False News On Social Media: A Data-Driven Survey

1. INTRODUCTION

• The introduction previews terminology, social platforms, human factors, real-world effects, open challenges, and how the survey’s focus compares with earlier surveys.It positions these topics as context for the survey’s study of false news on social media.Source: §1 · p. 1

1.1 Terminology

• The literature uses overlapping labels for deceptive information, but lacks an agreed precise definition; it distinguishes accidental misinformation from deliberate disinformation and also discusses hoaxes, satire, propaganda, click-bait, and junk news.Satire may cause harm when shared out of context, while junk news can describe a publisher’s overall content rather than a single article.Source: §1.1 · pp. 1–2

1.2 Social media platforms as news outlets

• Social media broadens false news’ impact as consumption shifts online, entry barriers fall, unfiltered low-quality outlets spread, trust declines, and advertising, ideology, bots, and trolls motivate or amplify disinformation [30].Financial and ideological motives are proposed for disinformation websites, while malicious agents such as bots and trolls also contribute to misinformation spread.Source: §1.2 · p. 2

1.3 Human factors

• Naive realism and confirmation bias hinder discernment; social conformity, echo chambers, and personalized feeds reinforce polarized sharing, while factual corrections may worsen misperception.Echo chambers are homogeneous clusters insulated from opposing perspectives, and the cited research identifies close-knit communities as a primary driver of misinformation diffusion.Source: §1.3 · p. 2

1.4 Effects on the real world

• False news has been linked to election impacts [26] [15], a stock-market crisis, Pizzagate-related violence, and vaccine mistrust during Ebola and Zika epidemics.Politics accounts for much of the attention on false news.Source: §1.4 · pp. 2–3

1.5 Challenges

• False news is difficult to distinguish from genuine reporting, cannot be comprehensively human-fact-checked at scale, and is constrained by scarce, incomplete training data and platform limits.These conditions also make early detection important to prevent false news from spreading.Source: §1.5 · p. 3

1.6 Survey Focus

• The survey focuses mainly on 2017–2018 English-language research on Twitter, Facebook, and Sina Weibo, excluding collaborative-platform misinformation, automated fact-checking, and social-bot research [30].A few studies from 2015 and 2016 are also included.Source: §1.6 · p. 3

2. PROBLEM FORMULATION AND METHODOLOGY

• The survey covers false-news detection, diffusion characterization, and mitigation in three parts.It reconstructs diffusion networks from users’ platform interactions and reviews mitigation work, including crowd-sourcing initiatives promoted by major platforms [27].Source: §2 · p. 3
• Detection research is classified by employed technique and considered features, spanning traditional machine learning, deep learning, and other models.Examples include Logistic Regression, Support Vector Machines, Random Forest, Convolutional and Recurrent Neural Networks, Matrix Factorization, and Bayesian Inference.Source: §2 · p. 3
• False-news detection is typically formulated as supervised binary classification using labeled news articles and related social-media posts.Labels are typically sourced from fact-checking organizations or researchers’ manual verification; features may include content, social context, and external knowledge bases.Source: §2 · p. 3

3.1 Content-based

• Horne et al. [24] examine article bodies and titles across true, false, and satire categories, finding real and false news differ substantially in title, while satire and false news are similar in content.In experiments comparing topic- and style-based features with a Random Forest classifier, style-based analysis alone was not effective for detecting false news.Source: §3.1 · pp. 4–5

3.2 Context-based

• Context-based methods use likes, user profiles, interactions, and sharing cascades; Tacchini et al. [68] report high accuracy with little training data, while Yu et al. model cascades for early detection.Wu et al. embed social graphs and use LSTMs to analyze propagation pathways.Source: §3.2 · p. 5

3.3 Content and Context-based

• Combined approaches incorporate news content and social context; Volkova et al. find falsification easier to identify than misleading strategies, and disinformation harder to predict than propaganda or hoaxes.Other approaches combine article text, social responses, and source users, or model publisher-news-user relationships [64].Source: §3.3 · p. 6

3.4 Promising directions

• The survey identifies poor generalization and limited comparability as key gaps, and favors broad linguistic features, social-response diffusion models, and combined psycho-linguistic and social signals.Perez-Rosas et al. require substantial annotated data; Liu et al. and Wu et al. use complex methods on limited datasets; Volkova et al. combine psycho-linguistic signals with Twitter responses.Source: §3.4 · p. 6

• Del Vicario et al. (2016) used the Facebook Graph API to collect a five-year span of data for a quantitative analysis of news consumption related to scientific and conspiracy-theory outlets.Source: §4 · p. 6

