Source-linked AI summary
#FailedRevolutions: Using Twitter to Study the Antecedents of ISIS Support
Walid Magdy, Kareem Darwish, Ingmar Weber
TL;DR
To understand the antecedents of expressed ISIS support, the paper analyzes Arabic Twitter users’ stances and histories. It uses name variants to classify support or opposition and examines pre-ISIS tweets; its classifier predicts later positions with 87% accuracy, while the authors identify frustration with Arab Spring missteps as a source of support.
Problem
The paper asks what distinguishes Arab Twitter users who explicitly support or oppose ISIS and what may prompt support, rather than assuming support means joining the group.
Source: Introduction · p. 2Method
The authors classify users by the ISIS name forms in their tweets, then analyze user histories and use pre-ISIS tweets to predict later expressed support or opposition.
Source: Introduction · p. 2, Classifying Tweeps · p. 5Results
The classifier predicted future support or opposition with 87% accuracy using pre-ISIS tweets, and hashtag patterns suggested frustration with Arab Spring missteps as a source of support.
Source: Conclusion · p. 10Takeaways & Limitations
Supporters’ Twitter activity was concentrated on voicing support, and their hashtags pointed to frustration with Arab Spring missteps.
Source: Conclusion · p. 10Because each user’s before-period ends at their first on-topic tweet, generalizing to the population is difficult, and early supporters may differ from later ones.
Source: Limitations · p. 10
Abstract
from arXiv · showhide
Within a fairly short amount of time, the Islamic State of Iraq and Syria (ISIS) has managed to put large swaths of land in Syria and Iraq under their control. To many observers, the sheer speed at which this "state" was established was dumbfounding. To better understand the roots of this organization and its supporters we present a study using data from Twitter. We start by collecting large amounts of Arabic tweets referring to ISIS and classify them into pro-ISIS and anti-ISIS. This classification turns out to be easily done simply using the name variants used to refer to the organization: the full name and the description as "state" is associated with support, whereas abbreviations usually indicate opposition. We then "go back in time" by analyzing the historic timelines of both users supporting and opposing and look at their pre-ISIS period to gain insights into the antecedents of support. To achieve this, we build a classifier using pre-ISIS data to "predict", in retrospect, who will support or oppose the group. The key story that emerges is one of frustration with failed Arab Spring revolutions. ISIS supporters largely differ from ISIS opposition in that they refer a lot more to Arab Spring uprisings that failed. We also find temporal patterns in the support and opposition which seems to be linked to major news, such as reported territorial gains, reports on gruesome acts of violence, and reports on airstrikes and foreign intervention.
Introduction
The paper studies what distinguishes Arab Twitter users who explicitly support or oppose ISIS, focusing on the antecedents of expressed support rather than enlistment. It examines users’ profiles and pre-ISIS histories to identify distinguishing language, predict later positions, and explore potential motivations.
Introduction
- The authors frame the question against research finding that people who join violent militant organizations are not necessarily psychologically disordered or otherwise easily distinguished by personal traits.The cited discussion instead describes grievance narratives in which violence is cast as confronting perceived structural injustices.Source: Introduction · p. 1
- The study asks what features of Arab Twitter users’ profiles distinguish those who later explicitly support or oppose ISIS, and what perceived injustices may prompt support.Its focus is support expressed on Twitter, not necessarily joining the group.Source: Introduction · pp. 1–2
- The paper reports three contributions: distinguishing language for current ISIS stances, an 87% accurate classifier of future support or opposition, and potential reasons for support and sources of opposition.The classifier uses users’ pre-ISIS data, according to the paper’s overview.Source: Introduction · p. 2
- The analysis first examines global pro- and anti-ISIS tweet trends in relation to external events, then studies users’ histories with emphasis on the pre-ISIS period.The overview names reports of territorial gains and videos of beheadings as events of interest.Source: Introduction · p. 2
Background
The background situates ISIS’s online presence within militant media strategies and studies of political polarization on social media. It distinguishes this paper by examining a wider set of users and looking back at the antecedents of support or opposition.
