Source-linked AI summary
Combating Fake News: A Survey on Identification and Mitigation Techniques
Karishma Sharma, Feng Qian, He Jiang, Natali Ruchansky, Ming Zhang, Yan Liu
TL;DR
Fake news spreads rapidly through social media and affects public opinion, creating a socially important and technically challenging problem. This survey synthesizes identification and mitigation techniques, compiles dataset characteristics, and outlines future research directions. It reports advances across the surveyed literature while noting practical scope constraints, including restricted access to Twitter content and metadata.
Problem
Fake news is a socially consequential and technically challenging problem whose spread on social media affects public opinion.
Method
The survey reviews identification and mitigation methods, including intervention techniques, and compiles characteristic features of available datasets.
Results
The surveyed literature demonstrates significant advances in fake news identification and mitigation, including preliminary evidence that a fake news game reduced perceived reliability and persuasiveness in a randomized field study of 95 high school students.
Takeaways & Limitations
The survey provides consolidated methods, datasets, limitations, and future directions for developing computational solutions to fake news.
Takeaways & Limitations
Twitter-based datasets generally provide tweet and user IDs because platform policy restricts public release of tweet contents and metadata.
Abstract
from arXiv · showhide
The proliferation of fake news on social media has opened up new directions of research for timely identification and containment of fake news, and mitigation of its widespread impact on public opinion. While much of the earlier research was focused on identification of fake news based on its contents or by exploiting users' engagements with the news on social media, there has been a rising interest in proactive intervention strategies to counter the spread of misinformation and its impact on society. In this survey, we describe the modern-day problem of fake news and, in particular, highlight the technical challenges associated with it. We discuss existing methods and techniques applicable to both identification and mitigation, with a focus on the significant advances in each method and their advantages and limitations. In addition, research has often been limited by the quality of existing datasets and their specific application contexts. To alleviate this problem, we comprehensively compile and summarize characteristic features of available datasets. Furthermore, we outline new directions of research to facilitate future development of effective and interdisciplinary solutions.
1 INTRODUCTION
Fake news has become a socially consequential and technically challenging problem as social media accelerates the speed and scale of false-information spread. This survey addresses the problem end to end by covering identification, mitigation, computational challenges, datasets, and future research directions.
- Social media has intensified fake news by accelerating the speed and expanding the scope of false-information dissemination.
- The survey covers both detection and mitigation, including intervention-based techniques, rather than treating fake news solely as a classification task.
- It focuses on computational challenges and compares existing methods that have made significant progress in addressing them.
- The survey consolidates datasets and summarizes their characteristic features to support dataset selection and future dataset collection.
- It concludes by consolidating recent advances and outlining concrete directions for future research.
2 DEFINITION, NATURE, ASSOCIATED CHALLENGES, KEY PLAYERS
The section defines fake news broadly as false information propagated through media, then examines its individual, social, technological, and institutional characteristics. It also surveys how content, user responses, sources, and datasets shape identification challenges.
- Definition: Fake news is defined as a news article or message propagated through media that carries false information regardless of means or motives.
- Definition: The broader definition covers fabricated, misleading, imposter, manipulated, false-connection, and false-context content.
- Nature and challenges: Individuals often struggle to distinguish false from true information, while ideological homophily and algorithmic personalization create echo chambers that can support fake-news spread.
- Identification signals: Detection exploits content characteristics, user engagements, propagation structure, temporal signals, and source or promoter behavior.
- Social context: Social-media growth expands the audience exposed to information and increases the potential for fake news to cause widespread damage.
2.4 Key Players
Fake-news mitigation involves multiple actors operating under competing demands for speed, reliability, openness, and adaptation. The section highlights challenges spanning public susceptibility, propagation, fact-checking, moderation, and adversarial behavior.
- Challenges: Fake-news spread is difficult to study and manage because multiple entities and technological platforms must coordinate while preserving rapid access to quality information.
- Challenges: Maliciously designed content exploits human cognition, emotions, and ideological biases, making both human and computational detection harder.
