Source-linked AI summary
Latent Space Models for Dynamic Networks
Daniel K. Sewell, Yuguo Chen
TL;DR
The paper addresses how to model and analyze longitudinal network relationships, including prediction and incomplete observations. It embeds actors as trajectories in a latent Euclidean space and estimates the model using Bayesian MCMC with a computational approximation. The model provides rich dynamic-network visualization and supports directed edges, missing data, future-edge prediction, and edge-attraction analysis, illustrated on classroom and congressional cosponsorship networks.
Problem
Predicting future links is an important problem, while longitudinal network data may contain missing edge values.
Method
The model embeds dynamic directed or undirected network data into latent Euclidean-space trajectories and estimates parameters and positions with Bayesian MCMC.
Results
The model provides rich visualization of network dynamics and supports directed edges, missing data, future-edge prediction, and edge-attraction detection, with applications to classroom and House cosponsorship data.
Takeaways & Limitations
The approach gives researchers insight into actor characteristics, network evolution, and groupings or communities in dynamic networks.
Takeaways & Limitations
Non-ignorable missing data, where the missingness of edges themselves is informative, is beyond the paper's scope.
Abstract
from arXiv · showhide
Dynamic networks are used in a variety of fields to represent the structure and evolution of the relationships between entities. We present a model which embeds longitudinal network data as trajectories in a latent Euclidean space. A Markov chain Monte Carlo algorithm is proposed to estimate the model parameters and latent positions of the actors in the network. The model yields meaningful visualization of dynamic networks, giving the researcher insight into the evolution and the structure, both local and global, of the network. The model handles directed or undirected edges, easily handles missing edges, and lends itself well to predicting future edges. Further, a novel approach is given to detect and visualize an attracting influence between actors using only the edge information. We use the case-control likelihood approximation to speed up the estimation algorithm, modifying it slightly to account for missing data. We apply the latent space model to data collected from a Dutch classroom, and a cosponsorship network collected on members of the U.S. House of Representatives, illustrating the usefulness of the model by making insights into the networks.
1 INTRODUCTION
The paper extends latent space modeling from static networks to longitudinal network data by representing actors as trajectories in a latent Euclidean space. Its Bayesian approach supports network evolution analysis, missing edges, prediction, and edge-attraction analysis.
- Motivation: Dynamic network analysis studies how relationships and structures evolve across time in many application areas.Examples include friendships, political loyalties, co-authorship, epidemiology, and terrorist networks.
- Related approaches: Existing dynamic network models include temporal ERGMs, stochastic actor-oriented models, and latent space approaches.The first two use common network structures or user-defined objective functions, whereas latent space models embed network information in a usually low-dimensional space.
- Latent space modeling: Latent space models represent local and global network structure while providing meaningful, interpretable visualizations.The approach also inherently incorporates transitivity and supports qualitative inference.
- Research gap: Prior latent-space work was largely static, and the most related dynamic method could analyze only undirected networks.Other temporal latent-space studies addressed change points or latent features, while Sarkar and Moore extended Euclidean latent spaces to dynamic networks.
- Contribution: The proposed model embeds dynamic directed or undirected networks as actor trajectories in latent Euclidean space and estimates parameters with Bayesian MCMC.It is designed to support missing-data handling, future-edge prediction, visualization, and detection of edge attraction.
- Scalability and applications: A computational approximation reduces the cost of MCMC estimation for large networks.The paper then demonstrates the model using Dutch classroom data and U.S. House cosponsorship data.
2 DYNAMIC LATENT SPACE MODEL
The model represents longitudinal networks as actor trajectories in a latent Euclidean space, with edge probabilities shaped by latent distance and actor-specific social reach. Its parameterization incorporates sender activity, receiver popularity, and temporal evolution.
- Edge formation: The model makes edges more likely when actors are closer in latent Euclidean space.The latent space can represent actor similarity or social relationships through inter-actor distance.
- Directed edges: β_IN and β_OUT determine whether receiver popularity or sender activity contributes more strongly to directed edge formation.β_IN > β_OUT makes receiver identity more important, while β_OUT > β_IN makes sender identity more important.
