Source-linked AI summary
Fake News Detection on Social Media: A Data Mining Perspective
Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang, Huan Liu
TL;DR
Fake news on social media is difficult to detect and poses societal risks. This survey organizes research on the problem through characterization and detection reviews, concluding that important data and research gaps remain.
Problem
Fake news detection on social media remains an early-stage research problem with challenging unresolved issues despite the societal harms of misinformation.
Method
The survey reviews fake news characterization and detection from a data mining perspective, covering features, models, datasets, metrics, and future directions.
Results
The review synthesizes existing detection approaches, evaluation resources, and open research directions for fake news on social media.
Takeaways & Limitations
Fake news detection research needs further investigation and improved datasets and methods to support continued progress.
Takeaways & Limitations
No agreed benchmark dataset exists, and no public dataset provides all features relevant to fake news detection.
Abstract
from arXiv · showhide
Social media for news consumption is a double-edged sword. On the one hand, its low cost, easy access, and rapid dissemination of information lead people to seek out and consume news from social media. On the other hand, it enables the wide spread of "fake news", i.e., low quality news with intentionally false information. The extensive spread of fake news has the potential for extremely negative impacts on individuals and society. Therefore, fake news detection on social media has recently become an emerging research that is attracting tremendous attention. Fake news detection on social media presents unique characteristics and challenges that make existing detection algorithms from traditional news media ineffective or not applicable. First, fake news is intentionally written to mislead readers to believe false information, which makes it difficult and nontrivial to detect based on news content; therefore, we need to include auxiliary information, such as user social engagements on social media, to help make a determination. Second, exploiting this auxiliary information is challenging in and of itself as users' social engagements with fake news produce data that is big, incomplete, unstructured, and noisy. Because the issue of fake news detection on social media is both challenging and relevant, we conducted this survey to further facilitate research on the problem. In this survey, we present a comprehensive review of detecting fake news on social media, including fake news characterizations on psychology and social theories, existing algorithms from a data mining perspective, evaluation metrics and representative datasets. We also discuss related research areas, open problems, and future research directions for fake news detection on social media.
1. INTRODUCTION
Social media has become a major news source because it is timely, inexpensive, and easy to share, but its lower news quality and powerful dissemination of fake news create distinctive detection challenges. This survey addresses those challenges by clarifying fake-news definitions and characteristics, reviewing detection methods, and identifying open issues and future directions.
- Motivation: Social media news consumption is increasing because it is timely, inexpensive, and easier to share and discuss than news from traditional organizations.62 percent of U.S. adults got news on social media in 2016, compared with 49 percent in 2012.
- Problem: Social media provides a powerful channel for disseminating lower-quality fake news, creating research problems that challenge traditional journalistic norms and automated detection.The problem is difficult because fake news on social media has distinctive characteristics compared with traditional news media.
- Problem: Fake news lacks an agreed definition, requiring clarification to guide detection research and future development.The survey distinguishes narrow and broad definitions that cover most existing definitions in the literature.
- Open issues and future directions: Because fake-news detection on social media remains in an early stage, the survey identifies open issues and proposes future research directions to improve detection and mitigation.The survey’s remaining sections address datasets, evaluation metrics, related research areas, open problems, and future directions.
- Survey contributions: The survey reviews fake-news characterization through psychology and social studies, then presents detection approaches and organizes representative methods into categories.It covers theories and properties of fake news alongside a principled overview of existing detection methods.
2. FAKE NEWS CHARACTERIZATION
This section characterizes fake news through competing definitions, psychological and social foundations, and social-media-specific patterns. It adopts the narrow definition of news that is intentionally and verifiably false while highlighting malicious accounts, echo chambers, and cognitive vulnerabilities.
- Definitions: Fake news is defined as a news article that is intentionally and verifiably false.The survey notes that no agreed definition exists and uses this narrow definition throughout.
- Media Evolution: The fake-news media ecology has shifted from newsprint to radio and television, then to online news and social media.Traditional fake news refers to the problem before social media had important effects on production and dissemination.
- Psychological Foundations: Humans struggle to differentiate real from fake news because psychological and cognitive factors influence belief and correction.Presenting true factual information to correct false information can fail to reduce misperceptions and may increase them among ideological groups.
- Social Media Characteristics: Social-media fake news is enabled by malicious accounts, including social bots, cyborg users, and trolls, whose creation is encouraged by low costs.Social bots are computer-controlled accounts that automatically produce content and interact with humans or other bots.
- Social Media Characteristics: Echo chambers selectively expose consumers to certain news, amplifying difficulties in dispelling fake news.Social credibility and frequency heuristics can make repeated or socially endorsed information seem credible even when it is fake.
3. FAKE NEWS DETECTION
Fake news detection on social media is formulated as a binary prediction task over social news engagements, using a two-phase framework of feature extraction and model construction. The framework combines news-content features with auxiliary social-context signals, including user, post, network, and stance information.
- Feature Extraction: News-content features cover publisher information, headlines, body text, images or videos, and linguistic or visual representations designed to capture discriminative characteristics.Linguistic representations include lexical, syntactic, domain-specific, and deceptive-writing-style features, while visual representations capture characteristics of images and videos.
