Abstract
While identifying misinformation spreaders is a well-studied problem, predicting the actual impact of their content remains a challenge. This paper presents a novel approach to predicting the influence of news-related tweets, categorized by retweet counts, for both real and fake news spreaders. By leveraging a Random Forest classifier integrated with user history and sentiment analysis, the proposed method achieves state-of-the-art performance, outperforming existing benchmarks and graph-based models. The findings reveal distinct spreading mechanisms: real news propagates through established authority, whereas fake news relies on social interaction.
Introduction
The ability to identify misinformation spreaders on social media has improved significantly, yet predicting the impact of their posts remains elusive. Not every tweet from a known misinformation actor goes viral. Current literature often focuses on binary classification (fake vs. real) or general retweet prediction, leaving a gap in understanding the specific influence of news articles shared by different types of users. This research addresses this gap by quantifying the "influence" of news spreaders through retweet count prediction, helping to assess the actual threat level posed by various actors.
Methods
The authors frame the problem as a multi-class classification task, grouping the number of retweets into five buckets: 0, 1-10, 11-100, 101-1000, and 1000+.
Feature Engineering
To power their model, the researchers engineered three categories of features:
- User Profile Features: Standard metadata like follower/following counts, account age, and "listed" count.
- History-based User Features: Aggregated data from the user’s last 500 tweets, including average engagement and emotional profiles using the NRC Emotion Lexicon and VADER.
- Tweet-based Features: Characteristics of the specific news tweet, including a novel metric: the cosine similarity between the current tweet’s sentiment/emotion and the user’s historical average.
Model Selection
Several models were tested, including K-Nearest Neighbors and Multi-layer Perceptrons. A Random Forest classifier with class weighting (to handle data imbalance) was found to perform best.
Datasets
The study utilized the Politifact dataset from the FakeNewsNet repository. Using the Twint scraper, the authors gathered historical data for 43,119 Fake News Spreaders (FNS) and 135,234 Real News Spreaders (RNS). Users with fewer than 100 tweets or those who shared both real and fake news were excluded to ensure data quality.
Results
The proposed approach established a new state-of-the-art for predicting news spreader influence. For Real News Spreaders (RNS), the model achieved a weighted F1 score of 0.931. For Fake News Spreaders (FNS), it achieved an F1 score of 0.853.
As shown in the table above, the Random Forest model significantly outperformed standard Graph Neural Network (GNN) approaches like GCN, GAT, and SAGE_GCN, which struggled with F1 scores ranging from 0.33 to 0.64. This suggests that for this specific task, rich feature engineering of user behavior is more predictive than the raw network structure of the follower graph.
Key Insights
Authority vs. Interaction
A key insight from the study is the divergence in drivers of influence. For Real News Spreaders, the most important features were account age and the number of URLs—suggesting that "established authority" drives their reach. For Fake News Spreaders, the key drivers were user interaction metrics (replies, following count, and total statuses), suggesting their influence is built on "social engagement" rather than institutional trust.
Predictive Asymmetry
It is consistently harder to predict the influence of fake news spreaders (F1: 0.853) than real news spreaders (F1: 0.931). This implies that the spread of misinformation may be more volatile or driven by external factors not captured in user history.
Conclusion and Future Work
This research demonstrates that user history and sentiment analysis are powerful tools for predicting the viral potential of misinformation. While the current study is limited to Twitter's retweet mechanics and the Politifact dataset, future work could explore other platforms like TikTok or Facebook and incorporate deeper semantic analysis of news content to further refine predictive accuracy.
Summary updated