Abstract
In social media and news ecosystems, the pathways through which information diffuses are often latent. Traditional models for network inference frequently assume uniform transmission rates, ignoring the critical role of content. This paper introduces TOPICCASCADE, a continuous-time model that reconstructs hidden social networks and predicts information spread by integrating both timing and content topics. By modeling transmission rates as a function of topic distributions, TOPICCASCADE significantly improves network recovery and event time prediction over existing state-of-the-art methods.
Introduction
Understanding how information spreads across digital platforms is vital for tracking trends, identifying influencers, and managing misinformation. While we can observe when a node (e.g., a blog or user) adopts a piece of information, the specific source or "path" is usually hidden. Previous research has attempted to infer these networks based solely on temporal data, often assuming a static spreading rate for all content.
However, information diffusion is inherently topic-sensitive. A technology-focused site might rapidly propagate gadget news but ignore or slowly report on political developments. TOPICCASCADE addresses this by simultaneously inferring hidden network structures and topic-specific transmission rates using observed timestamps and text.
Methodology
TOPICCASCADE utilizes survival analysis to model the "time-to-infection" between nodes. The core innovation lies in its topic-aware transmission model:
- Topic Modulation: Each information cascade (or "meme") is represented as a topic distribution using Latent Dirichlet Allocation (LDA). The transmission rate between two nodes is defined by the inner product of the meme’s topics and a learned "topic preference" vector specific to that edge.
- Rayleigh Distribution: The model employs the Rayleigh distribution to characterize transmission times. This choice captures the empirical observation that reposting probability typically peaks shortly after a source post before decaying.
- Convex Optimization: The learning objective is formulated as a convex optimization problem, ensuring that the algorithm converges to a global optimum.
- Regularization: To promote network sparsity—reflecting the reality that most nodes are not directly connected—the authors apply Group Lasso regularization. This effectively prunes irrelevant edges during the inference process.
Experimental Setup
The model was rigorously evaluated on both synthetic and real-world datasets:
- Synthetic Networks: Using Kronecker generators, the authors simulated three network topologies: Random (Erdős-Rényi), Core-Periphery, and Hierarchy.
- MemeTracker Dataset: This real-world dataset comprises millions of blog posts. The study focused on a sub-network of the top 500 sites, encompassing approximately 1,100 edges and 2,000 diffusion cascades.
Key Results
TOPICCASCADE demonstrates superior performance in both network structure recovery and temporal prediction:
- Network Recovery: In synthetic benchmarks, the model achieved significantly higher F1 scores than baselines like NETRATE and MoNET. With 10,000 cascades, it achieved nearly perfect reconstruction of the underlying network.
- Time Prediction Accuracy: On the MemeTracker data, TOPICCASCADE reduced the Mean Absolute Error (MAE) for transmission time prediction by nearly 50% compared to topic-agnostic models.
The results indicate that incorporating content context allows the model to better distinguish between likely and unlikely transmission paths, even when the number of observed cascades is limited.
Discussion and Future Work
TOPICCASCADE represents a significant step forward in making diffusion models context-aware. By treating the network as a dynamic structure "activated" by specific topics, it provides a more nuanced view of social influence.
Despite its strengths, the model currently assumes static network structures and topic preferences. Future iterations could incorporate temporal dynamics to account for evolving interests and relationship shifts. Additionally, addressing "content mutation"—where the topic distribution of a meme changes as it spreads—would further enhance the model's realism and predictive power.
Summary updated