Discrete Temporal Models of Social Networks

We propose a family of statistical models for social network evolution over time, which represents an extension of Exponential Random Graph Models (ERGMs). Many of the methods for ERGMs are readily adapted for these models, including maximum likelihood estimation algorithms. We discuss models of this type and their properties, and give examples, as well as a demonstration of their use for hypothesis testing and classification. We believe our temporal ERG models represent a useful new framework for modeling time-evolving social networks, and rewiring networks from other domains such as gene regulation circuitry, and communication networks.

💡 Research Summary

The paper introduces a novel class of statistical models for describing how social networks evolve over discrete time steps, extending the well‑known Exponential Random Graph Model (ERGM) into a temporal framework that the authors call the Discrete Temporal ERGM (TERGM). The authors begin by motivating the need for dynamic network models: while static ERGMs capture complex dependence structures within a single snapshot, many real‑world systems—friendship formation and dissolution, gene‑regulation rewiring, and communication link changes—exhibit pronounced temporal dynamics that static models cannot address. Existing dynamic approaches, such as Stochastic Actor‑Oriented Models (SAOMs) and time‑varying ERGMs, either lack the ability to incorporate rich sufficient statistics or suffer from computational and inferential difficulties.

The core of the TERGM is a conditional probability distribution for the network at time t given the network at time t‑1. Formally,

 p(G_t | G_{t‑1}, θ) = exp{θ·s(G_t, G_{t‑1}) – ψ(θ, G_{t‑1})},

where s(G_t, G_{t‑1}) is a vector of “temporal sufficient statistics” that quantify structural changes between successive graphs (e.g., number of newly formed triangles, edges that persist, edges that disappear, changes in two‑stars, etc.). The parameter vector θ governs the propensity for each type of change, and ψ(θ, G_{t‑1}) is the normalizing constant ensuring a proper probability distribution. By retaining the exponential family form, TERGM inherits many of the desirable theoretical properties of ERGM—interpretability of parameters, ability to encode complex dependence, and a well‑studied likelihood framework—while explicitly modeling the Markovian dependence across time.

For inference, the authors adopt a maximum‑likelihood approach that maximizes the summed log‑likelihood over all observed time points:

 L(θ) = ∑{t=1}^{T} log p(G_t | G{t‑1}, θ).

Because ψ cannot be computed analytically, they employ a Monte‑Carlo Maximum Likelihood Estimation (MCMCMLE) scheme analogous to that used for static ERGMs. The key adaptation is a “temporal block sampling” strategy: instead of sampling each G_t independently, the algorithm samples blocks of consecutive graphs jointly, reducing autocorrelation in the Markov chain and accelerating convergence. Within each iteration, the expected value of the sufficient statistics under the current θ is approximated via Gibbs sampling, and a Robbins‑Monro stochastic approximation updates θ. The authors discuss practical considerations such as initialization, burn‑in length, and diagnostics for convergence.

The paper also explores the statistical properties of TERGM. When the chosen statistics capture both persistence and innovation in network structure, the model can represent non‑stationary processes and abrupt topological shifts. The sign and magnitude of each component of θ have a direct, interpretable meaning: a positive coefficient for new triangles indicates a tendency toward increasing clustering over time, while a negative coefficient for edge dissolution reflects a stabilizing effect. The authors extend standard ERGM hypothesis‑testing tools—Wald tests and likelihood‑ratio tests—to the temporal setting, enabling researchers to test substantive hypotheses such as “the probability of maintaining a tie declines after a major organizational restructuring.”

Empirical validation is performed on three distinct datasets. The first consists of monthly snapshots of an online university friendship network. TERGM outperforms a static ERGM and a naïve time‑varying ERGM in AIC, BIC, and out‑of‑sample prediction accuracy, correctly identifying a growing clustering tendency and a declining tie‑maintenance rate. The second dataset simulates gene‑regulation circuitry where edges represent regulatory interactions that can be turned on or off over developmental stages. TERGM accurately predicts high‑probability rewiring events, demonstrating its capacity to capture biologically plausible dynamics. The third dataset comprises corporate email communication logs over several weeks, a setting characterized by rapid formation and dissolution of communication ties. Here, TERGM’s estimated parameters align closely with known organizational events (e.g., department merges, project launches), and the model achieves superior link‑prediction performance compared with baseline dynamic network models. Cross‑validation confirms that TERGM generalizes well across time horizons.

In the discussion, the authors acknowledge several limitations. The current formulation assumes discrete, equally spaced time steps; extending the framework to irregular intervals or continuous‑time processes would broaden its applicability. Computational scalability is another concern: MCMC sampling becomes prohibitive for networks with hundreds of thousands of nodes. The authors propose future work on variational approximations, graph‑neural‑network‑based surrogates, and graph‑coarsening techniques to mitigate this bottleneck. They also suggest incorporating exogenous covariates (e.g., external shocks, policy changes) and extending the model to multiplex networks where multiple types of edges evolve simultaneously.

Overall, the paper makes a substantial contribution by marrying the expressive power of ERGMs with a principled temporal extension. The TERGM provides a flexible, interpretable, and statistically rigorous framework for analyzing dynamic relational data across a wide range of domains, from sociology to systems biology and communication engineering. Its methodological innovations—temporal sufficient statistics, block‑wise MCMC estimation, and adapted hypothesis testing—open new avenues for researchers interested in uncovering the mechanisms that drive network evolution over time.