Presented at ICWSM 2020, this research introduces WoMG (Word-of-Mouth Generator), a framework designed to create synthetic social media data that accurately reflects real-world behavioral dynamics. Unlike traditional models that treat information diffusion as simple biological contagion, WoMG models the complex interplay between user interests, meme topics, and network structure. By aligning user interests with network topology using Non-negative Matrix Factorization (NMF), the framework allows researchers to simulate information spread by tuning parameters such as homophily and virality.
The Challenge of Synthetic Social Data
Studying the spread of misinformation or viral memes is often hindered by the lack of high-quality, accessible data. While synthetic social graphs can be generated, they often lack the content (topics) and the underlying motivations (user interests) that drive real-world sharing. Real human behavior is interest-driven: users share posts because they care about the topic and tend to follow others with similar interests—a phenomenon known as homophily. The core challenge lies in creating synthetic users whose interest vectors are mathematically consistent with their position in the social network.
The WoMG Framework
The WoMG framework operates in three primary stages to ensure realistic data generation:
- Interest Vector Generation: Each user is assigned a $k$-dimensional interest vector. The model uses Symmetric Non-negative Matrix Factorization (NMF) to ensure these vectors reflect the network topology, incorporating second-order proximity (shared neighbors) and a tunable randomness parameter to control homophily.
- Meme Generation: Using a Latent Dirichlet Allocation (LDA) topic model, WoMG generates "memes" characterized by a topic distribution and a bag of keywords. Each meme is assigned a "virality" score from a power-law distribution.
- Propagation Model: The framework employs a Topic-aware Linear Threshold Model. A user shares a meme if the social pressure from neighbors—weighted by mutual interest in the specific topic—exceeds a threshold determined by the meme’s virality.
Key Insights and Results
The authors validated WoMG using real-world datasets, including the Digg 2009 news story votes and the Associated Press (AP) Corpus. By tuning homophily and virality parameters, WoMG generated cascades that closely matched the macro-statistics (size-to-depth ratio) of the real Digg dataset, which was found to align most closely with a high-homophily, high-virality setting.
The Role of Homophily and Virality
A significant finding of the study is how the "shape" of a cascade—whether it is "flat" (broad reach from few sources) or "tall" (deep chains of sharing)—is driven by the interaction between homophily and virality resistance. Interestingly, high homophily can actually limit cascade size; if a meme is only relevant to a specific "echo chamber," it may spread deeply within that group but fail to reach the broader network.
Tunability and Consistency
The NMF approach proved superior to standard graph embedding techniques like node2vec, providing a wider range of tunable homophily. This allows for the simulation of diverse environments, from random networks to extreme echo chambers. Furthermore, experiments showed a clear linear relationship: the more similar two memes are in topic space, the more similar the sets of users who activate on them.
Limitations and Future Directions
While WoMG provides a robust tool for "in vitro" information diffusion studies, the authors note several areas for future work:
- Dynamic Networks: The current model assumes a fixed social graph, whereas real networks evolve as users follow or unfollow based on content.
- Opinion Dynamics: Future iterations could track how a user's underlying opinions change after exposure to cascades, rather than just tracking activations.
- Veracity Modeling: A promising direction involves modeling the distinct spreading patterns of "fake" versus "genuine" news, potentially requiring new parameters for user skepticism or automated bot behavior.
Summary updated