Source-linked AI summary
Studying Fake News via Network Analysis: Detection and Mitigation
Kai Shu, H. Russell Bernard, Huan Liu
TL;DR
Fake news spreads widely on social media, where intentional deception, noisy data, diverse users, and malicious accounts complicate analysis. This chapter reviews network properties and network types across content, social, and temporal dimensions, then surveys detection and mitigation methods. It concludes that network analysis can support fake-news detection and mitigation, while knowledge-network approaches remain constrained by incomplete and noisy knowledge graphs and provenance identification remains open and computationally difficult.
Problem
Fake news spreads rapidly on social media and is difficult to address because it is intentionally misleading, data are noisy and heterogeneous, and malicious accounts can amplify propagation.
Method
The chapter surveys content, social, and temporal network dimensions, network types, and representative network-based methods for fake-news detection and mitigation.
Results
The reviewed approaches use network relations and representations to perform fake-news detection and mitigation across the news dissemination ecosystem.
Takeaways & Limitations
Network analysis provides a framework for combining heterogeneous information about entities, relations, propagation, and credibility in studying fake news.
Takeaways & Limitations
Knowledge-network assessment is limited by incomplete and noisy graphs, while fake-news provenance paths remain an open problem and the associated optimization can be NP-complete at scale.
Abstract
from arXiv · showhide
Social media for news consumption is becoming increasingly popular due to its easy access, fast dissemination, and low cost. However, social media also enable the wide propagation of "fake news", i.e., news with intentionally false information. Fake news on social media poses significant negative societal effects, and also presents unique challenges. To tackle the challenges, many existing works exploit various features, from a network perspective, to detect and mitigate fake news. In essence, news dissemination ecosystem involves three dimensions on social media, i.e., a content dimension, a social dimension, and a temporal dimension. In this chapter, we will review network properties for studying fake news, introduce popular network types and how these networks can be used to detect and mitigation fake news on social media.
1 Introduction
Social media’s low-cost, rapid, and accessible news sharing enables fake-news propagation while creating detection challenges. The chapter organizes network-based detection and mitigation around content, social, and temporal relations.
- Social media accelerates news consumption and sharing through low cost, easy access, and rapid dissemination.
- Fake news is difficult to detect because it is intentionally misleading and social-media data are large-scale, multimodal, user-generated, noisy, and often anonymous.
- The news dissemination ecosystem comprises content, social, and temporal dimensions representing relationships among information, participants, and behavior over time.
- Network representation learning and knowledge-network operations can support fake-news detection by exploiting auxiliary social, temporal, interaction, and factual information.
- Network estimation and influence-minimization strategies are introduced for mitigating fake-news effects.
2 Network Properties
Network properties shape how fake news is encountered, interpreted, and spread on social media. Echo chambers, user roles, filter bubbles, and malicious accounts therefore matter for detection and mitigation.
- Like-minded social groups can form echo chambers that reinforce preferred narratives and facilitate fake-news belief.
- Persuaders spread fake news, gullible users are easily influenced, and clarifiers provide opposing viewpoints or true news.
- Filter bubbles personalize information using behavioral data, reducing exposure to contradictory viewpoints and reinforcing existing views.
- Social bots, trolls, and cyborg users can amplify fake-news dissemination, with some accounts controlled by computer algorithms.
3 Network Types
The chapter distinguishes homogeneous networks, which use one node and link type, from heterogeneous networks that encode multiple entities and relations. These structures represent social, content, credibility, factual, stance, and interaction information for fake-news analysis.
- Homogeneous Networks: Homogeneous networks include friendship, diffusion, and credibility networks, each potentially supporting fake-news detection and mitigation.
- Homogeneous Networks: Friendship networks represent social relations among users and reveal community structure underlying news spreading.
- Homogeneous Networks: Diffusion networks encode information-propagation direction, probability, and timing, helping represent spreading paths and temporal patterns.
- Homogeneous Networks: Credibility networks connect social-media posts through supporting or opposing relations to evaluate news truthfulness from post credibility scores.
- Heterogeneous Networks: Heterogeneous networks combine entity and relation types; common forms include knowledge, stance, and interaction networks.
