Heterogeneous Network Modelling

• Heterogeneous network modelling is a framework that retains diverse types of nodes, edges, and temporal variations to accurately reflect system complexities.
• It is applied in domains such as wireless communications, traffic systems, and social dynamics to model non-uniform interactions and layered structures.
• Methodologies span graph theory, stochastic processes, and simulation, offering insights into interference, clustering, and performance trade-offs.

Searching arXiv for recent and foundational papers on heterogeneous network modelling. Heterogeneous network modelling denotes a family of formalisms for representing systems in which entities, relations, dynamics, or layers are not uniform. Across the literature, heterogeneity may refer to node and edge types in multipartite graphs, tier-specific parameters in cellular systems, node-level status or costs in diffusion and formation models, mixed intersection types in traffic networks, or edge-dependent travel-time laws in stochastic transport (Chatterjee et al., 2023). The common methodological aim is to preserve structured non-uniformity rather than collapsing it into homogeneous abstractions. In the surveyed work, this leads to models based on typed graphs, layered and hybrid network representations, marked and non-Poisson point processes, latent spaces with heterogeneous sender and receiver effects, queueing-coupled spatial networks, and coarse-grained dynamical reductions (D'Angelo et al., 2018).

1. Conceptual scope and formal representations

A central distinction in heterogeneous network modelling is whether heterogeneity is encoded in graph semantics, spatial deployment, node attributes, temporal laws, or all of them simultaneously. In a graph-theoretic setting, a Heterogeneous Interaction Network is defined from disjoint node-type sets $V_1,\dots,V_K$, with $V=\bigcup_{k=1}^K V_k$, heterogeneous edge sets $E_{k\ell}\subseteq V_k\times V_\ell$, a full edge set $E=\bigcup_{k<\ell}E_{k\ell}$, and a weight function $\omega:E\to\mathbb N$, giving the quadruple $G=(V_1,\dots,V_K;E;\omega)$ or $G=(V,E,\omega)$ when typing is understood (Feng et al., 11 Jan 2026). In the bipartite case, this reduces to a weighted adjacency matrix $A\in\mathbb N^{N_1\times N_2}$ with $A_{ij}=\omega((i,j))$ if $(i,j)\in E$, and to higher-order adjacency tensors when more modes are retained (Feng et al., 11 Jan 2026).

A more general unifying construction is the Hybrid Layered Network (HLN), defined as

$V=\bigcup_{k=1}^K V_k$0

where $V=\bigcup_{k=1}^K V_k$1 is a finite set of layers, $V=\bigcup_{k=1}^K V_k$2, $V=\bigcup_{k=1}^K V_k$3 assigns objects to layers, $V=\bigcup_{k=1}^K V_k$4 assigns vertex types, and $V=\bigcup_{k=1}^K V_k$5 assigns edge types (Chatterjee et al., 2023). Classical heterogeneous graphs embed into HLN by choosing a single layer $V=\bigcup_{k=1}^K V_k$6, while multilayer and mixed-type systems are retained without flattening (Chatterjee et al., 2023). The paper proves that the sets of all homogeneous, heterogeneous and multi-layered networks are subsets of the set of all HLNs, and defines multilayer neighborhood, degree centrality, closeness centrality and betweeness centrality in this generalized setting (Chatterjee et al., 2023).

Other literatures formalize heterogeneity differently. In multidimensional latent-space models, one has $V=\bigcup_{k=1}^K V_k$7 binary adjacency matrices $V=\bigcup_{k=1}^K V_k$8 on a common node set, latent coordinates $V=\bigcup_{k=1}^K V_k$9, and layer-specific or shared sender and receiver effects $E_{k\ell}\subseteq V_k\times V_\ell$0 and $E_{k\ell}\subseteq V_k\times V_\ell$1, yielding a logistic link

$E_{k\ell}\subseteq V_k\times V_\ell$2

with $E_{k\ell}\subseteq V_k\times V_\ell$3 and $E_{k\ell}\subseteq V_k\times V_\ell$4 (D'Angelo et al., 2018). Here heterogeneity is not a graph type label but a layer-varying node effect.

