Temporal Exponential-Family Random Graph Models

• TERGMs are statistical models for dynamic networks that extend ERGMs with a discrete-time Markov process to capture temporal dependencies.
• They incorporate key network features such as tie formation, dissolution, stability, triadic effects, and latent community structures.
• Inference methods like MCMC-MLE and MPLE address computational challenges, enabling practical analysis of complex, time-varying relational data.

Temporal Exponential-Family Random Graph Models (TERGMs) are a class of statistical models that generalize the exponential-family random graph model (ERGM) framework to dynamic networks—that is, sequences of networks observed over time. TERGMs provide a principled basis for modeling the probabilistic evolution of network topology, permitting explicit control over endogenous network structures and temporal dependencies such as dyadic persistence, formation, and dissolution. They have become a cornerstone for modern longitudinal network analysis, facilitating both inference and prediction in domains characterized by complex, time-varying relational data.

1. Formal Definition and Core Structure

A Temporal ERGM (TERGM) models a discrete-time Markov process $\{Y_t\}_{t=1}^T$, where $Y_t$ is the network at time $t$ (typically represented as an adjacency matrix on a fixed set of $n$ nodes). The one-step conditional transition law is of exponential-family form: $$P(Y_t = y \mid Y_{t-1} = y', \theta) = \exp\left\{\theta^\top g(y, y') - \psi(\theta, y')\right\},$$ where

• $\theta \in \mathbb{R}^p$ is the vector of model parameters,
• $g(y, y') \in \mathbb{R}^p$ is a vector of sufficient statistics, which may depend on both present and lagged network states,
• $$\psi(\theta, y') = \log \sum_{y^* \in \mathcal{Y}} \exp\left\{\theta^\top g(y^*, y')\right\}$$ is the log-partition function (normalizer) ensuring proper probabilistic mass.

The joint likelihood is, by the first-order Markov property,

$$P(Y_2, \ldots, Y_T \mid Y_1) = \prod_{t=2}^T P(Y_t \mid Y_{t-1}, \theta).$$

Extensions accommodate exogenous covariates and higher-order Markov dependence, though the first-order case is standard in the literature (Lee et al., 2017, Fritz et al., 2019).

2. Sufficient Statistics and Model Components

The vector of sufficient statistics $g(y, y')$ encodes structural and temporal network features. Common choices include:

• Edge count (density): $Y_t$0 controls the baseline propensity for ties.
• Stability/persistence: $Y_t$1 rewards dyadic stability of both ties and non-ties.
• Formation/dissolution ("STERGM" notation):
  • Formation: $Y_t$2 counts newly formed edges.
  • Persistence: $Y_t$3 counts ties persisting from $Y_t$4 to $Y_t$5.
• Triadic and reciprocity effects: E.g., lagged transitivity or delayed reciprocity: $Y_t$6 (Lee et al., 2017, Fritz et al., 2019, Huang et al., 2022).

Additional user-specified statistics can capture complex dependencies, including motif closure, degree-distributions, or block structure. The flexibility to mix cross-sectional and temporal terms is a distinguishing advantage of TERGMs.

3. Extensions: Separable Models, Hierarchical Models, and Beyond

Several important subclasses and generalizations of TERGMs are prevalent:

• Separable TERGM (STERGM): Imposes a conditional independence structure between tie-formation and tie-dissolution processes within each timestep, factorizing the one-step transition law as a product of two ERGMs:

$Y_t$7

where $Y_t$8 (formation) and $Y_t$9 (dissolution). This structure provides separate control over tie incidence and duration (Krivitsky et al., 2010, Klumb et al., 2021).

• Mixture and Hierarchical TERGMs: Community structure and time-varying heterogeneity can be modeled through mixtures or block-wise TERGMs. In finite mixture TERGMs (used for dynamic community detection), nodes are assigned latent membership vectors, and dyads evolve according to community-specific TERGM parameters (Lee et al., 2017, Cao, 2017). Variational methods and tailored EM algorithms yield scalable inference.
• Partially/Non-separable models for valued networks: The PST ERGM partially decouples the increment and decrement processes for valued-edge dynamics, allowing for separate sufficient statistics for dyad value increases and decreases, but not full independence as in STERGM (Kei et al., 2022).
• Varying-coefficient TERGMs: These models relax time-homogeneity by allowing coefficients to vary smoothly over time via basis-spline expansion $t$0, estimated by penalized pseudo-likelihood (Lee et al., 2017).
• Triadic and social-learning extensions: TTERGMs add explicit triadic (transitivity, triangle formation) and social-learning (influencer/follower) terms, showing improved prediction in real-world temporal social networks (Huang et al., 2022).

4. Inference, Estimation Algorithms, and Scalability

Inference in TERGMs is nontrivial due to the intractability of the normalizing constants for high-dimensional network spaces. Established strategies include:

• MCMC-MLE: Stochastic approximation methods using Markov chain Monte Carlo to estimate expected sufficient statistics under the model, iterating updates for $t$1 until convergence (Leifeld et al., 2015, Huang et al., 2022).
• Maximum pseudo-likelihood estimation (MPLE): Approximates the likelihood by a product of conditional binary logistic regressions for dyads; computationally efficient and consistent under certain regimes, but often requires bootstrap correction for standard errors (Fritz et al., 2019, Cao, 2017).
• Variational EM and Minorization-Maximization (MM): Particularly for mixture models (community detection), variational methods approximate the intractable posterior over latent node memberships, and MM constructs tractable surrogates for nonconcave optimization in the variational parameters (Lee et al., 2017).
• Block or clusterwise models: Hierarchical approaches (e.g., THERGM) partition the network and fit cluster-specific TERGMs; scalable with parallelization and pseudo-likelihood (Cao, 2017).

