Abstract
Traditional social media analysis relies on explicit interactions like retweets or follows to map influence. However, these networks are often sparse or incomplete. We present the Latent Influence Network (LIN), a method to uncover hidden, directed influence links by analyzing the timing and nature of shared user activities. Using the Latent Influence Detection Framework (LIDET), we demonstrate that LIN significantly outperforms existing coordination-based and explicit networks, achieving up to 99% accuracy in behavioral prediction tasks across diverse datasets including COVID-19 stance detection and rumor propagation.
Introduction
On platforms like X (formerly Twitter), influence is typically measured through overt actions. Yet, many users are influenced by content they consume without direct interaction. Existing "Latent Coordinated Networks" (LCN) identify groups of similar users but fail to capture the directionality of influence—who is leading and who is following. This paper addresses this gap by proposing LIN, which maps the directional flow of information based on temporal activity patterns.
Methodology: The LIDET Framework
The authors introduce the Latent Influence Detection Framework (LIDET) to systematically identify and validate these hidden links.
The framework consists of four main stages:
- Candidate LIN Construction: Directed edges are created from User A to User B if B performs a shared activity (e.g., using a specific hashtag or URL) within a specific time window ($W$) after A.
- Parameter Optimization: The system iterates through various time windows and frequency thresholds to find the configuration that best reflects real-world influence.
- Node Embedding: Techniques like DeepWalk and Node2Vec convert the network structure into high-dimensional vectors representing user positions.
- Behavioral Validation: The resulting network is tested by its ability to predict future user behavior or stances (e.g., pro-vaccine vs. anti-vaccine) using machine learning classifiers.
Experimental Evaluation
The framework was evaluated across four diverse social media contexts:
- COVID-19 Vaccination: Stance detection (supportive, neutral, or opposed).
- Twitter15 & Twitter16: Standard benchmarks for rumor detection.
- IRA Dataset: Data involving coordinated Russian "troll" accounts.
- Digg: A news aggregator dataset used to test influence through "likes."
Results and Discussion
The LIN model demonstrated a clear advantage over both traditional Directed Retweet Networks (DRN) and undirected Latent Coordinated Networks (LCN).
Key findings include:
- Superior Accuracy: On the COVID-19 dataset, LIN achieved a staggering 99% classification accuracy, far exceeding the performance of traditional DRNs.
- Directionality as a Key Signal: By capturing the order of events rather than just similarity, LIN provides a more nuanced view of information flow than LCN models.
- Activity Sensitivity: Retweets and URL sharing were found to be the most potent signals for influence. Conversely, "Replies" were often too sparse to build a robust network, sometimes leading to lower performance.
Conclusion
The Latent Influence Network (LIN) provides a powerful tool for mapping the invisible structures of social media influence. By moving beyond explicit follower graphs, researchers can better understand how information spreads and how communities form around shared interests. Future work will focus on incorporating temporal decay to account for the waning strength of influence over time.
Summary updated