Multilayer Time-Varying Graphs

• Multilayer time-varying graphs are mathematical structures that integrate multiplex network layers with dynamic temporal evolution, facilitating unified analysis of complex systems.
• The framework leverages MultiAspect Graph formalism to support sub-determination, embedding, and application-specific adaptations for domains like transportation and dynamic interactions.
• Advanced methodologies such as statistical latent space models, sheaf-theoretic approaches, and graph-level deep learning enable accurate anomaly detection and time-series forecasting.

A multilayer time-varying graph is a mathematical framework for modeling networked systems that possess both multilayer (multiplex) and temporal structure. In these graphs, nodes, layers, and time are formalized as distinct “aspects,” enabling unified analysis of structural and dynamical phenomena such as interlayer coupling and temporal evolution. The MultiAspect Graph (MAG) abstraction is foundational, supporting representation, algorithmic analysis, and embedding methodologies that apply to complex systems ranging from transportation networks to time-evolving interaction graphs. Recent work extends these concepts to include advanced statistical modeling, sheaf-theoretic approaches for temporal morphisms, and state-of-the-art embedding and learning paradigms.

1. Mathematical Formalism of Multilayer Time-Varying Graphs

Let $p \in \mathbb{N}$ denote the number of aspects. A MultiAspect Graph (MAG) of order $p$ is a pair $H = (A, E)$, where $A = \{A_1, ..., A_p\}$ and each $A_i$ is a finite set corresponding to an aspect (e.g., physical nodes, layers, time instants) (Wehmuth et al., 2014, Wehmuth et al., 2015). The composite vertex set is $V(H) = A_1 \times \cdots \times A_p$. Edge set $E \subseteq V(H) \times V(H)$ consists of ordered pairs $(u,v)$, each $u,v \in V(H)$, specifying directed connections. For multilayer temporal graphs ($p=3$ or $p > 3$), aspects typically include nodes, layers, time instants, and possibly additional categorical dimensions (e.g., modalities or dataset periods).

MAGs are isomorphic to directed graphs on $|V(H)|$ vertices, so every walk, path, or classical graph operator (adjacency, BFS, DFS, shortest paths) lifts directly to the multilayer time-varying setting. The adjacency structure may be housed as a high-order tensor or, after flattening, as a sparse matrix indexed by composite vertex tuples (Wehmuth et al., 2014, Wehmuth et al., 2015).

2. Construction and Structural Analysis

The definition and construction of multilayer time-varying graphs depend on the application domain. For air transportation, each composite vertex is $$(\text{airport},\,\text{airline},\,\text{time},\,\text{period})$$; edges encode flights from one airport to another by a specific airline at precise times and data periods (Costa et al., 2017). To analyze network structure, sub-determination is applied: projecting the full MAG onto selected aspects to focus on routes (aggregating over airlines and time), flight patterns (retaining airport and time), or layer-specific features.

Key structural metrics such as K-core decomposition generalize to any MAG sub-determination, quantifying, for instance, the backbone of flight connectivity either in terms of core airports or frequencies of simultaneous flights. The unified MAG framework permits one to seamlessly alternate between multilayer, temporal, and aggregate views as needed (Wehmuth et al., 2015, Costa et al., 2017).

3. Embedding and Learning on Multilayer Time-Varying Graphs

Graph-level representation learning for dynamic networks requires methods compatible with the joint multilayer and temporal structure. One approach constructs a multilayer graph $M = (V_M, E_M)$ where $V_M = \bigcup_{t\in T} \{(v, t): v \in V\}$, and $E_M$ consists of both intra-layer (structural at time $t$) and inter-layer (temporal backtracking, e.g., $(v, t) \to (v, t-1)$) edges (Wang et al., 2023).

A modified random walk process mixes spatial transitions (within a timestamp) with temporally backtracking jumps; transition probabilities are governed by the "stay-in-layer" hyperparameter $\alpha$, node2vec-style hyperparameters $p, q$, and alias-sampling for computational efficiency. Walk traces form "sentences," which are aggregated by snapshot and encoded via a document-level LLM (Doc2Vec), yielding a $d$-dimensional embedding per graph snapshot. These embeddings support downstream tasks such as temporal similarity ranking, isomorphism testing, and anomaly detection using metrics like cosine similarity, Precision@K, Spearman's $\rho$, and MRR (Wang et al., 2023).