- Heterogeneous Networks: Knowledge networks support fact-checking by testing whether news claims can be inferred from linked entities and relations.
- Heterogeneous Networks: Stance networks aggregate supporting, opposing, denying, and questioning viewpoints to infer news veracity.
- Heterogeneous Networks: Interaction networks model publishers, news items, and users, whose characteristics and interactions can differentiate fake news.
4 Fake News Detection
Fake-news detection is formulated as classification using features extracted from content, users, credibility, publisher bias, and interactions. Network embedding combines these signals into latent representations for prediction.
- Fake-news detection consists of feature extraction followed by model construction that predicts a news label.
- Interaction networks embed different entity types in a shared latent space to provide representations for fake-news detection.
- News embedding uses non-negative matrix factorization to project document-word relations into a low-dimensional latent semantic space.
- User embedding applies non-negative matrix factorization to user-user adjacency information, using negative sampling for unobserved links.
- User-news embedding incorporates user credibility because lower-credibility users are described as more likely to spread fake news.
- Publisher-news embedding represents publisher characteristics and predicts partisan bias because highly biased publishers are more likely to publish fake news.
- The finalized model combines these components to learn latent representations of news and users for fake-news classification.
4.2 Temporal Diffusion Representation
Temporal diffusion representation models how users engage with news over time by combining engagement features with recurrent processing. The resulting news representation supports fake news detection and related tasks.
- Temporal diffusion networks: A diffusion network combines temporal user engagements with a friendship network.Each engagement records a user, posting time, and post, while the friendship network supplies social connections.
- Temporal representation learning: RNNs learn representations of sequential user engagements by processing standardized engagement features in temporal order.A fully connected embedding layer converts heterogeneous raw features before the RNN receives the sequence.
- News representation: The final hidden state is transformed into a news representation used for fake news detection and related tasks.The representation is defined as v_j = tanh(W_rh_m + b_r).
4.3 Friendship Network Embedding
Friendship network embedding captures users’ social structure, latent interests, and community organization, complementing temporal representations that omit direct user dependencies. Community-aware methods are motivated by polarized groups among fake news spreaders.
- Social representation: Friendship networks preserve users’ structural relationships, including first-order and higher-order proximity.These representations address the loss of direct user dependencies in temporal news representations.
- Social representation: Users connected through friendship networks are expected to share similar latent interests because social relationships commonly exhibit homophily.This motivates learning standardized user representations from network structure.
- Community structure: Community structure is important because fake news spreaders are likely to form polarized groups.Modularized Nonnegative Matrix Factorization models proximity, community mapping, and modularity.
- Credibility assumption: The credibility of a news event is assumed to relate closely to the credibility of its relevant social media posts.A news veracity score can be obtained by averaging the credibility scores of relevant posts.
Credibility Network Propagation
Credibility network propagation estimates news veracity by modeling post credibility and viewpoint relations. It initializes node scores, encodes supporting or opposing links, and iteratively propagates credibility until the posts’ values converge.
- Network construction: A credibility network contains post credibility values and viewpoint correlations between pairs of posts.The network optimizes node credibility and infers the credibility of corresponding news items.
- Network initialization: Initial node credibility scores come from pretrained classifiers, while links encode whether post viewpoints support or contradict one another.Same-viewpoint posts form supporting relations; different viewpoints form opposing relations.
- Viewpoint relations: Jensen-Shannon distance measures the difference between posts’ topic-viewpoint distributions.These distributions support determining whether posts express the same or opposing viewpoints.
- Network optimization: The optimization encourages supporting posts to have similar credibility and opposing posts to have contrasting credibility.A regularization parameter trades off relational smoothness against fidelity to initial values.
- Credibility propagation: Credibility values are updated iteratively through T(t) = μHT(t−1) + (1−μT0) until convergence.The final news credibility is the average credibility of its posts.
4.4 Knowledge Network Matching
Knowledge network matching uses external structured facts to assess whether claims in news content are truthful. Claims are represented as subject-predicate-object triples and checked by finding matching paths in knowledge networks.
- Knowledge-based fact-checking: Knowledge networks serve as auxiliary sources for fact-checking news claims.The goal is to match claims with facts represented in the networks.
