Abstract
Social networks are inherently dynamic, with relationships and attributes evolving in a continuous, asynchronous manner. Traditional modeling approaches often rely on discrete-time snapshots, which fail to capture the underlying process of change. This paper proposes a framework that treats social network evolution as a Continuous-Time Bayesian Network (CTBN). By leveraging Monte Carlo Expectation Maximization (MCEM) and Markov Chain Monte Carlo (MCMC) sampling, the authors provide a scalable method for parameter estimation and the inference of hidden social structures from indirect event data.
Introduction
The study of social network dynamics is critical for understanding phenomena such as the spread of behaviors or the formation of professional ties. However, most existing research utilizes static snapshots or discrete-time models, which ignore the continuous nature of real-world interactions. Sociological "actor-oriented" models, while continuous, face significant computational hurdles: the state space grows exponentially with the number of actors, and traditional estimation methods like the Method of Moments (MoM) require complete network snapshots, which are rarely available in practice.
Methodology
The authors bridge the gap between sociology and graphical models by framing social dynamics as a CTBN. This approach factorizes the complex system into local variables, allowing for more efficient modeling of dependencies over time.
MCEM for Partially Observed Networks
For scenarios where network snapshots are available but potentially sparse or asynchronous, the authors employ an importance sampling-based EM procedure. This allows for parameter estimation by considering the full continuous trajectory between observations rather than just the endpoints, making it more robust than traditional discrete methods.
Hidden Social Network Dynamics
A significant contribution of this work is the model for entirely hidden networks. In many real-world cases, the underlying social structure is unobserved; only communication events (e.g., emails) are visible. The authors model these events as "toggle variables" dependent on the hidden state of the network. To navigate the massive search space of possible trajectories, an MCMC approach using Metropolis-Hastings sampling is implemented to guide simulations toward trajectories consistent with observed events.
Experimental Results
The proposed methods were validated across synthetic and real-world datasets, demonstrating significant improvements in both accuracy and applicability.
Synthetic Validation and Algorithm Comparison
On synthetic data, the EM algorithm demonstrated superior robustness compared to the standard MoM used in sociology. As shown in the performance plots below, the EM approach consistently achieved higher log-likelihoods, particularly as the time interval between observations ($\Delta t$) increased. This highlights its effectiveness in data-scarce environments where traditional methods fail to capture the dynamics between snapshots.
Real-World Applications
The model was applied to the "50 Girls" dataset, a longitudinal study of friendship and substance use. The CTBN framework successfully handled the ~2,500 variables involved, demonstrating scalability beyond the reach of exact inference algorithms.
Furthermore, the hidden model was tested on the Enron email dataset to infer organizational structure from timestamps alone. The algorithm identified stable connections and quantified communication frequencies, finding that reciprocated ties led to more frequent interactions (roughly every 2 days versus every 3 days for unreciprocated ties). The evolution of such network structures can be visualized through adjacency matrices at different time points.
Conclusion
By integrating sociological models with the machinery of Continuous-Time Bayesian Networks, this work provides a powerful toolkit for analyzing social dynamics. The sampling-based inference methods offer improved accuracy over traditional techniques and enable the discovery of hidden social structures from indirect data. Future research will focus on improving the computational efficiency of MCMC sampling to scale these models to even larger global networks.
Summary updated