Multilayer Graph Representations

• Multilayer Graph Representations are mathematical constructs encoding multiple, heterogeneous relationships among aligned entities across distinct layers.
• They employ aggregation methods like convex aggregation and supra-adjacency to combine diverse data modalities while preserving crucial network structures.
• Advances in spectral analysis and deep learning have enabled robust clustering, embedding, and inference for scalable applications in real-world networks.

A multilayer graph, also commonly termed a multiplex or multilayer network, is a mathematical object encoding multiple, potentially heterogeneous, types of relationships among a set of entities. Each "layer" corresponds to one edge relation (or data modality), and nodes are typically aligned across layers. Such representations are fundamental in analyzing networks with multifaceted connectivity—examples include social networks with various kinds of ties, heterogeneous bio-networks, multichannel sensor data, and time-evolving graphs, as well as the internal representations of deep learning models. This article describes the principal mathematical models, spectral and learning-theoretic frameworks, and algorithmic implications of multilayer graph representations, with a focus on formal definitions, aggregation strategies, representation learning, and clustering, as well as connections to network science, optimization, and graph signal processing.

1. Formal Models of Multilayer Graphs

A multilayer graph consists of a set of entities $V = \{1,\dots,n\}$ and $L$ layers, each associated with an undirected or directed graph $G^{(\ell)} = (V, E^{(\ell)})$, $\ell=1,\dots,L$. The adjacency matrices $A^{(\ell)} \in \mathbb{R}^{n \times n}$ encode the intra-layer edge structure, which may be binary or weighted, and are typically symmetric and hollow for undirected simple graphs. The ensemble $\{A^{(\ell)}\}_{\ell=1}^L$ is known as the multilayer adjacency tensor in some contexts.

A more general formalization is the MultiAspect Graph (MAG), suitable for representing higher-order and time-varying systems, as $H = (A, E)$, where $A$ is a list of $p$ aspects and edges are $2p$-tuples forming $E \subseteq \times_{n=1}^{2p} A[(n-1 \bmod p)+1]$. This formalism enables flexible representation of multi-mode, time-resolved, or attribute-rich structures (Wehmuth et al., 2015). Also relevant are supra-adjacency constructions, where all possible node-layer pairs are expanded into a single, large graph with both intra- and inter-layer (coupling) edges (Sánchez-García et al., 2013).

One important distinction is whether inter-layer edges exist (connecting a node to its counterpart across layers) or if layers are only coupled via aggregation or implicit correspondences. In some domains, layers are node-aligned and share the same $V$; in others, node sets may differ.

2. Aggregation and Dimensionality Reduction Strategies

A central challenge in multilayer graph analysis is how to combine or aggregate the different layers for inference or representation learning. Several approaches are established:

• Convex Layer Aggregation: The most principled and widely adopted method is convex aggregation, where a weight vector $w \in \Delta_L$ (the (L−1)-simplex) combines adjacency matrices as $A(w) = \sum_{\ell=1}^L w_\ell A^{(\ell)}$ (Chen et al., 2016). This preserves positive-semidefiniteness and admits geometric and statistical interpretations. Searching over $w$ can adaptively emphasize more informative layers.
• Network and Results Aggregation (Embeddings): One may merge all layers into a single simple graph (network aggregation) or concatenate per-layer embeddings after running a node embedding method on each layer separately (results aggregation) (Liu et al., 2017).
• Supra-Adjacency and Quotient Graphs: The supra-adjacency matrix represents all node-layer pairs as vertices and models both intra- and inter-layer edges. Two spectral quotient constructions are important: the "network of layers" quotient collapses all nodes within a layer, yielding an $L \times L$ layer-coupling matrix; the "aggregate network" collapses all layer-instances of each node, reflecting overall connectivity (Sánchez-García et al., 2013).
• Tensor-Based Constructions: In applications such as graph signal processing, multilayer graphs are encoded as fourth-order adjacency tensors $$\mathcal{A} \in \mathbb{R}^{M \times N \times M \times N}$$, capturing the full range of intra- and inter-layer connections, naturally supporting spectral and multilinear analysis (Zhang et al., 2021).

The table below summarizes several aggregation methods and their properties, as established in (Liu et al., 2017, Chen et al., 2016, Sánchez-García et al., 2013):

Method	Inter-Layer Info	Dimensionality	Preserves Layer Identity
Convex aggregation	No	$n \times n$	No
Supra-adjacency	Yes	$nL \times nL$	Yes
Results aggregation	No	$d' |L|$	Yes (in embedding)
Layer co-analysis	Yes (random walk)	$d$	Partial

3. Spectral and Statistical Foundations

Spectral analysis lies at the core of multilayer graph representation, both for unsupervised clustering and as a basis for regularization or signal propagation:

• For convexly aggregated graphs, the (combinatorial or normalized) Laplacian $L(w) = D(w) - A(w)$ is constructed, and spectral clustering proceeds via the $K-1$ non-trivial smallest eigenvectors. The aggregated noise level $\theta(w)$ controls a sharp phase transition in cluster separability, with closed-form lower and upper bounds on the critical value $\theta^*$ (Chen et al., 2016).
• Joint Graph Learning and Embedding: Modern approaches jointly optimize per-layer Laplacians and a shared low-dimensional node embedding. The regularized objective involves smoothness across layers, sparsity ($\|\cdot\|_F^2$ penalties), and rank constraints (for exact multi-component structure), solved via alternating minimization (Gurugubelli et al., 2020).
• Latent Position and Probabilistic Models: The multilayer random dot product graph (MRDPG) extends latent position models to multiple graphs, sharing node embeddings across layers and supporting joint spectral embedding and probabilistic community detection (Jones et al., 2020). Hierarchical latent-variable mixtures formalize more general Bayesian multi-layer coupling (Oselio et al., 2013).
• Spectral Interlacing and Reduced Models: Eigenvalue interlacing guarantees that the spectra of network-of-layers and aggregate quotients bound the true multilayer spectrum, which has key implications for modeling diffusion, spreading, and synchronization, as well as for efficient computation of extremal eigenvalues (Sánchez-García et al., 2013).