- Claim matching: A news claim can be represented as a subject-predicate-object triple and checked by finding a matching path in a knowledge network.This provides a network-based procedure for assessing the truthfulness of claims.
Path Finding
Path finding evaluates knowledge-network paths connecting a claim’s subject and object to estimate truth. Specificity, path capacity, and edge costs support truth evaluation through flow optimization, but incomplete or noisy networks require additional reconstruction tasks.
- Knowledge-stream path finding: Knowledge streams are sets of paths from subject s to object o that can be evaluated to estimate a claim’s truth value.Claims involving more specific entities are described as more likely to be true.
- Knowledge-stream path finding: The specificity measure uses entity degrees, while path evaluation can assign truth value 1 when s already exists or optimize for a shortest connecting path.
- Flow optimization: Each edge carries knowledge capacity and usage cost, and the objective finds maximum knowledge flow between s and o at minimum cost.A path’s capacity is limited by its bottleneck edge.
- Limitations: Incomplete or noisy knowledge networks can prevent exact entity-path matches, so entity resolution and link prediction are needed alongside path finding and flow optimization.
Stance Network Aggregation
Stance networks aggregate user, post, and news relationships to infer news veracity. Semi-supervised updates combine labeled news, user beliefs, news priors, and explicit or inferred stance signals.
- Network representation: A stance network connects users, social media posts, and news items, treating stance links as signals for aggregating and inferring news veracity.
- Semi-supervised aggregation: User and news items are represented with Beta distributions whose parameters are iteratively updated from like actions and prior constants.The priors encode users’ tendencies to believe fake or true news and the ratio of fake or true news.
- Semi-supervised aggregation: The model is trained semi-supervised using labeled fake and true news alongside unlabeled news assigned neutral labels.
- Veracity prediction: Aggregated stance information updates the parameters used to predict news veracity through q_j.
- Implicit stance: Implicit stances can be inferred from posts with a supervised model, and the resulting scores serve as inputs for fake-news classification.Existing work often uses handcrafted linguistic or deep latent features, while fewer studies use a network perspective.
5 Fake News Mitigation
Network-based mitigation seeks to reduce fake-news influence by identifying influential sources and spreaders, estimating affected populations, blocking users, and promoting true-news exposure. These approaches rely on diffusion and intervention models, but provenance discovery remains difficult in large-scale networks.
- Mitigation goals: Fake-news mitigation minimizes spreading scope by identifying provenances and persuaders, estimating affected populations, blocking users, and launching mitigation campaigns.
- Identifying provenances: The Information Provenance problem selects at most k sources maximizing propagation utility to a target set P.
- Identifying provenances: Provenance paths are usually unknown, and prior approaches are presented as methods that may be adapted for fake-news research.
- Identifying provenances: Under the independent cascade model, provenance-path utility is the product of propagation probabilities, but extracting the maximum-utility subgraph is NP-complete.The problem is difficult because only a few target nodes may be given and diffusion graphs are usually large-scale.
- Identifying provenances: Existing provenance approaches use node centralities, including degree and closeness, to identify likely transmitters.
- Identifying persuaders: Leadership-based approaches model users and posts bipartitely, while selecting K leaders corresponds to a cardinality-constrained submodular maximization problem with a constant-factor hill-climbing approximation.
- Estimating impact: Impact estimation adapts network scale-up methods, but linked social-media data can make estimates from different methods non-independent.
- Influence minimization: Intervention includes blocking users under edge-specific activation probabilities and campaigns that increase true-news exposure for users exposed to fake news.Blocking can reduce sender out-links, receiver in-links, or activation probabilities; campaigns aim to make users less likely to believe fake news completely.
6 Summary
This chapter reviews recent network-analysis approaches to fake news on social media, organizing dissemination around content, social, and temporal dimensions. It uses these network types to explain methods for fake-news detection and mitigation.
- Fake-news dissemination involves content, social, and temporal dimensions whose relations motivate network analysis.These dimensions characterize the entities and dependencies involved during dissemination.
- The chapter introduces representative network-analysis methods for detecting fake news on social media.
- The chapter also presents representative methods for mitigating fake-news dissemination using networks formed from these dimensions.