- Challenges: False information can rapidly reach large audiences, while human fact-checking organizations struggle to keep pace with propagation.
- Moderation: Moderation strategies must avoid overly aggressive restrictions that could harm platforms or limit access to high-quality information and collaborative sharing.
- System goals: Effective systems should combine timely detection with interventions that contain confirmed false information and quantify exposure impacts.
- System goals: Detection should use information incrementally, balancing confidence and accuracy against the timeliness of mitigation.
- Adversarial behavior: Adversaries create new or disposable accounts and coordinate promotion campaigns, requiring systems to adapt dynamically to changing strategies.
3 OVERVIEW OF METHODS
The survey organizes existing work into content-based identification, feedback-based identification, and intervention-based solutions. These categories cover classification from information content, user responses, and active containment or impact mitigation.
- Content-based methods: Content-based methods classify news according to the content of the information being verified.
- Feedback-based methods: Feedback-based methods classify news using the user responses it receives on social media.
- Intervention-based solutions: Intervention-based solutions actively identify and contain false-information spread while mitigating exposure impacts.
4 CONTENT-BASED IDENTIFICATION
Content-based identification distinguishes fake from true news using linguistic cues, engineered features, syntactic analysis, and neural models. These approaches offer interpretable or automated signals but face limits in generalizability, contextual coverage, and automation.
- Hand-engineered cues: Cue-based methods design linguistic and platform-specific features to distinguish fake from true content.Examples include punctuation, sentiment, regular expressions, hashtags, mentions, and enquiry or correction patterns.
- Scientific Content Analysis (SCAN): SCAN applies subjective deception-related content and structural criteria but requires trained professionals, limiting automation.The approach was developed for assessing credibility in transcripts or written statements.
- Limitations of cue-based methods: Linguistic cue sets generalize poorly across topics, languages, domains, and platforms because relevant language and features vary by setting.Designing exhaustive regular-expression patterns and platform-specific features also requires substantial effort.
- Linguistic analysis: n-gram methods represent contiguous words or phrases and classify texts using term-frequency vectors, but alone they miss finer-grained writing-style information.Naive Bayes and SVM classifiers were trained with tokenized n-gram term-frequency inputs.
- Part-of-Speech tags: POS-tag frequencies provide syntactic features and can form a strong baseline, although they capture less information than word-based approaches.In fake-review data, the POS approach outperformed the best human judge but performed below n-grams.
- Probabilistic Context Free Grammar: PCFG-based features encode deeper syntax, but cannot capture context-sensitive information across sentences in longer fake-news texts.Variants differ in which production rules and hierarchical context they include.
5 FEEDBACK-BASED IDENTIFICATION
Feedback-based identification treats user responses, comments, and propagation patterns as potentially more informative than article content. These signals include engagement structure, timing, textual reactions, and user profiles.
- Feedback-based identification: Feedback-based approaches exploit user responses, comments, and propagation patterns because article content may be crafted to avoid detection.The approach regards secondary social-media information as a potential source of stronger evidence than article text alone.
5.1 Hand-engineered features
Hand-engineered feedback features combine user, textual, propagation, and temporal signals for fake-news detection. Their main limitation is restricted generality and complexity of the captured feature space.
- Feature types: Hand-engineered feedback features include user-based, text-based, propagation-based, and temporal-pattern signals.Examples include registration age, follower counts, mention proportions, and retweet-tree depth.
- Limitation: Feature engineering incorporates diverse information but remains limited in the generality and complexity of its feature space.This constrains how broadly the engineered representation can capture feedback signals.
5.2 Propagation pattern analysis
Propagation-based fake news detection uses diffusion cascades, user engagements, and propagation dynamics as signals. Methods range from computationally intensive tree kernels to recursive neural networks and epidemiological process models, each with distinct limitations.
- Propagation cascades: Diffusion cascades represent source posts and their re-shares as propagation trees whose structures and dynamics can support fake news detection.Detection methods analyze paths of re-shares and other propagation patterns in social media engagements.