This diversity of definitions implies that “heterogeneous network” is not a single model class. It is a modelling stance: preserve the asymmetries induced by types, tiers, layers, roles, status distributions, or temporal mechanisms when these affect observables.

2. Spatial stochastic models for heterogeneous wireless systems

In wireless communications, heterogeneous network modelling is dominated by stochastic geometry. A generic $E_{k\ell}\subseteq V_k\times V_\ell$5-tier heterogeneous cellular network in the downlink is modelled by independent planar Poisson point processes $E_{k\ell}\subseteq V_k\times V_\ell$6 of density $E_{k\ell}\subseteq V_k\times V_\ell$7, with tier-specific transmit powers $E_{k\ell}\subseteq V_k\times V_\ell$8, path-loss exponents $E_{k\ell}\subseteq V_k\times V_\ell$9, and Rayleigh fading gains $E=\bigcup_{k<\ell}E_{k\ell}$0 (Tanbourgi et al., 2014). Users associate with the base station offering the highest long-term average received power $E=\bigcup_{k<\ell}E_{k\ell}$1, and the resulting interference can be analyzed through joint Laplace transforms that retain cross-spectrum dependence (Tanbourgi et al., 2014).

A related framework considers $E=\bigcup_{k<\ell}E_{k\ell}$2 open-access tiers and $E=\bigcup_{k<\ell}E_{k\ell}$3 closed-access tiers, each deployed as independent homogeneous PPPs, with arbitrary fading distributions satisfying $E=\bigcup_{k<\ell}E_{k\ell}$4, path-loss $E=\bigcup_{k<\ell}E_{k\ell}$5, and connectivity rules based on max-SINR, nearest-BS, MIRP, or MBRP association (Madhusudhanan et al., 2014). Closed-form or semi-analytic coverage expressions are then derived by Laplace-transform methods, including the interference-limited equal-$E=\bigcup_{k<\ell}E_{k\ell}$6 regime where coverage simplifies to explicit ratios involving $E=\bigcup_{k<\ell}E_{k\ell}$7 terms (Madhusudhanan et al., 2014). The same formalism yields average ergodic rate and average per-tier load (Madhusudhanan et al., 2014).

The Poisson assumption is analytically convenient but not universal. Non-Poisson models introduce repulsion through the Matern Hard-Core Process and Determinantal Point Process, or attraction through Neyman–Scott cluster processes (Chun et al., 2015). For example, a Matern hard-core process $E=\bigcup_{k<\ell}E_{k\ell}$8 is obtained from a PPP $E=\bigcup_{k<\ell}E_{k\ell}$9 of intensity $\omega:E\to\mathbb N$0 by imposing a hard-core radius $\omega:E\to\mathbb N$1, leading to intensity $\omega:E\to\mathbb N$2 and a nontrivial pair correlation function $\omega:E\to\mathbb N$3 (Chun et al., 2015). Interference Laplace transforms are then modified through pair-correlation or Fredholm-determinant machinery rather than the PGFL of a homogeneous PPP (Chun et al., 2015).

Several papers explicitly model dependencies between tiers rather than treating each tier as independent. Hänggi’s dependent four-tier HetNet places Tier 2 nodes on Voronoi edges of the macrocell tessellation, Tier 3 nodes at thinned Voronoi vertices, and Tier 4 either as an independent homogeneous PPP or a Cox/cluster process over population centers (Haenggi, 2013). The construction yields stationary intensities $\omega:E\to\mathbb N$4 for Tier 2 and $\omega:E\to\mathbb N$5 for Tier 3, and is motivated by “worst-cover” locations in the macrocell Voronoi diagram (Haenggi, 2013). The reported simulations compare covered area fraction $\omega:E\to\mathbb N$6 across coverage-oriented and baseline deployments; for example, Example 1 reports uncovered area $\omega:E\to\mathbb N$7 at $\omega:E\to\mathbb N$8, while Example 2 reports $\omega:E\to\mathbb N$9 under an independent two-tier baseline (Haenggi, 2013).

