Source-linked AI summary
Polarization and Fake News: Early Warning of Potential Misinformation Targets
Michela Del Vicario, Walter Quattrociocchi, Antonio Scala, Fabiana Zollo
TL;DR
Misinformation on social media is associated with polarization and confirmation bias, creating a need to identify potential fake-news targets early. The paper introduces a framework that detects polarizing content and incorporates user-behavior information into fake-news classification. On Italian Facebook data, it reports 77% accuracy for identifying susceptible topics and presents the approach as a basis for real-time monitoring, while noting limitations in identifying disinformation or propaganda.
Problem
Polarization and confirmation bias are linked to misinformation spreading on social media, motivating earlier identification of topics that may become fake-news targets.
Method
The framework extracts topic, sentiment, and user-behavior features to identify polarizing content and use topic susceptibility to misinformation in fake-news classification.
Results
77% accuracy identifies polarizing topics susceptible to misinformation on a massive Italian Facebook dataset.
Takeaways & Limitations
The approach could support real-time monitoring that warns about delicate topics and represents a step toward mitigating misinformation and smoothing polarization online.
Takeaways & Limitations
Identifying disinformation or propaganda requires due caution, and fake or unsubstantiated information may also be disseminated by official newspapers.
Abstract
from arXiv · showhide
Users polarization and confirmation bias play a key role in misinformation spreading on online social media. Our aim is to use this information to determine in advance potential targets for hoaxes and fake news. In this paper, we introduce a general framework for promptly identifying polarizing content on social media and, thus, "predicting" future fake news topics. We validate the performances of the proposed methodology on a massive Italian Facebook dataset, showing that we are able to identify topics that are susceptible to misinformation with 77% accuracy. Moreover, such information may be embedded as a new feature in an additional classifier able to recognize fake news with 91% accuracy. The novelty of our approach consists in taking into account a series of characteristics related to users behavior on online social media, making a first, important step towards the smoothing of polarization and the mitigation of misinformation phenomena.
1 INTRODUCTION
The paper frames polarization and confirmation bias as linked to misinformation, motivating an early-warning framework for polarizing content and future fake-news topics. It evaluates this approach on Italian Facebook data while emphasizing caution about automated truth classification.
- Problem context: Dissenting information is often ignored, limiting the effectiveness of debunking, fact-checking, and similar responses.
- Motivation and contribution: The authors position trustworthy information gathering as a scientific challenge for meaningful participation in public debate and societal decision making.
- Motivation and contribution: 77% accuracy identifies polarizing topics susceptible to future misinformation in a massive Italian Facebook dataset.The dataset includes more than 300K news items from official newspapers and 50K posts from websites disseminating fake or unsubstantiated information.
- Motivation and contribution: The framework uses polarization and confirmation bias as a proxy for determining potential targets of hoaxes and fake news.
- Scope and caution: The approach is presented as extensible to other social networks and microblogging platforms, but the authors caution against relying on machine-learning algorithms alone to separate truth from falsehood.
- Scope and caution: The authors acknowledge that identifying disinformation or propaganda requires due caution despite the approach's benefits for misinformation.
2 RELATED WORK
Prior work addresses misinformation through behavioral, content, and credibility-based approaches, but polarization and user behavior remain central considerations. This paper contributes early detection of potential fake-news topics, behavior-aware features, and a platform-extensible framework.
- Prior work: Prior studies target misinformation through suspicious behavior, automated debunking, credibility assessment, and hoax detection across several platforms.
- Polarization and behavior: Users tend to focus attention on limited sets of pages, producing sharp community structures among news outlets.
- Polarization and behavior: Structural properties of content-propagation cascades alone are insufficient for automatic misinformation classification, motivating inclusion of user behavior.
- Contributions: The framework identifies polarizing content promptly to predict future fake-news topics and supplies susceptibility to misinformation as a classifier feature.
- Contributions: The paper claims the first work addressing early detection of possible fake-news topics.
- Contributions: New features account for how news are presented and perceived on the social network.
- Contributions: The framework is designed to extend easily to different social media platforms.
3 A FRAMEWORK FOR THE EARLY WARNING
The framework identifies polarizing content and potential misinformation targets by combining content sentiment, user engagement, and entity-level features, then evaluates classifiers for early warning.
- Framework overview: The four-phase pipeline collects official and fake-source data, extracts entities and sentiment, defines behavioral features, and classifies topics susceptible to misinformation.It is designed for prompt use across platforms such as Facebook and Twitter.
- Features: A user is engaged with entity e when more than 95% of their comments are on posts containing e.A user may be engaged with more than one entity.
- Features: The framework measures presentation distance, mean response distance, controversy, perception, and captivation for each entity.Presentation distance compares the maximum and minimum sentiment; mean response distance compares post and comment sentiment means.