- Problem Formulation: Social news engagements represent how news spreads among users and their posts over time, with each engagement recording a user, post, and timestamp.If an article has no engagements, its timestamp is set to Null and the user represents the publisher.
- Problem Formulation: Fake news detection predicts whether an article is fake or not from its social news engagements, defining F : E → {0, 1}.The binary formulation models fake news as distortion bias on information manipulated by the publisher.
- Detection Framework: The proposed data mining framework has two phases: feature extraction represents content and auxiliary information, while model construction learns to distinguish fake from real news.Social-media detection supplements news content with social context because engagement data provides information about news veracity.
- Feature Extraction: Social-context features exploit user-driven engagement patterns, including user, group, post, network, and stance-based signals, to infer news veracity.These signals capture spreader characteristics, aggregated group properties, public reactions, network patterns such as echo chambers, and users’ viewpoints in relevant posts.
4. ASSESSING DETECTION EFFICACY
Assessing fake news detection efficacy requires suitable annotated datasets and evaluation metrics. Existing datasets lack complete feature coverage and have important limitations, while metrics such as precision, recall, F1, accuracy, and AUC assess different aspects of classifier performance.
- Datasets: Annotated news data are difficult to produce because determining veracity usually requires domain expertise and careful analysis of claims, evidence, context, and authoritative reports.
- Datasets: Public datasets provide different sources, labels, and features, but no existing dataset supplies all possible features of interest.BuzzFeedNews, LIAR, BS Detector, and CREDBANK illustrate this variation through journalist, fact-checking, browser-extension, and crowdsourced annotations.
- Datasets: Dataset limitations complicate detection, including narrow source coverage, headlines-only content, short statements, non-publisher speakers, and labels that may not represent human annotation.
- Evaluation Metrics: Most approaches formulate fake news detection as classification and evaluate classifiers using accuracy, precision, recall, F1, ROC curves, and AUC.
- Evaluation Metrics: AUC is especially suited to imbalanced fake news classification because it is more statistically consistent and discriminating than accuracy.
5. RELATED AREAS
Related research areas clarify how fake news detection differs from rumor analysis, truth discovery, clickbait detection, and social spammer detection. These areas contribute complementary concepts and signals, including rumor veracity, source credibility, headline-content inconsistency, and user or network activity.
- Section Scope: The related-area review distinguishes each task by its goals and popular methods in relation to fake news detection.The comparison covers rumor analysis, truth discovery, clickbait detection, and social spammer detection.
- Rumor Analysis: Rumor analysis covers rumor detection, tracking, stance classification, and veracity classification, whereas fake news detection focuses on identifying intentionally false information.Rumors may be true, false, or unverified when circulated.
- Truth Discovery: Truth discovery identifies true facts from conflicting sources by jointly estimating source credibility and object truthfulness.Its methods rely on contradicting sources rather than directly exploring fact claims, and can benefit fake news detection in different scenarios.
- Clickbait Detection: Clickbait detection uses linguistic features from teaser messages, linked webpages, and tweet metadata to identify curiosity-driven headlines.Headline-content inconsistency can help detect fake news, although not all fake news includes clickbait headlines.
- Social Spammer Detection: Social spammer detection targets malicious users coordinating attacks and primarily extracts features from user activities and social network information.The rise of social bots has also increased the circulation of false information.
6. OPEN ISSUESANDFUTURERESEARCH
The paper organizes open issues and future research directions for social-media fake news detection into four categories: data-oriented, feature-oriented, model-oriented, and application-oriented. Proposed directions address benchmark data, psychological and multimodal features, more effective models, and diffusion-focused applications.
- Research framework: The survey identifies four research directions: data-oriented, feature-oriented, model-oriented, and application-oriented.These categories are presented as promising directions from a data mining perspective.
- Data-oriented: A comprehensive, large-scale benchmark dataset is needed because no existing dataset includes resources to extract all relevant features.The data-oriented agenda also considers dataset, temporal, and psychological characteristics.
- Data-oriented: Quantitative study of psychological factors remains limited, including how to capture and use echo-chamber effects for fake news detection.The survey also identifies intention detection as promising but limited because research often focuses on authenticity while ignoring intention.
- Feature-oriented: Feature-oriented research should determine effective features from news content and social context, using linguistic-based and visual-based techniques for text information.The survey distinguishes two major data sources: news content and social context.
- Model-oriented: Future model research should improve feature use and effectiveness through aggregation, probabilistic, ensemble, and projection methods.Earlier approaches commonly incorporated extracted features into supervised classifiers and selected the best-performing classifier.
- Application-oriented: Application-oriented research should study fake news diffusion and intervention, including detecting clarifiers, persuaders, and malicious social bots.Spreaders may be human, bot, or cyborg; clarifiers express skeptical or opposing viewpoints, whereas persuaders support fake news.
7. CONCLUSION
The article examines fake news on social media, whose growing use for news consumption has also enabled harmful dissemination. It reviews the problem through characterization and detection phases.
- Conclusion: Social media is increasingly used for news consumption instead of traditional news media.The passage links this shift to the increasing popularity of social media.
- Conclusion: Social media has also been used to spread fake news with strong negative impacts on individuals and broader society.The passage identifies harms to both individual users and society more broadly.
- Conclusion: The article reviews the fake news problem in two phases: characterization and detection.The characterization phase introduces basic concepts and principles of fake news.