Coupled Graph Neural Networks for Predicting the Popularity of Online Content

Qi Cao, Huawei Shen, Jinhua Gao, Bingzheng Wei, Xueqi ChengarXiv 2019 · OpenReview

Introduction

Predicting whether online content will go "viral" is a critical yet challenging task for social media platforms and marketers. The difficulty lies in the complex "cascading effect," where information spreads through a social network via iterative interpersonal influence. This paper presents CoupledGNN, a novel framework that leverages two intertwined Graph Neural Networks (GNNs) to capture both user activation states and the dynamics of influence spread, significantly improving popularity prediction accuracy.

The Challenge of Cascading Effects

Traditional methods for predicting popularity often rely on hand-crafted features (like follower counts) or sequence-based deep learning models that treat cascades as simple temporal paths. However, these approaches frequently overlook the underlying social structure—the "empty spaces" of inactive users who represent potential future paths for the cascade.

The cascading effect: active users (red) influence neighbors, potentially leading to future activations (yellow) among previously inactive users (white).

To accurately predict the final reach of a post, a model must understand not just who has already shared it, but who is likely to share it next based on their position in the network and the influence exerted by their active neighbors.

The CoupledGNN Framework

The core innovation of this work is the CoupledGNN architecture, which uses two distinct but communicating GNNs to model the diffusion process.

The CoupledGNN architecture, featuring intertwined State and Influence GNNs connected by gating mechanisms.

Dual-Network Architecture

  1. State Graph Neural Network: This component models the probability of a user transitioning from inactive to active. It aggregates the influence from a user's neighbors to determine their current state.
  2. Influence Graph Neural Network: This component models the latent influence spread. It recognizes that influence is not uniform; it varies between users and evolves as the cascade progresses.

The Coupling Mechanism

The two networks are linked through two primary gating mechanisms:

  • Influence Gate: Uses influence representations to weight the impact of a neighbor's activation on a target user.
  • State Gate: Ensures that only users who are currently active can exert influence on their neighbors.

By stacking $K$ layers of these coupled GNNs, the model naturally simulates information spreading $K$ hops away from the initial seed nodes.

Experimental Evaluation

The authors evaluated CoupledGNN on a massive real-world dataset from Sina Weibo, involving millions of users and hundreds of millions of relationships.

Key Findings

  • Superior Performance: CoupledGNN achieved a 10% improvement in Mean Relative Square Error (MRSE) over DeepCas, the previous state-of-the-art deep learning method.
  • Structural Awareness: The model's performance peaked at $K=3$ layers, aligning with the observation that 99.76% of activated users in the Weibo dataset were within three hops of the early adopters.
  • Robustness: The framework maintained high accuracy even when 20% of the social network edges were randomly removed, demonstrating its resilience to incomplete graph data.

Conclusion

CoupledGNN represents a significant step forward in popularity prediction by explicitly modeling the interplay between user states and interpersonal influence. While future work could integrate temporal dynamics and content analysis (NLP), the current structural-aware approach provides a robust and highly accurate foundation for understanding how information goes viral in complex social networks.

Summary updated

BibTeX
@misc{cao2019coupled,
  author = {Qi Cao and Huawei Shen and Jinhua Gao and Bingzheng Wei and Xueqi Cheng},
  title = {{Coupled Graph Neural Networks for Predicting the Popularity of Online Content}},
  year = {2019},
  url = {https://openreview.net/forum?id=vO23OpMD7A},
}
Figure