Abstract
User behaviors in social networks are microscopic with fine-grained temporal information. Predicting a macroscopic quantity based on users’ collective behaviors is an important problem. However, existing methods often struggle to bridge the gap between individual micro-level dynamics and aggregate macro-level outcomes. This paper introduces MIC2MAC, a unifying framework that bridges this gap using Jump Stochastic Differential Equations (SDEs). By accounting for the discrete nature of individual actions, MIC2MAC provides a mathematically rigorous way to calculate the probability distribution of future outcomes, outperforming existing models in both accuracy and computational efficiency.
Introduction
In our hyper-connected world, we are inundated with "micro" data: timestamps of every retweet, song play, and product purchase. A central challenge for researchers and businesses is using this data to make "macro" predictions, such as forecasting the total number of viewers for a show or determining when a meme will reach a specific sharing threshold.
Current approaches generally fall into two categories. One group fits complex models to individuals and relies on slow, often inaccurate "brute force" simulations. The other group models the population as a whole, losing the rich detail inherent in micro-data. MIC2MAC serves as the "missing link," mathematically connecting these two scales in a principled, general-purpose framework.
Methods
The MIC2MAC framework operates in three stages to link individual events to population-level trends:
- Micro Modeling: Individual behaviors are modeled using Temporal Point Processes (e.g., Hawkes processes for viral events). Parameters are learned efficiently from historical data via convex optimization.
- Jump SDE Linkage: The core innovation is the formulation of a Jump Stochastic Differential Equation. Unlike standard continuous SDEs, Jump SDEs account for discrete "jumps"—where each individual action causes a specific increase in the macro quantity.
- Macro Inference: Instead of random simulations, the framework derives a differential equation describing the evolution of the macro quantity's probability distribution. This is solved numerically using a scalable Runge-Kutta algorithm, enabling rapid calculation of expected values and event timings.
Key Results and Insights
The framework was evaluated on diverse real-world datasets, including IPTV watching history, Reddit discussions, and MemeTracker hyperlink cascades.
- Superior Accuracy: MIC2MAC consistently outperformed state-of-the-art baselines. In popularity prediction, it achieved a 10% accuracy improvement over SEISMIC. For time-based tasks, it was approximately 2x more accurate than leading macroscopic competitors.
- Predictive Stability: As shown in Figure 2, the model's Mean Absolute Percentage Error (MAPE) grows more slowly than competitors as the prediction window extends, demonstrating its robustness for long-term forecasting.
- Stochasticity as a Feature: Unlike models that discard "randomness" for mathematical simplicity, MIC2MAC demonstrates that retaining the stochastic nature of jumps leads to significantly better predictions.
Conclusion
MIC2MAC provides a unified mathematical framework for popularity and influence prediction. By modeling the full probability distribution, it can predict not just "how many" events will occur, but also "when" specific thresholds will be met. While current limitations include deterministic approximations for intensities and the assumption of static networks, the framework opens new avenues for research into intertwined micro-macro dynamics.
Summary updated