- Features: Controversy, perception, and captivation use data-dependent thresholds δp, δr, and ρe to convert entity-level measures into indicators.Captivation is set to 1 when the engaged-user fraction ue reaches ρe, and 0 otherwise.
- Features: Thresholds are selected by comparing all entities with disputed entities above candidate values for each feature.The resulting pairs are (Eδp, Dδp), (Eδr, Dδr), and (Eρe, Dρe).
4 A REAL USE-CASE: FACEBOOK
The Facebook use-case applies the framework to Italian official and fake-news pages, extracting entities, sentiment, and behavioral features to identify controversial topics and potential misinformation targets. The analysis finds distinct polarization patterns, rapid transitions from official news to fake news, and strong classification performance using behavioral features.
- 4.1 Data collection: The dataset combines Italian official newspapers with websites disseminating hoaxes, unsubstantiated information, or fake news, using posts, likes, and comments collected from 31.07–12.12 2016.Pages were manually checked, and the data collection used public Facebook pages.
- 4.2 Topic extraction and sentiment analysis: Topics are represented as extracted entities, while the second sample additionally includes sentiment scores from comments.Entities may denote persons, works, organizations, places, events, or concepts.
- 4.3 Features: Presentation distance separates entities into low-distance uncontroversial and high-distance controversial groups, enabling data-dependent thresholds for the analysis.The reported thresholds are δp(e) = 1.1 for E1 and δp(e) = 0.98 for E2.
- 4.3 Features: Higher response distance is associated with a lower probability that an entity is disputed, whereas about 90% of the most viral posts are disputed.The response-distance ratio decreases, while the disputed-entity ratio increases with the share of engaged users.
- 4.4 Classification: The classifiers accurately recognize undisputed topics but perform less well on disputed ones; LOG and NN achieve particularly low false-positive rates for disputed entities.The study compares LIN, LOG, SVM, KNN, NN, and DT after balancing the classes through resampling.
- 4.4 Classification: Presentation distance is the best-performing feature in nearly all cases, while presentation distance, response distance, controversy, perception, and captivation consistently rank among the strongest features.The exception is logistic regression on E2, where response distance appears among the first three features.
- 4.5 Insights: Entities first appearing in official news generally emerge in fake news within about 24 hours, and about 50% of roughly 2K official-news entities also appear in hoaxes or fake news.The same passage reports that this subset typically appears on fake news only after first appearing in official news.
5 CAN WE DETECT FAKE NEWS?
The framework's predicted features are added to feature sets for fake-news classification on Facebook posts, alongside structural, semantic, user-based, and sentiment-based features. Across samples and classifiers, the results show high classification performance, with predicted features among the best-performing features and a best accuracy of 0.91.
- Feature construction: The framework output is converted into four predicted features based on the number and rate of predicted disputed entities across samples E1 and E2.The features use the respective parameters and best classifiers from the early-warning framework.
- Evaluation design: Experiments A and B evaluate classifiers by progressively adding semantic, user-based, sentiment-based, and predicted features to structural features on samples P1 and P2.Experiment A tests one added category at a time, whereas experiment B adds categories sequentially.
- Classification results: 0.91 is the best accuracy, achieved by logistic regression on P2, with precision of 0.88 for fake news and 0.94 for not fake information.The classifiers generally detect not-fake information more accurately, while both false positive rates are significantly low, especially for logistic regression.
- Classification results: The accuracy pattern has exceptions for KNN on P1, where step 2 is highest, and DT on P1, where accuracy decreases at the final step.These exceptions qualify the general effects of feature addition.
- Feature ranking: Predicted features rank among the best-performing features in both samples, while structural and semantic features are most represented and mean presentation distance appears in both samples.Feature relevance is assessed using AUC and forward stepwise selection.
6 CONCLUSIONS
The paper presents a framework for timely identification of polarizing content to predict future fake-news topics and support fake-news classification. Validated on official news and hoaxes from Facebook, the approach is presented as promising for real-time monitoring, while its extension to other platforms is described as straightforward and its use for disinformation or propaganda requires caution.
- Contributions: The framework enables timely identification of polarizing content to predict future fake-news topics and build a fake-news classifier.The authors describe this as the paper's central contribution.
- Validation: The methodology is validated on a massive dataset of official news and hoaxes collected from Facebook.The conclusions frame the validation specifically around Facebook data.
- Scope: The approach is described as straightforward to extend to other social media platforms.This is presented as a scope of applicability rather than a reported cross-platform evaluation.
- Limitations: Identification of disinformation or propaganda remains a limitation requiring caution because fake or unsubstantiated information may also be diffused by official newspapers.The conclusion explicitly retains this limitation despite the reported results.
- Implications: The results are presented as promising for systems that monitor information flow in real time and issue warnings about delicate topics.The authors connect the framework's output to warning about potential fake-news topics.