5. FALSE NEWS MITIGATION

• Proposed interventions range from curbing highly active, likely bot users to coordinating user flags with fact-checking organizations [27], an approach platforms have offered as a way to combat disinformation without censorship alerts.The survey notes that crowd input can help identify malicious news and prevent its spread.Source: §5 · pp. 8–9
• CURB [27] selects stories for fact-checking to reduce non-credible-news spread, with theoretical guarantees, and its accuracy is highly sensitive to the crowd’s ability to spot misinformation.The authors formulate the problem using temporal point processes and stochastic differential equations, and evaluate precision and misinformation reduction on the Rumors datasets.Source: §5 · p. 8
• Tschiatschek et al. (2018) select a small verification set, learn individual users’ flagging accuracy, and use a propagation-agnostic algorithm that accounts for adversarial spammers.Experiments in a simulated Facebook environment examine learning users’ flagging behavior while considering spammers who promote false news.Source: §5 · p. 8
• Vo et al. (2018) first study active Twitter users sharing fact-checks and propose URL recommendations to encourage these “guardians” to spread credible information.Using Hoaxy tweets referring to fact-checking organizations, they analyze user characteristics and compare a social-structure-aware recommendation model with state-of-the-art collaborative filtering.Source: §5 · pp. 8–9

Conclusions

• False news spreads faster and more broadly than truth on social media, while social bots and echo chambers play important roles in diffusion networks.The authors note that only a few substantial research contributions exist, most focused on characterizing misinformation diffusion, in agreement with.Source: §6 · p. 9
• Textual psycho-linguistic signals aid false news detection, but content alone may be insufficient; social-dimension features should also be considered.Source: §6 · p. 9
• Without agreed gold-standard datasets or research guidelines, ad-hoc collections have proliferated, and detection techniques often lack comparisons and discussion of their consequences.Source: §6 · p. 9
• The many recent contributions show that the research community has responded promptly and can build on prior results to combat false news.Source: §6 · p. 9
• Future work should deepen network-perspective detection insights, coordinate research across disciplines, and develop real-world applications to help combat false news.Source: §6 · p. 9

A. DATASETS

• Fact-checking datasets form a rich but heterogeneous collection, often serving similar objectives through slightly different tasks.Source: A. DATASETS · p. 13
• The survey introduces referenced datasets with brief descriptions, sources, and main references, and summarizes their features in Table 2.Source: A. DATASETS · p. 13
• The survey also presents other interesting datasets beyond those introduced first.Source: A. DATASETS · p. 13

A.1 BuzzFeedNews

• Two collections introduced by Silverman (2016) contain news-item titles and sources and are used in [24].The collections are among BuzzFeed News journalists’ verified false- and true-news datasets shared on Facebook in 2016 and 2017 by hyperpartisan and mainstream news media.Source: §A.1 · p. 13

A.2 BuzzFeed-Webis

• BuzzFeed-Webis adds full content and attached multimedia for shared articles, and was used in.Source: §A.2 · p. 13

A.3 DeClare; A.4 FakeNewsAMT

• DeClare [47] combines Snopes, PolitiFact, and NewsTrust articles with true and false claims, while FakeNewsAMT joins mainstream articles, MTurk-generated false news, and GossipCop claims.Both collections are used for false news detection.Source: §A.3 · p. 13, §A.4 · p. 13

A.5 FakeNewsChallenge

• Fake News Challenge provides thousands of headlines and documents for four-label document-level stance detection, inspired by sentence-level stance work and used in [51].An additional analysis appears in.Source: §A.5 · p. 13

A.6 FakeNewsNet

• FakeNewsNet [62] pairs false and true articles from Snopes and BuzzFeed shared on Twitter with news content and social context, and was employed in [64].Its information includes source, body, multimedia, user profiles, and follower/followee data.Source: §A.6 · p. 13

A.7 Hoaxy

• Hoaxy continuously monitors Twitter diffusion of articles from disinformation and fact-checking sites since 2016 and supports custom data collections; it is used in.Source: §A.7 · p. 13

A.9 Liar

• Liar is a collection of short labeled political statements from PolitiFact for false news classification, employed in.Source: §A.9 · p. 14

A.10 SemEval-2017 Task8

• SemEval-2017 Task8 [11] provides Twitter tweets and replies organized into conversations for stance and veracity resolution of social media content, and was used in [47].Source: §A.10 · p. 14

A.11 Rumors

• Rumor datasets contain Twitter and Sina Weibo retweet and discussion cascades labeled as rumors or non-rumors, used for false news detection and mitigation in [27].They were originally conceived for rumor detection and resolution.Source: §A.11 · p. 14

A.12 Others

• A.12 adds BuzzFace, Facebook disinformation tracking [28], fact-checker and outlet lists [71], and highly shared 2016 election articles.BuzzFace extends prior BuzzFeed datasets with Facebook story comments and webpage content; the IFCN signatory list collects fact-checking organizations operating in different countries.Source: §A.12 · pp. 14–15