WeSeer: Visual Analysis for Better Information Cascade Prediction of WeChat Articles

Quan Li, Ziming Wu, Lingling Yi, Kristanto Sean N, Huamin Qu, Xiaojuan MaIEEE TVCG 2018 · DOI · arXiv

TL;DR

This paper introduces WeSeer, a visual analytics system designed to improve the prediction of information cascade popularity on WeChat. By integrating visual reasoning with a point process-based model (SEISMIC), WeSeer allows domain experts to understand and fine-tune prediction parameters, leading to more accurate and interpretable results for various article types.

What Problem Does This Paper Address?

Social media platforms like WeChat are flooded with information, making it challenging for articles to gain user attention. Predicting which articles will become popular ("viral") is crucial for recommendation systems and for identifying potentially harmful content like rumors. Existing predictive models often rely on complex, manual feature engineering or act as "black boxes," making them difficult to understand, debug, and improve. Furthermore, these models often struggle to adapt to the diverse propagation patterns of different types of articles, leading to inaccurate predictions.

Methods

The core of WeSeer is an enhancement of the SEISMIC point process-based model. SEISMIC, while more interpretable than feature-based models, has limitations in parameter tuning and adaptability. WeSeer addresses these by:

  • Speed-Adjusted Infectiousness: Instead of a fixed decay factor for article infectiousness over time, WeSeer uses the article's propagation speed to dynamically adjust infectiousness. This accounts for the fact that WeChat articles don't always become stale and can even have delayed outbreaks.
  • Bounded Mean Degree Estimation: WeSeer allows users to interactively estimate the mean degree of the network (representing the average number of friends a user has) within a "safe zone" derived from the adjusted infectiousness. This provides a more flexible and article-specific estimation of this crucial parameter.
  • Visual Reasoning Interface: The WeSeer system provides a multi-view interface that visualizes the prediction process. This includes:
  • Prediction View: Shows propagation speed, infectiousness (original and adjusted), and compares predictions from SEISMIC and WeSeer.
  • Exploration View: Allows users to perform "what-if" analyses by simulating the deletion or addition of users to understand their impact on infectiousness and mean degree.
  • Propagation View: Visualizes sharing channels and hierarchy, providing insights into user behavior.

Datasets

The paper uses real-world WeChat article data for training and evaluation. Specifically, they collected:

  • 7509 articles published in the first week of March 2016, with sharing histories until the first week of April 2016.
  • A ground-truth list of the top 100 most shared articles.
  • 814 newly published WeChat articles from September 21, 2017, with over a week of propagation data.

Key Insights and Ideas

  • Dynamic Article Behavior: WeChat articles don't always follow a simple "stale over time" decay. Some can have delayed outbreaks or sustained popularity, requiring models that adapt to these varied patterns.
  • The Power of Visual Interaction: Experts can gain deeper insights into the SEISMIC model's behavior and parameter effects by visually exploring predictions and performing "what-if" scenarios, rather than just looking at raw output.
  • "Big Nodes" vs. "Small Nodes": The influence of users with many friends ("big nodes") versus those with fewer ("small nodes") significantly impacts article propagation and the estimation of network parameters.
  • Adaptability is Key: A one-size-fits-all approach to parameter tuning (like a fixed mean degree) is insufficient. WeSeer's ability to dynamically estimate parameters based on article-specific propagation data is crucial for accuracy.

Key Results

  • Improved Prediction Accuracy: WeSeer consistently achieved better prediction accuracy (lower Absolute Percentage Error - APE) compared to the original SEISMIC model, especially in the early stages of an article's propagation. For example, WeSeer could provide a prediction in 20 minutes, while SEISMIC took 150 minutes for a popular article.
  • Enhanced Breakout Article Coverage: WeSeer was significantly better at identifying articles that would become highly popular (top 100). It covered approximately 50% of top articles within 60 minutes, compared to SEISMIC's 10%.
  • Expert Validation: Domain experts from WeChat found WeSeer effective for understanding model mechanisms, exploring parameters, and improving prediction results, with interactions described as "effective and efficient."

Limitations and Future Work

  • Inductive Article Content: The current system does not utilize the actual content of the articles (text, sentiment, etc.), which could further improve predictions.
  • Sharing Probability: While WeSeer infers differences in article penetration, explicitly modeling individual user sharing probabilities could enhance accuracy.
  • Complete Social Graph: Access to a complete WeChat social graph would allow for more precise calculation of newly exposed users.
  • Daily Posting Cycles: The system doesn't currently account for daily patterns in user activity.
  • "Slow Onset" Articles: Articles with very delayed outbreaks, occurring long after the initial observation window, might still be missed. The authors suggest dynamically resizing this window.
  • Model Generalization: Adapting WeSeer to other point process-based models might require revisions to the parameter exploration interface.

Summary updated

BibTeX
@article{li2018weseer,
  author = {Quan Li and Ziming Wu and Lingling Yi and Kristanto Sean N and Huamin Qu and Xiaojuan Ma},
  title = {{WeSeer: Visual Analysis for Better Information Cascade Prediction of WeChat Articles}},
  journal = {IEEE Trans. Vis. Comput. Graph.},
  year = {2018},
  doi = {10.1109/tvcg.2018.2867776},
}
Figure