indicating greater endogenous activity.Hashtags active before the peak are less prone to viral spreading and instead reflect anticipatory behavior.
- Seeding: Hashtags concentrated after the peak tend to have more seeders, suggesting propagation fueled by exogenous publicity or mass-media communication.Seeders are users who adopt a hashtag without prior use by followed users.
- Follower adoption: β does not depend strongly on hashtag class, with a median value of about 0.02 for the fraction of followers adopting the hashtag.The study interprets β as a measure of meme infectiousness.
- Sampling caveat: The estimated propagation parameters are affected by follower-network sampling, although β is relatively stable and γ is more sensitive.Because the network sampling is fixed, the authors consider cross-hashtag comparisons of γ legitimate.
- Persistence: The average user-level spreading interval τ is similar across classes except for peak-day activity, which has the lowest value.τ measures the hours between a user's first and last tweet containing the hashtag, with τ = 0 for single-use hashtags.
6. DISCUSSION
The study finds robust dynamical classes of hashtag attention that correspond to distinct content semantics and different balances between endogenous propagation and external seeding. It also presents a scalable classification approach while noting that further validation is needed for applications.
- Dynamical classes: Four robust hashtag classes capture the four ways activity is distributed before, during, and after the day of peak usage.The classification remains stable under small perturbations.
- Semantic and behavioral patterns: Pre-peak hashtags usually concern scheduled events or specific times, whereas symmetric profiles appear associated with endogenous social-network propagation.Post-peak tails correspond to unexpected events or exogenous driving.
- Content and propagation: Hashtags about swine flu and the Oscars show different popularity profiles despite high media attention, while both exhibit high external seeding and relatively low endogenous propagation.The authors link the differing dynamics to the hashtags' different social semantics.
- Applications and scope: Simple daily-popularity parameters yield a scalable classification that can support discovery of behavioral patterns in large-scale activity records.The parameters require past activity and lack predictive power.
- Applications and scope: Using the classes for implicit temporal tagging or timeline annotation requires extensive validation beyond the paper's scope.The authors explicitly identify validation as unfinished work.
A. HASHTAG USAGE
The hashtag examples are organized into four activity-profile classes based on whether usage occurs before, after, at, or both before and after the popularity peak. The examples span events, media, disruptions, conventions, sports, political topics, games, and applications.
- Activity before peak: The activity-before-peak class includes scheduled events, contests, charity events, conventions, sports, and other anticipated occasions.Examples include the Shorty Awards, a radio-show anniversary, a Brazilian technology festival, and the World Baseball Classic.
- Activity after peak: The activity-after-peak class includes disruptions, media topics, Twitter applications, games, and job-related hashtags.Examples include Amazon's ranking controversy, a school shooting, a television competition, and a counter-movement to #FollowFriday.
- Activity at peak: The activity-at-peak class includes holidays, media finales, disruptions, political events, sports, and advertising-related hashtags.Examples include April Fools' Day, a Sydney blackout, the NFL draft, the State of the Union, and the Oscars.
- Activity before and after peak: The activity-before-and-after-peak class includes conventions, games, disruptions, political meetings, technology expos, media topics, and applications.Examples include 25C3, bushfires, CeBIT, Davos, and the Save Chuck campaign.