Introduction
Predicting the popularity of information—such as the number of retweets for a post or citations for a scientific paper—is a critical task for social media management and recommendation systems. While existing models effectively capture spatiotemporal patterns from an initial observation period, they often struggle to account for the "invisible" future growth trends that occur after observation ends.
The authors introduce CasFT (Cascade Future Trend), a framework designed to model these underlying growth dynamics explicitly. By treating the future not as a black box but as a continuous evolution of momentum, CasFT achieves significant accuracy improvements over state-of-the-art methods.
Methodology
The CasFT framework operates through three primary stages: historical modeling, future trend interpretation, and final prediction.
- Historical Spatiotemporal Modeling: The model utilizes a dual-graph approach to capture both local structural patterns (via GraphWave) and global social network context (via NetSMF). A self-attention mechanism encodes the timing of past events to create a robust representation of the observed cascade.
- Interpretation of Future Trends: This is the core innovation of CasFT. It employs Neural Ordinary Differential Equations (ODEs) to model the growth rate of popularity as a continuous-time function, allowing for smooth extrapolation from the observed past. To handle the inherent uncertainty of future events, a Conditional Diffusion Model is used to generate a sequence of potential incremental popularity values.
- Prediction: The historical features and generated future trends are fused and processed through a Multi-Layer Perceptron (MLP) to produce the final popularity forecast.
Experiments and Results
CasFT was evaluated on three diverse datasets: Twitter (hashtag cascades), APS (long-term physics citations), and Sina Weibo (repost cascades).
Key Findings
- Performance Gains: CasFT consistently outperformed baselines like CasFlow and CTCP. Notably, on the Twitter dataset with a 1-day observation window, it improved the Mean Squared Logarithmic Error (MSLE) by 19.36% over the best baseline.
- Momentum as a Physical Property: The results validate that modeling the growth rate (via ODEs) rather than just raw counts allows the model to capture the "momentum" of information spread.
- Generative Advantage: By "imagining" plausible future trajectories through diffusion, the model gains a richer feature set than standard discriminative models that map inputs directly to a single output.
Conclusion
CasFT represents a significant step forward in popularity prediction by bridging the gap between observed data and invisible future trends. While the combination of Neural ODEs and Diffusion models increases computational complexity, the gains in accuracy and robustness across different platforms and timeframes demonstrate the power of modeling continuous growth dynamics. Future work may involve integrating external shocks, such as breaking news, to further refine the model's predictive capabilities.
Summary updated