Source-linked AI summary
Real-Time Classification of Twitter Trends
Arkaitz Zubiaga, Damiano Spina, Raquel Martínez, Víctor Fresno
TL;DR
Prior work examined social trends over the long term, while their real-time lists offer little context about what triggered them. This paper classifies trends early using language-independent social-diffusion features, which outperform tweet-content features and require no external data. The authors also identify behavioral patterns associated with trend types.
Problem
Prior research focused on long-term social conversations, leaving little known about trend origins and early categorization despite the value of understanding what triggered a trend.
Source: INTRODUCTION · p. 1Method
The paper introduces four trend-trigger categories and uses straightforward language-independent features based on social diffusion to classify trends as they emerge.
Source: INTRODUCTION · p. 2, CONCLUSIONS · p. 13Results
The proposed features produce more accurate classification than tweet-content features, using tweets sent before the topic trended and no external data.
Source: CONCLUSIONS · p. 13Takeaways & Limitations
Feature analysis reveals behavioral patterns associated with trend types, and the features can help discriminate the trigger behind a trending topic.
Source: CONCLUSIONS · p. 14Twitter’s trend-detection algorithm is not public, so the study cannot establish that its trend identification is optimal; evaluating alternative algorithms was outside its scope.
Source: DISCUSSION · p. 12
Abstract
from arXiv · showhide
Social media users give rise to social trends as they share about common interests, which can be triggered by different reasons. In this work, we explore the types of triggers that spark trends on Twitter, introducing a typology with following four types: 'news', 'ongoing events', 'memes', and 'commemoratives'. While previous research has analyzed trending topics in a long term, we look at the earliest tweets that produce a trend, with the aim of categorizing trends early on. This would allow to provide a filtered subset of trends to end users. We analyze and experiment with a set of straightforward language-independent features based on the social spread of trends to categorize them into the introduced typology. Our method provides an efficient way to accurately categorize trending topics without need of external data, enabling news organizations to discover breaking news in real-time, or to quickly identify viral memes that might enrich marketing decisions, among others. The analysis of social features also reveals patterns associated with each type of trend, such as tweets about ongoing events being shorter as many were likely sent from mobile devices, or memes having more retweets originating from a few trend-setters.
INTRODUCTION
Twitter trends can reveal current events, but their term lists do not explain what triggered them. This work proposes early classification of trends by trigger using efficient, language-independent social-diffusion features.
INTRODUCTION
- Trend lists reveal terms discussed unusually often but provide no context about their triggers, despite potential value for users, journalists, governments, and marketers.Prior work focused on long-term conversations, leaving little research on early trend origins and type classification.Source: INTRODUCTION · p. 1
- The paper develops an automatic classifier to categorize trends as they emerge, aiming to provide end users with a filtered, real-time view.Its features are straightforward, language-independent, require no external data, and have low computational cost.Source: INTRODUCTION · p. 2
- The proposed features outperform a baseline using tweet content, while feature-distribution analysis reveals behavioral patterns associated with each trend type.The paper presents the approach as efficient and suitable for real-time use.Source: INTRODUCTION · p. 2
BACKGROUND
The background describes Twitter’s interaction conventions and trending topics, then explains that the study uses Twitter-provided trends as classifier input. These conventions include retweets, hashtags, and the platform’s real-time trend list.
- Twitter’s short tweets and rapid sharing support quick message creation and broad dissemination; its interaction syntax includes mentions, replies, retweets, and hashtags.Retweets copy an original tweet and can be reshared at multiple levels; hashtags group tweets around topics or common interests.Source: Twitter · p. 2
Trending Topics
- Twitter trending topics are terms discussed much more than usual, reflecting a recent increase in use rather than routine popularity.The platform defines them as immediately popular topics; the extraction algorithm is not public.Source: Trending Topics · pp. 2–3
- A trending topic consists of its term or phrase and the stream of tweets containing it; an example illustrates an early news tweet, resharing, a reply, and a link-sharing tweet.Source: Trending Topics · p. 3
- The study takes Twitter-provided trending topics as input and controls for trend detection, focusing its analysis on classifying trends after detection.Source: Trending Topics · p. 3
TYPOLOGY OF TRENDING TOPICS
The paper defines four trigger categories for trending topics and builds a manually annotated dataset from Twitter trends. The dataset supports classification experiments across a multilingual collection of topics.