- Propagation tree kernels: Propagation tree kernels compare cascades by recursively combining similarities between node features and subtrees.Node features include engagement timestamps, comments or replies, and user metadata; classification uses an SVM with the defined tree kernel.
- Propagation tree kernels: Tree-kernel methods improved detection performance over methods that do not use propagation patterns but require pairwise similarities across all trees.This computational cost can prohibit large-scale application to fake news detection.
- Propagation tree neural networks: Recursive neural networks process propagation trees using bottom-up or top-down traversal to extract features for classification.Bottom-up models aggregate leaf information at the root, whereas top-down models combine representations at the leaves after information flows along tree paths.
- Propagation tree neural networks: Both recursive architectures outperform propagation tree kernels and non-tree-structured baselines, while the top-down approach performs better in the reported comparison.The authors attribute this result to its assumed direction of information flow through the diffusion cascade.
- Propagation process modeling: The SEIZ model partitions users into susceptible, exposed, infected, and skeptic groups and estimates propagation dynamics from observed cascades.Its parameters are fitted using nonlinear least squares, but it assumes identical rate constants and probabilities for all users and lacks specific validation for fake news detection.
5.3 Temporal pattern analysis
Temporal pattern analysis examines how the timing, rate, and intervals of engagements differ between true and fake news. Approaches use interval-based handcrafted features or recurrent neural networks, with scalability and feature-design limitations.
- Motivation: Temporal methods leverage differences in the length, rate, and intervals of user engagements to detect fake news.The goal is to extract intricate temporal interactions and variations explicitly.
- Temporal feature models: Handcrafted temporal models divide ordered posts into fixed intervals and aggregate text, user, and propagation features together with changes between adjacent intervals.Example features include punctuation, hashtags, verified-user proportions, followers, reshares, and comments.
- Temporal feature models: These primitive variation models require handcrafted features and restrict temporal variation analysis to numerical features.This limits which properties of engagements can be represented through temporal changes.
- Temporal recurrent neural networks: Recurrent neural network approaches feed ordered sequences of interval-level feature vectors into a model that builds on information from previous timesteps.This enables temporal patterns to be modeled sequentially rather than only through manually specified variations.
- Temporal recurrent neural networks: Ruchansky et al. represent each interval using engagement count, time since the previous non-empty interval, average user features, and textual content.The textual representation can be learned from raw engagement text using doc2vec.
- Temporal recurrent neural networks: An LSTM classifies fake news from these engagement sequences on Twitter and Weibo datasets that lack article contents.The interval count and time gap jointly measure the frequency and distribution of an article’s responses.
- Limitations: Very large response cascades may require sampling or pruning, which can distort their temporal patterns.This creates an efficiency-related boundary for representing cascades with many user responses.
5.4 Response text analysis
Response-text analysis uses user comments and generated responses to supplement article content for fake-news detection, including early detection and joint stance–veracity modeling.
- User responses can reveal article information and may contain negative or questioning reactions useful for detecting dubious articles.
- Deep attention: Attention mechanisms highlight representative response words and provide interpretability, although they do not specifically improve detection performance.
- User response generation: Qian et al. generate sampled user responses from historically learned article-conditioned distributions when no responses are yet available.
- User response generation: Generated response features concatenated with CNN article representations improve detection over using the article representation alone.
- Stance detection: Stance-aware models jointly classify responses as support, deny, query, or comment and classify veracity using shared and task-specific GRU layers.
- Stance detection: Stance annotation remains time intensive even when crowd-sourced.
5.5 Response user analysis
Response-user analysis represents users through profiles, behaviors, and group interactions, while addressing generalization and early-detection limitations in feedback-based methods.
- User features: User features derive from social-network profiles or behavioral patterns in content sharing and responses.
- User features: Behavioral representations support classification by modeling recurring patterns among users engaged with fake or true posts.