4. Representation Learning and Embedding Architectures

Deep and shallow embedding methods for multilayer graphs have recently attracted great interest:

• Layer-Aware Embeddings: Node2vec-based embedding pipelines can be extended with random walks that traverse both within- and across-layer edges, with a tunable parameter controlling the trade-off (layer co-analysis) (Liu et al., 2017).
• Graph Neural Architectures: Multilayer representations are crucial in GNN design. Multi-Level Attention Pooling (MLAP) produces per-layer embeddings and fuses them, addressing the locality/globality trade-off and oversmoothing, with significant classification gains (Itoh et al., 2021).
• Multiplex Embedding and Inter-Layer Link Prediction: The MultiSAGE architecture generalizes GraphSAGE, introducing parallel neighborhood aggregators for intra- and inter-layer neighbors, and achieves improved inter-layer link prediction performance (Gallo et al., 2022).
• Community Search with Layer-Shared and Layer-Specific Embeddings: Recent models (e.g., EnMCS) utilize separate but coupled graph diffusion and neural feature encoders to obtain both layer-shared and layer-specific embeddings, yielding robust methods for retrieving query-driven communities that simultaneously exploit consensus and heterogeneity across layers (Wang et al., 4 Jan 2025).

5. Clustering, Optimization, and Inference

Clustering in multilayer graphs leverages the above representations, but requires methods that can preserve or exploit cluster structure while fusing multiple relations:

• Spectral Clustering via Convex Aggregation: A critical noise threshold governs the recoverability of true underlying clusters; robust theoretical guarantees hold, including exact separation when cluster sizes are balanced (Chen et al., 2016).
• Contrastive and Regularized Learning: Effective resistance and eigenvalue penalties can enforce sparsity and community structure in the learned representative graph, with fully differentiable objectives amenable to projected gradient methods. These regularizers translate to interpretable controls on the learned cluster number and graph connectivity (Gheche et al., 2020).
• Multi-View Consensus and Subspace Fusion: For deep subspace clustering, building a multilayer graph from all encoder layers and aligning eigenbases (subspaces) yields strong improvements, both empirically and theoretically, over single-layer approaches (Sindičić et al., 2024).

6. Extensions, Applications, and Theoretical Insights

• Temporal and Streaming Extensions: Multilayer graph frameworks extend naturally to temporal and streaming domains, with structures such as multilayer stream graphs representing edges as events indexed by time, node, and layer. Layer-specific centrality can be rigorously quantified in these dynamic contexts (Parmentier et al., 2019).
• Graph Signal Processing: Tensor-based representations of multilayer graphs generalize the notion of graph Fourier transforms. These are essential for image and signal processing applications, enabling superior denoising, compression, and segmentation compared to layer-wise or mean-graph approaches (Zhang et al., 2021).
• Multilayer Graphlets and Motif Analysis: The compositional structure of multilayer graphs admits the enumeration and analysis of subgraphlets, automorphism orbits, and their algebraic dependencies, supporting the comparative analysis of complex network models and empirical motifs (Sallmen et al., 2021).
• Parametric Synthesis and Optimization: In engineering contexts (e.g., broadband network design), multilayer graphs encode physical and logical overlays. Optimization of resource allocation, subject to flow-aggregation and self-similar traffic models, can be rigorously formulated as nonlinear programs solved by gradient descent (Ageyev et al., 2012).

7. Limitations, Open Problems, and Directions

Key limitations and ongoing challenges include:

• Scalability: Supra-adjacency and tensor formulations incur significant computational and memory overhead for large-scale or high-aspect networks. Approximate or low-rank methods are often necessary.
• Nonlinear and Higher-Order Couplings: Convex aggregation cannot capture nonlinear inter-layer phenomena, and most models assume pairwise (or at most multilayer) interactions, whereas domain instances may exhibit complex dependency structures.
• Heterogeneous Node Sets and Missing Data: Many models require that all layers are node-aligned; relaxing this assumption to accommodate partial overlap is an open challenge.
• Parameter Sensitivity and Model Selection: Aggregation weights, regularization strengths, and the choice of subspace dimension or cluster number require careful tuning, and principled model selection criteria remain an active area.
• Interpretability: While layer weights and co-analysis parameters admit interpretation, deeper neural models for multilayer graphs raise new questions on explainability and robustness.

In summary, multilayer graph representations provide a comprehensive and mathematically rigorous way to encode and analyze complex networked systems with multiple interaction types, time-evolution, or layered semantics. Ongoing research spans foundations, scalable algorithms, and applications in statistical inference, learning, and beyond (Chen et al., 2016, Liu et al., 2017, Itoh et al., 2021, Sánchez-García et al., 2013, Gurugubelli et al., 2020, Jones et al., 2020, Gallo et al., 2022, Gheche et al., 2020, Zhang et al., 2021, Parmentier et al., 2019, Sindičić et al., 2024).