Abstract
Understanding the complex interactions within dynamic multilayer networks is critical for advancements in various scientific domains. Existing models often fail to capture such networks' temporal and cross-layer dynamics. This paper introduces a novel Tensor State Space Model for Dynamic Multilayer Networks (TSSDMN), utilizing a latent space model framework. TSSDMN employs a symmetric Tucker decomposition to represent latent node features, their interaction patterns, and layer transitions. Then by fixing the latent features and allowing the interaction patterns to evolve over time, TSSDMN uniquely captures both the temporal dynamics within layers and across different layers. The model identifiability conditions are discussed. By treating latent features as variables whose posterior distributions are approximated using a mean-field variational inference approach, a variational Expectation Maximization algorithm is developed for efficient model inference. Numerical simulations and case studies demonstrate the efficacy of TSSDMN for understanding dynamic multilayer networks.
Introduction
In our increasingly interconnected world, data frequently manifests as multilayer networks—complex systems where a single set of entities (nodes) interact through multiple distinct types of relationships (layers). For instance, in international relations, countries interact through trade, diplomatic cooperation, and military conflict. These networks are inherently dynamic, evolving over time as relationships shift and events unfold.
Modeling these systems presents two significant challenges. First, capturing cross-layer dynamics is difficult: how does a diplomatic incident in one layer influence future trade relations in another? Second, models must balance stability and change. While many existing frameworks assume nodes move through a latent space, in reality, intrinsic node characteristics (such as a country's geography or a person's core personality) are often stable, while their behavioral interaction patterns fluctuate. The Tensor State Space Model for Dynamic Multilayer Networks (TSSDMN) addresses these gaps by separating stable node traits from time-varying interaction behaviors.
The TSSDMN Framework
The core of the TSSDMN approach is a symmetric Tucker decomposition applied to the network's edge probabilities. This framework allows the model to decompose the complex tensor of interactions into interpretable components.
Symmetric Tucker Decomposition
The model represents the log-odds of a connection at time $t$ as a product of three primary components:
- Latent Node Factors ($\mathbf{C}_1$): A nonnegative matrix representing intrinsic node characteristics, shared across all layers and time points.
- Layer-Transition Matrix ($\mathbf{C}_2$): A matrix that captures the structural relationships and influences between different layers.
- Dynamic Core Tensor ($\mathcal{Z}_t$): A time-varying component that represents the specific interaction patterns between latent factors at time $t$.
By imposing nonnegativity constraints on the node factors, the model ensures identifiability—meaning the learned parameters are unique and meaningful—and allows these factors to naturally represent community membership weights.
State-Space Dynamics
Unlike models that move nodes through latent space, TSSDMN evolves the interaction patterns themselves. The core tensor $\mathcal{Z}_t$ follows a third-order tensor autoregressive model. This allows the framework to capture both "intra-layer" trends (historical patterns repeating within a single layer) and "cross-layer" trends (the history of one layer influencing the future of another).
Model Inference and Identifiability
To handle the high dimensionality and complexity of the tensor state-space, the authors developed a Variational Expectation-Maximization (VEM) algorithm. This approach utilizes a mean-field approximation to efficiently estimate the posterior distributions of the latent variables. Furthermore, the paper provides formal proofs regarding the model's identifiability, ensuring that the separation of static traits and dynamic behaviors is mathematically sound. From a theoretical perspective, the Bayesian estimator used in TSSDMN achieves the minimax optimal convergence rate, providing a rigorous foundation for its performance.
Experimental Evaluation
The efficacy of TSSDMN was evaluated across synthetic and real-world datasets:
- Synthetic Benchmarks: TSSDMN consistently outperformed baseline models (such as EDMN and DSBM) in terms of Mean Square Error (MSE) for latent probability estimation and Area Under the Curve (AUC) for edge prediction.
- Geopolitical Analysis (ICEWS): Using data on interactions between 30 countries (2009–2017), the model's dynamic core tensor successfully identified major historical shifts. Significant "jumps" in interaction patterns corresponded precisely with the onset of the Arab Spring in 2011 and the Crimea Crisis in 2014.
- Cybersecurity (UNSW-NB15): In modeling network traffic (Normal vs. various attack types), TSSDMN achieved near-perfect AUC scores (0.99+). It proved particularly adept at identifying malicious activity despite the sparse and intermittent nature of such events.
Conclusion and Future Work
TSSDMN offers a powerful paradigm shift by modeling behavioral evolution rather than positional shifts. By fixing nodes in a latent space and evolving their interaction styles, the model provides superior interpretability and predictive accuracy.
Future research directions include extending the model to handle inter-layer edges (direct connections between different nodes in different layers) and adapting the framework for weighted networks to capture interaction intensity. Additionally, optimizing the computational scalability of tensor operations will be essential for applying TSSDMN to massive, web-scale datasets.
Summary updated