Definition of a Typology of Trending Topics
- The typology distinguishes news, ongoing events, memes, and commemoratives, developed through trend monitoring and researcher discussion.Trends that could fit multiple categories were assigned the single category judged to fit best.Source: Definition of a Typology of Trending… · pp. 3–4
- News covers trends caused by newsworthy events; ongoing events arise from live discussion, while memes spread for appeal without being apparently newsworthy or mainstream.Source: Definition of a Typology of Trending… · pp. 3–4
- Commemoratives include birthday congratulations, anniversaries, and memorial days; the authors note that some trends could fit multiple categories but choose one best-fitting label.Source: Definition of a Typology of Trending… · p. 4
Dataset
- Monitoring Twitter’s top trends and searching for their tweets from March 1–7, 2011 yielded 1,036 topics and 567,452 tweets from 348,757 users.Topics averaged about 548 associated tweets and appeared in multiple languages; every topic included tweets in more than one language.Source: Dataset · pp. 4–5
- Four everyday Twitter users annotated trending topics, with each topic labeled by three annotators: 616 ongoing events, 251 memes, 142 news, and 27 commemoratives; Fleiss’ kappa was 0.597.The topics and their annotations were publicly available for further research.Source: Dataset · p. 5
CHARACTERIZATION OF TRENDING TOPICS
This section introduces 15 social features intended to capture diffusion patterns for early trend classification, then examines how features and tweet vocabularies vary across four trend types.
Features of Trending Topics
- The proposed 15 social features are straightforward, language-independent, and cheap to compute for real-time classification; 10 use averages and five use diversity measures.The diversity measures use Shannon’s index to represent variation in feature values across a trend.Source: Features of Trending Topics · pp. 5–6
Understanding the Features
- Retweets are more frequent for memes and, to a lesser extent, news, while some memes and commemoratives have higher hashtag ratios.Across the four trend types, the feature distributions vary, suggesting that the features can discriminate among types to some extent.Source: Analysis of Twitter Features · pp. 6–7, Figure 1 · p. 8
- Ongoing-event tweets are shorter; commemorative trends use more question marks, while news has the lowest question-mark use and high exclamation use.The authors suggest short ongoing-event tweets may reflect limited attention or time, or the effort of typing on a handheld device.Source: Analysis of Twitter Features · pp. 7–8, Figure 1 · p. 8
- User diversity is low across all four classes; retweeted-user diversity is lowest for memes, followed by news, and highest for ongoing events and commemoratives.The authors interpret the meme pattern as indicating a clearer contribution from fewer influential users or trend-setters.Source: Figure 2 · p. 8, Analysis of Twitter Features · p. 8
- Memes have more links than news, which may break on Twitter before links are available on the Web.The authors connect this pattern to the motivation for discovering breaking news on Twitter.Source: Analysis of Twitter Features · p. 8, Figure 1 · p. 8
- Topic repetition is more pronounced for commemoratives, while some memes and ongoing events spread faster and have higher tweet-per-second frequencies.Repetition does not distinguish news, ongoing events, and memes in the same way.Source: Analysis of Twitter Features · p. 8, Figure 1 · p. 8
- Memes often have low language diversity, whereas ongoing events have the highest; news and commemoratives show higher vocabulary diversity but not higher hashtag diversity.The authors suggest retweets help preserve a meme’s starting language and that longer, wordier news tweets may contribute to vocabulary diversity.Source: Figure 2 · p. 8, Analysis of Twitter Features · p. 9
- Top-term patterns include live-action vocabulary for ongoing events, congratulatory and time-related terms for commemoratives, less-specific meme vocabulary, and frequent named entities in news.The authors note that some terms can overlap across classes, as illustrated by different Charlie Sheen trends.Source: Analysis of Text in Tweets · p. 9
AUTOMATIC CLASSIFICATION OF TRENDING TOPICS
The section evaluates four-class trend classification with one-vs-all SVMs, comparing social Twitter features, tweet text, and a committee combining both. Twitter features and especially the committee perform better overall than bag-of-words, while class-wise precision varies.