ISIS and the Syrian Conflict Online
- ISIS’s social-media strategy includes multilingual accounts and videos, with military successes and implementation of Sharia prominent in its propaganda.Prior accounts attribute recruitment success largely to military victories, particularly the capture of Mosul.Source: ISIS and the Syrian Conflict Online · p. 2
- Earlier Twitter research on Syria identified four communities, including Islamists and ISIS supporters, while this study sought a wider user set and historical antecedents.The cited network analysis covered 652 accounts before ISIS’s major territorial gains; this study reports a 57,000-user analysis drawn from an initial 165,000.Source: ISIS and the Syrian Conflict Online · p. 2
- A study of Twitter reactions to the 2013 Sarin Gas Attacks found no immediate opinion polarization and sparse communities that could not produce substantial social pressure for opinion change.Other cited work used geo-tagged Syrian tweets to evaluate bias across data sources and sampling methods.Source: ISIS and the Syrian Conflict Online · p. 3
Using Online Data to Study Political Polarization
- Prior work on Twitter polarization and political violence describes differences between camps, but the cited studies do not examine antecedents or what leads users to support one side.The paper positions its pro- versus anti-ISIS analysis as a study of online polarization.Source: Using Online Data to Study Political… · p. 3
Data Collection
The study collected Arabic tweets mentioning ISIS and inferred users’ stances from the name forms they used. It then assembled user histories and applied selection criteria intended to retain engaged users with clear positions.
Data Collection
- The collection began with Arabic tweets found through continuous Twitter REST API searches, then gathered tweets mentioning ISIS by full-name or abbreviated variants.The paper notes that the Streaming API language operator would have sharply reduced the data because it was applied after sampling down to 1% of tweets.Source: Data Collection · p. 3
- The researchers collected 3.1 million Arabic tweets mentioning ISIS from more than 250,000 users between October 13 and December 31, 2014.They obtained historic tweets for 165,000 of 180,000 users identified during the initial period; 15,000 accounts were unavailable by November 17.Source: Data Collection · p. 3
- A native Arabic-speaking judge labeled random samples of 1,000 tweets per name-form condition as positive, negative, or neutral to assess stance.Positive tweets praised the group, spread its messages, or reported positively; negative tweets attacked it or spread negative news, while neutral tweets merely reported news or included spam.Source: Data Collection · p. 3
- Full-name variants indicated support in 93% of tweets, while abbreviated-name variants generally indicated opposition in 77%.Among acronym-using tweets, 7.5% were supportive and 15% showed no specific leaning; Table 1 summarizes stance percentages by name form.Source: Data Collection · p. 3, Table 1 · p. 4
- The authors retained 56,960 tweeps who authored at least 10 ISIS-mentioning tweets out of 3,200 tweets and strictly used either the long or abbreviated ISIS name in at least 70% of their ISIS-mentioning tweets.These conditions were meant to ensure that study users were engaged in the topic and had voiced clear positions toward ISIS.Source: Data Collection · p. 4
Trends
Pro- and anti-ISIS tweet patterns shifted over time and peaked around different kinds of news. Their hashtags reflected distinct interests, with one shared focus on the Sinai attack.
Trends
- The fraction of pro-ISIS tweets steadily increased while the anti-ISIS fraction decreased, although the groups’ tweet counts were only weakly correlated (Pearson’s r = 0.27).The authors say it is unclear whether the increase in pro-ISIS tweets was organic.Source: Trends · p. 4, Figure 3 · p. 5
- Pro-ISIS tweets generally peaked with propaganda releases and military achievements, while anti-ISIS tweets peaked with reports of hostage killings, torture, and Yazidi women’s enslavement.The event-linked shifts are observable, but their strength is unclear and causality is difficult to determine.Source: Trends · p. 4
- Pro-ISIS hashtags covered threats and victory claims, media outlets and releases, territorial news, and message-spreading through popular hashtags.Examples include #Furqan Foundation and #we are all Moaz.Source: Trends · p. 4, Figure 4 · p. 5
- Anti-ISIS hashtags exposed violence, threatened ISIS, attacked its sympathizers and leaders, and sometimes used pro-ISIS hashtags to attack the group.Examples include #ISIS stones woman and #banning ISIS on Twitter.Source: Trends · p. 4, Figure 4 · p. 5
- The groups’ hashtag interests were generally different, but both focused on the Sinai attack through #Sinai State and #Egypt cemetary for ISIS.The authors suggest ISIS’s brutal tactics may galvanize opposition, but say their effectiveness as a recruitment tool is unclear.Source: Trends · pp. 4–5
Tweeps
The section retrospectively distinguishes likely ISIS supporters from opponents using pre-declaration tweets, then compares their hashtag themes, examples, locations, and tweeting patterns.