Restriction to dyad-wise independent statistics allows linear-time parameter updates and enables models to scale to hundreds of nodes and dozens of time points (Lee et al., 2017). For STERGM, closed-form or dyad-wise approximations exist in the sparse or dyad-independent limit (Klumb et al., 2021).

5. Continuous-Time Limits and Theoretical Properties

A major development has been the precise connection between discrete-time TERGMs and continuous-time Markov processes on graph space:

• Continuous-time Markov chain (CTMC) limits: The CTMC formulation yields a process whose infinitesimal generator $t$2 induces transition rates $t$3 for graph state changes. Construction ensures detailed balance and a stationary distribution coinciding with a cross-sectional ERGM, under suitable rate parameterization (Butts, 2022).
• Separable temporal models in continuous time (CSTERGM): The CSTERGM captures constant-rate formation/dissolution mechanisms and recovers the exact ERGM equilibrium with prescribed edge duration under mild assumptions. The continuous-time limit validates various STERGM approximations and provides a mechanistic interpretation for temporal tie dynamics (Klumb et al., 2021, Butts, 2022).
• Implications: The equilibrium (stationary) distribution constrains long-run network law; cross-sectional data alone cannot distinguish between dynamic processes with the same stationary law. Dwell-time, event interval, and tie duration distributions are dynamic diagnostics.

6. Applications, Empirical Performance, and Extensions

TERGMs have been applied extensively in sociology, economics, neuroscience, and computational biology. Examples include:

• International arms transfers: Edge-propensity and inertia strongly regulate temporal dynamics; STERGM decomposes exogenous (e.g., GDP) and endogenous (degree, transitivity) effects on formation vs. persistence (Fritz et al., 2019).
• Email communication networks: Repetition, reciprocity, and triadic closure all show distinct behavior in tie formation and dissolution phases, which is revealed by STERGM (Fritz et al., 2019).
• Brain connectivity and legislative co-voting: Varying-coefficient TERGMs surface long-term heterogeneity and regime shifts in structural parameters (Lee et al., 2017).
• Dynamic valued networks: PST ERGMs accommodate count-valued edge trajectories, enabling parameter interpretation for edge increment/decrement forces (Kei et al., 2022).
• Predictive accuracy: Augmented models (TTERGM) that incorporate triadic and learning terms outperform standard TERGM and block models in longitudinal out-of-sample prediction (Huang et al., 2022).

The workflow typically involves exploratory choice of sufficient statistics, model selection (e.g., conditional-likelihood BIC for mixture models), fitting by MCMC-MLE or MPLE, simulation-based goodness-of-fit diagnostics, and predictive evaluation via ROC/AUC or explicit link-forecasting (Lee et al., 2017, Huang et al., 2022).

7. Limitations, Open Challenges, and Future Directions

TERGMs exhibit certain limitations:

• Computational cost: Full MCMC likelihood estimation is slow for large, richly dependent networks; scalable methods favor dyad-wise independence or maximum pseudo-likelihood at the expense of inferential accuracy (Klumb et al., 2021).
• Model degeneracy: As in static ERGMs, overparametrized dynamic models can suffer from degeneracy, placing nearly all probability on trivial (empty or complete) networks; remedy often involves smoothing high-order effects or careful term selection (Fritz et al., 2019).
• Interpretability: While memory statistics are flexible, improper specification may confound persistence with lack of new ties; separable modeling (as in STERGM or PST ERGM) partially resolves this ambiguity (Krivitsky et al., 2010, Kei et al., 2022).
• Inference from network samples: Most formulations assume full observation of each snapshot; extensions to partially observed, sampled, or egocentrically reported data rest on adaptation of ERGM methods (Butts, 2022).
• Continuous-time and micro-level process ambiguity: TERGMs do not specify the underlying micro-dynamics within inter-observation intervals. Richer actor-oriented and event-based models provide an alternative but introduce additional assumptions (Leifeld et al., 2015, Butts, 2022).

Recent progress focuses on the development of hybrid models, tighter connections between discrete- and continuous-time frameworks, and scalable approximations for dense and high-dimensional temporal network data (Butts, 2022, Klumb et al., 2021, Huang et al., 2022).

---

Key References:

• Model-Based Clustering of Time-Evolving Networks through Temporal Exponential-Family Random Graph Models (Lee et al., 2017)
• A Separable Model for Dynamic Networks (Krivitsky et al., 2010)
• Approximations for STERGMs Based on Cross-Sectional Data (Klumb et al., 2021)
• Varying-coefficient models for dynamic networks (Lee et al., 2017)
• Triadic Temporal Exponential Random Graph Models (TTERGM) (Huang et al., 2022)
• Continuous Time Graph Processes with Known ERGM Equilibria (Butts, 2022)