Experimental Setup
- Four one-vs-all SVMs classify the four trend types, and the class with the largest classifier margin becomes the prediction.The committee sums the two representations’ outputs for each class before selecting the class with the largest sum; one example shifts a news/meme disagreement to ongoing event.Source: Experimental Setup · pp. 10–11
- Experiments averaged results across 10 random training selections, each using 600 training trends and 436 test trends.The two representations were Twitter features and bag-of-words text.Source: Experimental Setup · p. 10
- Twitter features use 15 dimensions, whereas the bag-of-words representation uses 512,943 dimensions for 1,036 trends and is more computationally expensive.Twitter features are straightforward to compute and their dimension count does not grow with the number of trends represented.Source: Experimental Setup · p. 10
- Classifier committees improved Twitter-feature performance by 2.8%, while Twitter features exceeded bag-of-words accuracy by 3.2%.The accuracy comparison is reported in Table 6.Source: Table 6 · p. 11, Evaluation · p. 11
- Twitter features performed better for ongoing events and especially memes, with the meme precision gap exceeding 9%; bag-of-words performed better for news and commemoratives.The collection included only 27 commemorative trends, which may explain their low precision and could leave room for improvement with more examples.Source: Evaluation · p. 11
- Cohen’s kappa rankings matched the accuracy rankings, indicating that the accuracy improvement was not a consequence of random guesses.The authors use kappa to assess the impact of randomness on agreement between predicted and observed categorizations.Source: Evaluation · p. 12
DISCUSSION
The discussion frames categorized trends as inputs for audience-specific systems and notes both a granularity trade-off and limits on trend-detection validation.
DISCUSSION
- News, ongoing-event, meme, and commemorative labels could serve journalism, event summarization, marketing, social-science research, and analysis of commemorative conversations.The authors describe these as examples of systems that could benefit, not an exhaustive list.Source: DISCUSSION · p. 12
- The four-category typology may be too coarse for some applications, but its categories can be subdivided, such as memes by fan type or news by diffusion source.The paper contrasts multiple contributors with a few authoritative sources as possible diffusion subcategories.Source: DISCUSSION · p. 12
- Trend inputs came from Twitter’s own detector, whose unpublished algorithm prevents the authors from certifying it as optimal; testing alternatives was outside scope.Future work would examine how trend types evolve over subsequent minutes or hours, including possible shifts between types.Source: DISCUSSION · p. 12
RELATED WORK
Prior work addressed tweet classification, credibility, trend dynamics, and specialized trend taxonomies; this paper positions its contribution as worldwide trigger typing at trend onset.
RELATED WORK
- Prior research covered Twitter use, diffusion, and tweet content; tweet-level studies used language-dependent English features, while credibility classification used user and sentiment features.Sriram et al. (2010) reported that their features outperformed bag-of-words for five tweet classes.Source: RELATED WORK · p. 13
- Trend studies often focused on detection or summarization; Asur et al. (2011) distinguished early trend-setters from propagators, while hashtag research modeled meme probability.The hashtag work used temporal behavior, author distribution, and word splitting.Source: RELATED WORK · p. 13
- Other work examined fan- and anti-fan-driven music trends as five meme types, and categorized 99 trending events into Breaking, Ongoing, Meme, Commemorative, or Unknown.The 99-event categorization considered Twitter trends and Bing Popular Now queries.Source: RELATED WORK · p. 13
- A complementary local-event taxonomy separates endogenous trends emerging within Twitter from exogenous trends caused by real events, using a 72-hour-before-and-after-peak window averaging 1,000 tweets.The distinction is presented by Naaman, Becker, and Gravano (2011).Source: RELATED WORK · p. 13
- This work presents a high-level typology for worldwide trending topics and identifies their trigger in real time, rather than analyzing a trend’s whole lifetime.The authors describe these as novel aspects relative to the reviewed work.Source: RELATED WORK · p. 13
CONCLUSIONS
The authors introduce four trigger categories and language-independent social-diffusion features, which outperform tweet-text content in classification using only pre-trend tweets.
CONCLUSIONS
- The paper organizes Twitter trends into news, ongoing events, memes, and commemoratives, represented by 15 language-independent features based on averages and diversity.The features characterize social diffusion of the trending topic.Source: CONCLUSIONS · p. 13
- The proposed social-diffusion features provided more accurate SVM classification than tweet-text content and require only tweets sent before a topic trended, with no external data.The feature count stays fixed as the collection grows, unlike the content-based representation.Source: CONCLUSIONS · p. 13
- Qualitative analysis of the features offers insight into how different happenings spread on social media and supports distinguishing trend triggers.Source: CONCLUSIONS · p. 14