- Social reach: Actor-specific radii r_i represent social reach, influencing each actor’s tendency to send and receive edges.A larger radius implies a greater propensity to form and receive ties.
- Model representation: Longitudinal network data are represented as trajectories of actors’ latent positions in a low-dimensional Euclidean space.The latent positions form a state-space representation of a time series of graphs.
- Model specification: The social-reach parameterization links individual effects to latent distance rather than treating them as constant additive effects.The paper states that this differs from the alternative parameterization rather than forming a subset of it.
- Social reach: If d_ijt < min(r_i, r_j), edge probability exceeds 1/2; if d_ijt > max(r_i, r_j), it is below 1/2.When d_ijt = r_i = r_j, the edge probability equals 1/2.
3 ESTIMATION
The paper estimates latent trajectories and model parameters with Bayesian posterior sampling using Metropolis-Hastings within Gibbs. It also addresses computational scalability through a case-control likelihood approximation and reorients sampled trajectories for identifiability.
- Posterior sampling: Bayesian inference targets the posterior distribution of latent trajectories and model parameters.The posterior is sampled with a Metropolis-Hastings within Gibbs algorithm.
- Prior specification: Inverse gamma priors are chosen for conjugacy, while the Dirichlet prior accommodates the constrained radii.The radii are sampled jointly with a Dirichlet proposal because of their constraint.
- Posterior sampling: The algorithm updates latent positions, variance parameters, β_IN, β_OUT, and r_1:n iteratively.Latent positions and regression-like parameters use Metropolis-Hastings proposals, while variance parameters use inverse gamma full conditionals.
- Identifiability: Rotations, reflections, and translations leave the posterior invariant, so sampled trajectories require Procrustes reorientation.The transformation preserves distances between actors across time points.
- Scalability: Case-control likelihood approximation reduces static-network log-likelihood cost from O(n^2) to O(n).The approximation replaces a quadratic summation with Monte Carlo sampling under a sparsity assumption.
- Scalability: The approximation is modified so it can be applied to longitudinal networks with missing edge values without knowing all links beforehand.The modification decomposes the log likelihood differently to avoid requiring every y_ijt in advance.
4 MISSING DATA
The paper treats missing network edges within a Bayesian missing-data framework and incorporates them into posterior sampling. Its discussion assumes missingness is MCAR or MAR, excluding non-ignorable missingness.
- Missing-data setup: Missing edges are represented separately from observed edges using the sampling pattern D.Complete data comprise (Y(obs), Y(mis), D), while observed data comprise (Y(obs), D).
- Missingness assumptions: The paper distinguishes missing completely at random (MCAR), missing at random (MAR), and non-ignorable missingness.MCAR and MAR are defined by whether the sampling pattern depends on observed or unobserved edges.
- Missingness assumptions: The analysis assumes missing edges are either MCAR or MAR because non-ignorable missingness is beyond the paper’s scope.Non-ignorable missingness occurs when the missingness pattern depends on the unobserved edges themselves.
- Ignorable sampling: Under ignorable sampling, inference can use the posterior conditional on observed edges without conditioning on D.The stated conditions include prior independence of sampling-pattern parameters and a product parameter space.
- Posterior sampling with missing data: Within MH-within-Gibbs sampling, each missing edge receives an additional Bernoulli draw based on the model’s edge probability.The full conditionals for latent positions and parameters remain unchanged when current missing-edge values are included.
5 PREDICTION
The paper predicts both future network edges and latent positions, using posterior information rather than simple plug-in parameter estimates. The weighted prediction approach outperforms the simpler plug-in method by reducing unnecessary parameter uncertainty.
- The prediction task jointly targets the future adjacency matrix YT+1 and latent positions XT+1.
- The model estimates future latent positions as posterior expectations conditioned on the observed network history.
- Future edge probabilities are estimated by averaging over posterior draws rather than conditioning only on posterior mean parameters.
- The method approximates the joint future distribution of an edge and the corresponding future latent positions, then marginalizes over latent-position uncertainty.
- The posterior-weighted method outperforms the simpler plug-in method because it uses fewer estimated parameters and introduces less prediction uncertainty.