Further dependence is introduced by Poisson hole processes in two-tier mmWave HetNets, where macrocell BSs form a PPP $G=(V_1,\dots,V_K;E;\omega)$0 and small-cell BSs are drawn from a baseline PPP $G=(V_1,\dots,V_K;E;\omega)$1 but excluded from circular-sector holes $G=(V_1,\dots,V_K;E;\omega)$2 around each MBS (Sattari et al., 2018). This PHP captures spatial separation between macro and small cells, incorporates blockage and directional beamforming, and yields semi-closed expressions for distance distributions, association probabilities, and coverage probabilities (Sattari et al., 2018).

A separate line of work studies equivalence rather than direct deployment realism. In a general marked Poisson network, each BS carries i.i.d. marks $G=(V_1,\dots,V_K;E;\omega)$3, and the propagation loss

$G=(V_1,\dots,V_K;E;\omega)$4

induces an inhomogeneous PPP on $G=(V_1,\dots,V_K;E;\omega)$5 (Blaszczyszyn et al., 2013). The marked propagation process can be replaced, from the perspective of a typical user, by an isotropic inhomogeneous PPP with constant propagation parameters and radial density

$G=(V_1,\dots,V_K;E;\omega)$6

which absorbs variability in powers, shadowing, path-loss constants, and path-loss exponents into spatial density and distance-dependent threshold laws (Blaszczyszyn et al., 2013). This suggests that distinct heterogeneous deployments may be equivalent under user-centric propagation observables even when their physical parameters differ.

3. Interference, traffic, and performance laws

A defining feature of heterogeneous wireless models is that infrastructure heterogeneity is often coupled to traffic heterogeneity and interference correlation. In the downlink under frequency diversity, resources are partitioned into two subsets $G=(V_1,\dots,V_K;E;\omega)$7 and $G=(V_1,\dots,V_K;E;\omega)$8 of sizes $G=(V_1,\dots,V_K;E;\omega)$9 and $G=(V,E,\omega)$0, with a two-block independent-fading model but shared BS locations across blocks (Tanbourgi et al., 2014). The interference terms $G=(V,E,\omega)$1 and $G=(V,E,\omega)$2 are therefore correlated, with single-tier joint Laplace transform

$G=(V,E,\omega)$3

and rate coverage is derived by conditioning on serving distance and serving tier (Tanbourgi et al., 2014). The paper reports that frequency-diversity gain increases rate-coverage by 40–90% for typical targets, that the maximal gain occurs when $G=(V,E,\omega)$4, and that ignoring the correlation across subbands overestimates rate-coverage by about 3–6% (Tanbourgi et al., 2014).

The interference field itself can be modelled by a hybrid construction consisting of a fixed circular typical cell, tier-dependent guard regions, one dominant interferer per tier, and PPP shot noise outside the guard regions (Jr et al., 2012). Out-of-cell interferers of tier $G=(V,E,\omega)$5 form an independent PPP $G=(V,E,\omega)$6 of density $G=(V,E,\omega)$7, each tier has guard radius

$G=(V,E,\omega)$8

and the total interference is approximated by a Gamma distribution with parameters

$G=(V,E,\omega)$9

after matching the first two moments (Jr et al., 2012). This Gamma approximation yields closed-form success probability expressions involving the regularized Gauss-hypergeometric function and simplifies average-rate calculation while retaining co-tier and cross-tier interference (Jr et al., 2012).

When traffic itself is heterogeneous, the classical independent-PPP user model becomes insufficient. “HetHetNets” introduces two statistical controls for UE distributions: a coefficient of variation $A\in\mathbb N^{N_1\times N_2}$0 of a Voronoi-cell-area-based inter-UE distance measure, and a normalized cross-moment $A\in\mathbb N^{N_1\times N_2}$1 between UEs and a BS-induced potential field (Mirahsan et al., 2015). The inter-UE measure is $A\in\mathbb N^{N_1\times N_2}$2, the area of the Voronoi cell around user $A\in\mathbb N^{N_1\times N_2}$3, and the coefficient of variation is

$A\in\mathbb N^{N_1\times N_2}$4

A continuous potential field $A\in\mathbb N^{N_1\times N_2}$5, with $A\in\mathbb N^{N_1\times N_2}$6 at the center of each BS’s power-weighted Voronoi cell and $A\in\mathbb N^{N_1\times N_2}$7 on every cell edge, yields

