HID: Hierarchical Multiscale Representation Learning for Information Diffusion

Honglu Zhou, Shuyuan Xu, Zuohui Fu, Gerard de Melo, Yongfeng Zhang, Mubbasir Kapadia

Predicting the spread of information—be it a viral meme, a news story, or a marketing campaign—is a fundamental challenge in social network analysis. While deep learning has advanced the field of information diffusion modeling, existing approaches often struggle with the inherent complexity and multi-layered structure of human interactions.

Researchers from Rutgers University, in a paper presented at IJCAI-20, introduced HID (Hierarchical Multiscale Representation Learning for Information Diffusion). HID is a model-agnostic framework designed to enhance existing diffusion models by incorporating multiscale modeling into user representation learning.

The Challenge: Capturing the Forest and the Trees

Most current state-of-the-art models create user "embeddings"—vector representations—at a single, individual scale. This approach often misses macro-level patterns, such as community-level behaviors, and is computationally expensive. Furthermore, training these deep models from scratch can lead to suboptimal solutions (local minima) due to poor initial parameter guesses.

The HID Framework

HID addresses these issues through a two-way learning process: Upscaling and Downscaling.

Overview of the HID framework showing the upscaling and downscaling processes.

1. Upscaling (Macro-level Abstraction)

The framework first groups users into "hyper-users" based on a $k$-order diffusion proximity matrix. Unlike traditional structural graphs, this matrix analyzes actual diffusion paths to identify users with similar information-adoption behaviors. Using clustering techniques like Hierarchical Agglomerative Clustering (HAC), HID creates progressively coarser versions of the network.

2. Downscaling (Knowledge Transfer)

Learning begins at the coarsest scale, where the reduced number of entities allows for rapid training. These coarse representations then serve as the initialization for the next, finer scale. This "warm-start" prevents the model from getting lost in individual-level noise early on, leading to more stable and accurate embeddings.

3. Model Agnosticism

HID acts as a wrapper. It can be integrated with various base diffusion models, such as CDK, CSDK, or Forest, providing them with a superior starting position and a hierarchical "map" of the network.

Experimental Evaluation

The researchers tested HID on three real-world datasets with varying characteristics:

  • Memetracker: Meme and news phrase propagation across blogs.
  • Twitter: Tweet timestamps and authorship.
  • Digg: User voting patterns for front-page stories.

Key Results and Efficiency

The integration of HID consistently improved both the predictive accuracy and the training efficiency of the base models.

Comparison of training time and MAP accuracy for various models with and without the HID framework.
  • Accuracy Boost: HID-enhanced models outperformed standalone baselines across all datasets. Notably, on the Digg dataset, it achieved a 14.76% improvement in Mean Average Precision (MAP).
  • Efficiency Gains: Training times were significantly reduced. As the number of scales ($s$) increased, training time decreased nearly linearly with the reduction in the training corpus size.
  • Behavioral vs. Structural Modeling: HID outperformed general multiscale graph embedding techniques (e.g., HARP), demonstrating that respecting the temporal and behavioral nature of diffusion paths is superior to relying solely on static network structure.

Conclusion and Future Directions

HID demonstrates that hierarchical abstraction is a powerful tool for modeling complex social processes. While there is a "sweet spot" for coarsening—beyond which performance may drop due to loss of detail—the framework offers a robust way to scale up virality prediction. Future work may involve incorporating content-based representations and exploring diverse clustering strategies for different network topologies.

Summary updated

Abstract

Multiscale modeling has yielded immense success on various machine learning tasks. However, it has not been properly explored for the prominent task of information diffusion, which aims to understand how information propagates along users in online social networks. For a specific user, whether and when to adopt a piece of information propagated from another user is affected by complex interactions, and thus, is very challenging to model. Current state-of-the-art techniques invoke deep neural models with vector representations of users. In this paper, we present a Hierarchical Information Diffusion (HID) framework by integrating user representation learning and multiscale modeling. The proposed framework can be layered on top of all information diffusion techniques that leverage user representations, so as to boost the predictive power and learning efficiency of the original technique. Extensive experiments on three real-world datasets showcase the superiority of our method.
BibTeX
@inproceedings{zhou2020hid,
  author = {Honglu Zhou and Shuyuan Xu and Zuohui Fu and Gerard de Melo and Yongfeng Zhang and Mubbasir Kapadia},
  title = {{HID: Hierarchical Multiscale Representation Learning for Information Diffusion}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/468},
}
Figure