Abstract
Studies on detecting and understanding the spread of unreliable news on social media have identified key characteristic differences between reliable and unreliable posts. These differences in language use also vary in expression across individuals, making it important to consider personal factors in unreliable news detection. The application of personalization methods for this has been made possible by recent publication of datasets with user histories, though this area is still largely unexplored. In this paper we present approaches to represent social media users in order to improve performance on three tasks: (1) classification of unreliable news posts, (2) classification of unreliable news spreaders, and, (3) prediction of the spread of unreliable news. We compare the User2Vec method from previous work to two other approaches; a learnable user embedding layer trained with the downstream task, and a representation derived from an authorship attribution classifier. We demonstrate that the implemented strategies substantially improve classification performance over state-of-the-art and provide initial results on the task of unreliable news prediction.
Introduction
Traditional automated misinformation detection systems primarily focus on the content of individual posts—the "what." However, linguistic expression is deeply personal; the stylistic nuances of skepticism or endorsement vary significantly between individuals. This research argues that by ignoring a user's historical behavior and unique linguistic signature, models overlook a critical dimension: the "who." By shifting focus toward modeling the user, we can better identify "unreliable information spreaders"—individuals who habitually propagate content from low-credibility sources—and even predict future spreading behavior.
Methodology and Personalization Strategies
To capture a user's unique digital fingerprint, the study explores four distinct personalization strategies integrated into a unified detection framework:
- User2Vec: Generates user embeddings by aggregating word embeddings (GloVe or Word2Vec) from a user's entire historical corpus, capturing general vocabulary and thematic tendencies.
- Authorship Attribution (AA): Utilizes SBERT to encode posts, which are then processed through a feed-forward layer to identify stylistic signatures. This approach tests if spreaders possess a unique "stylistic fingerprint."
- Task Embeddings: A learnable embedding layer that optimizes a unique vector for every user ID directly during the training of the downstream detection task.
- T-BERT: A baseline method using a truncated BERT model to process the first 512 tokens of a user's post history.
Experimental Setup
The models were evaluated using two longitudinal datasets:
- FACTOID: A Reddit-based dataset with 4,150 users and 3.4 million posts covering contentious topics like COVID-19 and climate change.
- Twitter Dataset: A collection of 3,541 users and 2.6 million tweets, recompiled from previous misinformation studies to allow for historical analysis.
Key Insights and Results
The experimental results demonstrate that personalization significantly outperforms content-only or non-personalized baselines.
Style and Sentiment as Indicators
Authorship Attribution (AA) emerged as a powerful signal, suggesting that the way a user writes is often as indicative of unreliability as what they write. Furthermore, linguistic analysis revealed a "sentiment downward spiral": as users post more unreliable content, their language becomes increasingly negative and emotive.
Performance Gains
Personalized approaches yielded substantial improvements over previous state-of-the-art (SOTA) methods:
- User-Level Detection: On the FACTOID dataset, an SVM combined with User2Vec achieved an F1 score of 0.75 (up from the previous SOTA of 0.61). On Twitter, AA with Logistic Regression reached an F1 of 0.82.
- Predictive Modeling: The study provides pioneering results in predicting future behavior. On the Twitter dataset, the Task Embedding approach achieved an F1 score of approximately 0.85 for predicting whether a future post would be unreliable.
Conclusion and Future Directions
While personalization offers a significant leap in detection accuracy, challenges remain regarding label noise (relying on domain-level credibility) and the need for multi-modal analysis (incorporating images and video). Furthermore, the ethical implications of identifying "unreliable spreaders" necessitate careful implementation to avoid surveillance or the suppression of legitimate discourse. Future research will focus on integrating social network dynamics to understand how user interactions further influence the spread of misinformation.
Summary updated