Abstract
Predicting user engagement in information cascades is a critical task for social media analysis, yet existing models often struggle with the trade-off between interpretability and predictive flexibility. We present DANTE (Discriminative probabilistic ANytime user Engagement prediction), a novel framework that bridges this gap. By combining split-population survival analysis with a discriminative learning objective, DANTE enables accurate, interpretable, and "anytime" predictions for arbitrary observation and forecast windows without retraining. Our evaluation across four real-world datasets demonstrates that DANTE achieves state-of-the-art performance while providing principled uncertainty quantification.
Introduction
In social media ecosystems, understanding the dynamics of information cascades—how content spreads from user to user—is essential for applications ranging from viral marketing to the mitigation of misinformation. Predicting future engagement levels is inherently challenging due to the temporal nature of diffusion.
Existing approaches typically fall into two categories: generative models based on temporal point processes, which offer interpretability but often lack predictive accuracy, and discriminative deep learning models, which are accurate but function as "black boxes." A significant limitation of current discriminative models is their lack of temporal flexibility; a model trained for a specific observation window (e.g., 3 hours) cannot be easily adapted to a different window (e.g., 30 minutes) without extensive retraining. DANTE addresses this "anytime" challenge by providing a single model capable of handling arbitrary observation periods ($t_c$) and forecast horizons ($\Delta t$).
The DANTE Framework
DANTE's architecture is founded on two primary innovations:
- Split-Population Survival Analysis: Unlike standard survival models that assume all individuals will eventually engage, DANTE incorporates a "split-population" approach. It explicitly models user susceptibility ($\pi$) using a logistic function of features and the hazard rate (the instantaneous likelihood of engagement) using influence parameters and memory kernels (e.g., Power-law or Weibull). This allows the model to account for the fact that many users may never interact with a specific cascade.
- Discriminative Objective Function: The authors derive a novel objective function that serves as an upper bound to the generative likelihood. This objective emphasizes predicting whether an event occurs within a specific window, allowing the model to function like a classifier while maintaining the probabilistic rigor of a point process.
To ensure scalability, DANTE utilizes a Consensus ADMM (Alternating Direction Method of Multipliers) optimization procedure, enabling parallelized training across large-scale networks.
Experimental Evaluation
The model was rigorously tested on four diverse real-world datasets: Irvine (student social network), LastFM (music listening histories), Digg (news story votes), and Memes (phrase/link propagation). For datasets with explicit social graphs, graph embeddings were leveraged to enhance the model's understanding of user susceptibility.
Key Results and Insights
As shown in the experimental results, DANTE demonstrates significant advantages over both traditional and deep learning baselines:
- Anytime Flexibility: DANTE's ability to learn underlying hazard and susceptibility functions allows it to be applied to any $t_c$ and $\Delta t$ at test time. This makes it uniquely suited for real-time monitoring where data availability windows are dynamic.
- State-of-the-Art Accuracy: Using the Squared Log Error (SLE) metric, DANTE consistently matches or outperforms state-of-the-art models like CasFlow. Notably, while CasFlow requires separate models for each time configuration, DANTE achieves these results with a single, unified model.
- Interpretability and Uncertainty: Unlike deep learning "black boxes," DANTE's parameters represent explicit causal influence strengths between users. Furthermore, it produces a full Probability Mass Function (PMF), enabling the generation of 95% confidence intervals—a critical feature for risk assessment in sensitive applications like rumor control.
Conclusion and Future Work
DANTE represents a significant step forward in anytime engagement prediction. While the current model assumes specific functional forms for memory kernels and relies on high-quality input features, future work will explore non-parametric hazard functions and optimizations for massive-scale social networks. By bridging the gap between generative interpretability and discriminative accuracy, DANTE provides a robust tool for understanding and predicting the complex dynamics of information diffusion.
Summary updated