For time-series graph forecasting, evolving multi-scale GNNs synthesize hierarchical/multi-scale graph construction with time-varying adjacency learning, segmenting temporal data to capture dynamic dependencies per layer/scale. Neural modules (MLP, GRU) are employed to update node-embeddings and adjacency matrices over time and scales; mix-hop propagation aggregates higher-order neighborhood information. Learned representations facilitate both single/multi-step prediction and visualization of dynamic structure (Ye et al., 2022).

4. Algorithmic and Computational Considerations

MAGs enable direct reuse of classical graph algorithms on high-dimensional (composite vertex) digraphs. Degree, BFS, and DFS algorithms are adapted by operating on flattened adjacency matrices and applying selection/projection matrices to extract aspect-specific or "sub-determined" slices. Avoidance of spurious subgraph structure during aspect aggregation is handled by reachability-limited search (Wehmuth et al., 2015).

Construction of multilayer time-varying graphs is scalable; adjacency and incidence structures are sparsified in practice. Complexity scales with the number of composite vertices and edges, but empirical results show sub-linear wall-time growth up to graphs with $10^6$ edges, due to parallelization during walk generation and asynchronous stochastic optimization in embedding models (Wang et al., 2023).

5. Statistical Models and Dynamical Frameworks

Advanced latent space models formalize the joint evolution of multilayer and temporal structure by treating observed graphs $\{Y^k_t\}$ as generated by both shared ($X_t$) and layer-specific ($\Lambda_k$, $\delta^i_{k,t}$) latent variables (Loyal et al., 2021). Statistical inference leverages Pólya-Gamma augmentation and structured mean-field variational methods, yielding scalable closed-form updates for node trajectories, degree effects, and homophily while enabling interpretation of shared vs layer-specific dynamics.

Recent categorical and sheaf-theoretic frameworks extend multilayer time-varying graphs to functorial structures ("narratives"), where multiplex graphs are modeled as functors from a schema category to sets and their temporal evolution is encoded as sheaves or cosheaves on a poset of time intervals (Bumpus et al., 2024). The persistent-cumulative dichotomy is formalized as an adjunction between cosheaves and sheaves, with temporalization of network invariants via limits/colimits of instantaneous graphs.

6. Applications and Case Studies

Multilayer time-varying graph models have been applied to:

• Structural analysis of transportation networks (e.g., Brazilian airline routes, carrier adaptation strategies during economic crises) with MAG, enabling per-carrier, temporal, and multi-scale decompositions (Costa et al., 2017).
• Dynamic interaction and forecasting in multivariate time series, with evolving graph learning frameworks outperforming static alternatives (Ye et al., 2022).
• International relations and infectious disease spread, with dynamic eigenmodels parsing both shared and layer-specific temporal trajectories, degree dynamics, and latent homophily (Loyal et al., 2021).
• Anomaly detection and temporal graph isomorphism, utilizing low-dimensional graph-level embeddings derived from multilayer walks and document-level LLMs (Wang et al., 2023).

7. Advantages, Limitations, and Future Directions

Advantages of the multilayer time-varying graph paradigm include: lossless fusion of structural and temporal data, direct applicability of classical and modern algorithms, scalable embedding for retrieval and anomaly detection, and extensibility to high-order networks. The MAG formalism is object-agnostic and supports efficient algorithmic pipelines and theoretical guarantees (Wehmuth et al., 2014, Wehmuth et al., 2015, Bumpus et al., 2024, Wang et al., 2023).

Limitations are observed with respect to the storage and indexing demands for large composite vertex sets (mitigated by sparsity), indirect modeling of long-range temporal dependencies (requiring sufficiently deep random walks or recurrent inference), resource intensity of document-level embedding, and the need for cross-validation on critical hyperparameters. The selection of sub-determination aspects or temporal gluing intervals remains application-driven. Sheaf-theoretic and categorical approaches offer promising unification and generalization, with future directions including automatic model selection, extension to weighted/directed networks, and further integration of dynamical systems and narrative morphism theory (Bumpus et al., 2024, Loyal et al., 2021).

---

References:

(Wehmuth et al., 2014): On MultiAspect Graphs (Wehmuth et al., 2015): MultiAspect Graphs: Algebraic representation and algorithms (Costa et al., 2017): A Multilayer and Time-varying Structural Analysis of the Brazilian Air Transportation Network (Wang et al., 2023): Graph-Level Embedding for Time-Evolving Graphs (Ye et al., 2022): Learning the Evolutionary and Multi-scale Graph Structure for Multivariate Time Series Forecasting (Loyal et al., 2021): An Eigenmodel for Dynamic Multilayer Networks (Bumpus et al., 2024): Towards a Unified Theory of Time-Varying Data