Classifying Tweeps
- Pre-declaration tweets separated potential supporters from opponents, and the classifier distinguished the groups with relatively high accuracy.The analysis used tweets before users first explicitly stated their position; the training sample included 7,225 supporters and an equal number of opponents.Source: Classifying Tweeps · p. 5, Table 2 · p. 6
Pre-ISIS hashtags
- Future supporters’ hashtags emphasized Arab Spring uprisings, opposition to regional regimes, conflict areas, and civil-rights concerns, with few religious references.The authors also noted hashtags expressing disillusionment with the status quo.Source: Classifying Tweeps · p. 6
- Future opponents’ hashtags were mostly general, while political tags often supported the UAE and Egyptian regimes; Gaza was a shared concern with future supporters.Other examples expressed support for Gaza or hostility toward the Muslim Brotherhood.Source: Classifying Tweeps · pp. 6–7
Post-declaration hashtags
- Among current supporters, hashtags reflected animosity toward Middle Eastern regimes and referenced ISIS media, names, battles, and organizations.The analysis excluded tweets mentioning ISIS when identifying these post-declaration hashtags.Source: Classifying Tweeps · p. 7
- Current opponents’ hashtags included support for Syrian rebel or Islamic groups, support for some regional regimes, opposition to Islamic groups, and Shia-related references.The authors attributed the rebel-group theme to ISIS attacks on opposition groups in Syria.Source: Classifying Tweeps · p. 7
Example tweets and tweeps
- Examples of eventual supporters included people associated with the Egyptian and Libyan revolutions, opposition to Egypt’s coup, or support for militant groups in Syria.Examples of eventual opponents included supporters of Muslim rebel groups in Syria, Egyptian-regime supporters, and a Shia fighter against ISIS.Source: Example Tweets and Tweeps · p. 7
Geographic distribution
- Only 4.4% of pro-ISIS users could be mapped to a country, versus 24% of anti-ISIS users; the authors caution that pro-ISIS country patterns are inconclusive.The country-ratio analysis included only countries with at least 10 mapped pro-ISIS users and was described as proof of concept.Source: Geographic Distribution of Tweeps · pp. 8–9
Focus and time analysis
- Pro-ISIS users devoted 20% of their tweets to ISIS versus 4.5% for opponents; nearly 32% of pro-ISIS users joined Twitter in September–October 2014, coinciding with US bombing.The authors said other factors may have contributed to the account-creation spike and that further investigation was needed.Source: Figure 6 · p. 9, Focus and Time Analysis of Tweeps · p. 9
Limitations
The study focuses on ideological supporters rather than fighters and excludes Western-directed recruiting and press-release messaging. Missing accounts, Twitter’s history limit, and the user base constrain its scope and generalizability.
Limitations
- The study concerns ideological supporters, not fighters on the ground, and does not identify who will engage in violence.The authors describe their focus as the “soil” on which violent behavior can flourish.Source: Limitations · p. 9
- The Arabic-language focus excludes recruiting and press-release messages deliberately aimed at Western audiences, which fall outside the study’s antecedents question.The authors describe studying such marketing efforts as interesting in its own right but outside the paper’s motivation.Source: Limitations · p. 9
- Nearly 15,000 of 180,000 tweeps lacked historic data, likely biasing the sample toward less offensive users; histories also stop at users’ latest 3,200 tweets.The authors do not expect users’ deletion of individual old tweets to be frequent.Source: Limitations · p. 9
- Twitter users may not represent the general population, although surveys may also miss people unwilling to voice militant support openly.The authors note that Twitter’s perceived anonymity lends itself to studying such phenomena.Source: Limitations · p. 9
- Representativeness could potentially improve by inferring support likelihood for a broader user base, including users whose ground truth is unknown.The proposed classifier would use content and network features and could apply to users who might never express an opinion on the topic.Source: Limitations · pp. 9–10
- Because each user’s before/after cutoff depends on their first on-topic tweet, population-level generalization is difficult and supporter characteristics may shift over time.The authors say classifier performance suggests there may not be many temporal subgroups with vastly different characteristics, but identify this as future work.Source: Limitations · p. 10
Conclusion
Using Twitter data, the study examines antecedents of ISIS support and opposition. Pre-ISIS tweets predict future expressed positions, while supporter and opposition patterns differ in their topics and temporal associations.
Conclusion
- 87% accuracy in predicting future ISIS support or opposition using only pre-ISIS tweets.The prediction concerns future expressed positions, not necessarily joining ISIS.Source: Conclusion · p. 10
- Supporters’ hashtags point to frustration with failed Arab Spring revolutions as a major source of support.The authors contrast this with opposition linked to other rebel groups, Middle Eastern regimes, and Shia sectarianism.Source: Conclusion · p. 10
- Opposition to ISIS is linked with support for other rebel groups, mostly in Syria, that ISIS has targeted.The authors also link opposition with support for existing Middle Eastern regimes and Shia sectarianism.Source: Conclusion · p. 10
- At the global temporal level, spikes in support or opposition show some connection to reports of territorial gains or losses and accounts of torture or forced female enslavement.The authors describe these as connections, not causal effects.Source: Conclusion · p. 10
- The clearest temporal trend was that supporters seemed to join Twitter to voice support, reflected in their temporal concentration and topical focus.Source: Conclusion · p. 10
- The study also reports geographical and temporal trends among both pro- and anti-ISIS tweeps.Source: Conclusion · p. 10