6 EDGE ATTRACTION
The paper introduces edge attraction as an influence of one actor on another’s network-position movement, modeled through an extension of the latent-space transition equation. Posterior probabilities and trajectory geometry are used to detect and visualize this influence from edge information.
- Edge attraction is defined as one actor’s attracting influence on another actor’s edges, distinct from influence on friendships or behaviors.
- The transition model represents influence as an increased tendency for actor i to move through latent space toward actor j.
- The extended model retains the Markov property and permits inference using the same MCMC output obtained under the random-walk transition model.
- No edge attraction corresponds to µ = 0, whereas attraction is represented by µ > 0 and evaluated using posterior probability for the zero-attraction case.
- Large trajectory projections toward actor j produce smaller posterior probabilities that µ = 0, supporting detection of influence.
- Meaningful edge attraction requires the influencing actor to bring the influenced actor within at least one actor’s social circle.
- Influence visualization uses von Mises angle distributions around actors, with darker regions indicating higher directional probability.
- Plots can become overcrowded when multiple actors influence the same actor, motivating an averaged summary distribution.
7 SIMULATIONS
Simulation studies evaluate parameter recovery, model fit, latent-position recovery, edge-attraction detection, prior sensitivity, and computational approximation. The model recovers simulated structure well, detects attraction accurately, and substantially reduces computation with little fit loss.
- The mean (sd) estimates were 0.9172 (0.06207) for βIN, 2.045 (0.1438) for βOUT, and 0.9298 (0.06402) correlation for radii.
- The model’s AUC values were extremely high and uniformly exceeded Sarkar and Moore’s method for both directed and undirected networks.
- Estimated pairwise latent-distance ratios were narrow and centered near 1, indicating posterior-mean positions close to the simulated truth.
- Mean (sd) sensitivity and specificity were 0.952 (0.0316) and 0.832 (0.129) with attraction, while specificity without attraction was 0.868 (0.116).
- Sensitivity analyses produced AUC values from 0.9407 to 0.9858, averaging 0.9621, indicating robustness to σ2 prior hyperparameters.
- The approximation reduced mean computation time by 68.6% (0.835) while changing mean AUC from 0.9679 (0.0109) to 0.9618 (0.0048).
8 REAL DATA ANALYSIS
Applications to a Dutch classroom and a U.S. House cosponsorship network show that latent trajectories reveal demographic, ideological, predictive, and influence-related structure. The method also handles missing edges and outperforms the comparison method in reported prediction settings.
- Dutch classroom data: The Dutch classroom data contain four asymmetric friendship matrices, with missing edges arising when students were absent from surveys.
- Dutch classroom data: The Dutch classroom comparison yielded AUC 0.917 for the proposed method versus 0.8456 for Sarkar and Moore’s method.
- Dutch classroom data: In the classroom latent space, same-gender friendships were more prevalent, and the two non-Dutch students were close together.
- Dutch classroom data: The highest-ranked student moved toward the social-space center, whereas the lowest-ranked student moved away from it.
- Dutch classroom data: Edge attraction was detected for four classroom actors; actor 25 was influenced by nine others, with attraction waning as he formed his own social circle.
- Cosponsorship data: In the cosponsorship data, 30.1% to 31.1% of members were unrepresented at each time point, producing 51.2% to 52.5% missing edges.
- Cosponsorship data: Popularity was more important than activity level in cosponsorship edge formation, with βIN = 0.974 and βOUT = 0.147.
- Cosponsorship data: Republicans and Democrats occupied different halves of latent space, while denser central regions corresponded to more moderate ideologies.
9 DISCUSSION
The latent space model represents dynamic networks in ways that reveal network evolution, actor characteristics, and global and local structure. It also supports directed edges, missing data, future prediction, edge-attraction visualization, computational approximation, and extensions beyond binary relations.
- The model visualizes dynamic networks to reveal actor characteristics, network evolution, and overall groupings or communities.
- It handles directed edges and missing data while supporting prediction of future latent positions and edges.
- The model detects and visualizes edge attraction using network edge information.
- An approximation method produces statistically meaningful estimates in a computationally efficient way.
- Although the paper analyzes directed graphs, the methods can also model undirected graphs.
- The binary-edge formulation can be generalized to other dyadic data by changing the link function, such as using a log link for counts.