- User features: Hand-engineered user features may be platform-specific and should be complemented by user-group behavior analysis.
- User group analysis: Group analysis constructs a weighted user-user co-engagement graph and applies singular value decomposition to obtain lower-dimensional user representations.
- User group analysis: Integrating group-behavior scores improves detection performance, while suspiciousness scores correlate positively with engagements in true and fake cascades.
- User group analysis: New fake accounts and social bots may lack collected behaviors, preventing direct generalization of extracted user features.
- Feedback models trained on late propagation snapshots perform worse early, when fewer user responses are available and states cannot be updated incrementally.
6 INTERVENTION-BASED SOLUTIONS
Intervention-based solutions seek to limit fake-news impact through corrective cascades, competing true-news diffusion, adaptive multi-stage control, network monitoring, and crowd-sourced fact-checking.
- Intervention methods dynamically interpret propagation and update actions rather than classifying news from a static social-media feature snapshot.
- Corrective interventions: Greedy seed selection under Independent Cascade or Linear Threshold diffusion aims to decontaminate at least a β-fraction of exposed users with true news.
- Corrective interventions: Corrective diffusion occurs after fake-news spread, and large damage may make the required, unfixed number of seeds prohibitively costly.
- Competing cascades: Competing-cascade strategies introduce true news while misinformation propagates, formulating influence blocking as optimal true-news dissemination.
- Multi-stage intervention: Multi-stage point-process interventions adapt to observed dynamics, distinguish exposure from sharing, and optimize user incentives under budget and activity constraints.
- Multi-stage intervention: The adaptive method assumes fake news has already been identified and its propagation is being tracked, which is nontrivial.
- Network monitoring: Network monitoring must adapt to changing topologies and can be expensive on large networks.
- Crowd-sourcing: Only 1.3 % of users showed measurable and repeatable skills in reporting actions performed by fake accounts.
7 EXISTING DATASETS
The survey compiles datasets for fake-news research and organizes their characteristics by task context, information content, and user responses. It emphasizes that dataset quality and application differences affect method development and evaluation.
- Dataset quality has often limited the development of novel fake-news detection solutions.
- The compilation covers datasets spanning political statements, open-domain news, multiple label schemes, content types, temporal information, propagation information, and collection sizes.
- Tables 2, 3, and 4 summarize datasets by task and application context, information-content features, and collected user responses.
- The tables report task labels, content domains, annotation schemes, claim counts, label distributions, collection periods, and available engagement features.
8 CONCLUSION AND FUTURE WORK
The survey identifies continuing practical challenges in fake-news identification and mitigation and proposes directions involving dynamic knowledge, intervention strategies, and intent-focused datasets. It also reports preliminary evidence that an active inoculation strategy reduced perceived reliability and persuasiveness of fake news.
- The survey reports significant advances but identifies unresolved practical challenges for computational fake-news identification and mitigation.
- Dynamic knowledge bases: Dynamic knowledge bases must be regularly and automatically updated to reflect changes in a fast-paced world.
- New intervention strategies: Effective intervention strategies require studying how user actions and utilities relate to impacts across different networked environments.
- New intervention strategies: A fake-news game provided preliminary evidence of reduced perceived reliability and persuasiveness in a randomized field study of 95 high school students.
- Datasets for intent detection: Intent-focused datasets could distinguish truly fake news from related categories such as satire and opinion news more finely than binary labels.
A APPENDIX
The appendix consolidates classification-accuracy results for content-based and feedback-based fake-news detection methods across representative datasets. The comparison is explicitly non-exhaustive, and some methods cannot be evaluated when required dataset information is unavailable.
- The appendix consolidates classification accuracy for several fake-news detection methods across representative Twitter and Weibo datasets.
- The reported method-and-dataset list is not exhaustive, and evaluation may be impossible when dataset features required by a method are unavailable.
- Table 5 covers content-based methods that do not use user responses at test time.
- Table 6 covers feedback-based methods involving feature engineering, temporal information, and propagation.