$A\in\mathbb N^{N_1\times N_2}$8

A three-step “social-attractor” construction parameterized by $A\in\mathbb N^{N_1\times N_2}$9 then generates UE patterns spanning the feasible region of $A_{ij}=\omega((i,j))$0 (Mirahsan et al., 2015). The reported system-level findings are that as $A_{ij}=\omega((i,j))$1 increases, both coverage probability and mean user rate monotonically increase; when $A_{ij}=\omega((i,j))$2, increasing $A_{ij}=\omega((i,j))$3 hurts performance; and when $A_{ij}=\omega((i,j))$4 is already high, further clustering can help (Mirahsan et al., 2015).

At a larger scale, queueing-theoretic performance laws have been derived for large heterogeneous cellular networks where call arrivals form a space-time Poisson process of intensity $A_{ij}=\omega((i,j))$5, each call carries a data-volume mark $A_{ij}=\omega((i,j))$6, and each BS serves its shadow-perturbed Voronoi cell under processor sharing (Blaszczyszyn et al., 2014). With load $A_{ij}=\omega((i,j))$7, mean number of users

$A_{ij}=\omega((i,j))$8

and mean throughput $A_{ij}=\omega((i,j))$9, the coupled cell-load equations define a semi-analytic workflow: static Monte Carlo of geometry and shadows, iterative solution of the fixed-point system for $(i,j)\in E$0, classical PS-queue formulas, and Palm averaging (Blaszczyszyn et al., 2014). The resulting macroscopic laws express average traffic per cell, average load, and mean throughput in terms of $(i,j)\in E$1, $(i,j)\in E$2, $(i,j)\in E$3, $(i,j)\in E$4, $(i,j)\in E$5, the peak-rate function $(i,j)\in E$6, and the spatial-average SINR distribution (Blaszczyszyn et al., 2014).

These results collectively show that in heterogeneous wireless modelling, “heterogeneity” is not confined to BS tiers. It also enters through dependence between tiers, nonuniform demand, interference correlation across frequency, space, and time, and queueing feedback between load and SINR.

4. Dynamical heterogeneity on networks

Outside communications, heterogeneous network modelling often focuses on how non-uniform node or edge characteristics alter diffusion, consensus, and transport. The Heterogeneous Opinion-Status model (HOpS) is defined on a static undirected graph $(i,j)\in E$7 with binary opinions $(i,j)\in E$8 and fixed statuses $(i,j)\in E$9 (Tupikina, 2017). At each discrete time step, an active node $V=\bigcup_{k=1}^K V_k$00 chooses a random neighbor $V=\bigcup_{k=1}^K V_k$01 and adopts $V=\bigcup_{k=1}^K V_k$02’s opinion with probability

$V=\bigcup_{k=1}^K V_k$03

The status vector $V=\bigcup_{k=1}^K V_k$04 introduces node-level heterogeneity into the contagion dynamics (Tupikina, 2017). On a linear chain with a special initial condition, the opinion boundary executes an asymmetric random walk with probabilities $V=\bigcup_{k=1}^K V_k$05 and $V=\bigcup_{k=1}^K V_k$06, enabling analytic hitting probabilities, spectral gaps, and mixing-time estimates (Tupikina, 2017). On star networks, symmetry reduces the state space to $V=\bigcup_{k=1}^K V_k$07 macrostates $V=\bigcup_{k=1}^K V_k$08, and absorption probabilities and mean times follow by fundamental-matrix techniques (Tupikina, 2017).

A continuous-time transport counterpart is the Heterogeneous Continuous-Time Random Walk (HCTRW), defined on a connected graph $V=\bigcup_{k=1}^K V_k$09 with stochastic matrix $V=\bigcup_{k=1}^K V_k$10 and edge-dependent travel-time densities $V=\bigcup_{k=1}^K V_k$11 (Tupikina et al., 2018). The generalized Montroll–Weiss equation yields in Laplace space

$V=\bigcup_{k=1}^K V_k$12

so structural heterogeneity is encoded in $V=\bigcup_{k=1}^K V_k$13 and temporal heterogeneity in $V=\bigcup_{k=1}^K V_k$14 (Tupikina et al., 2018). Exponential waiting times correspond to regular edges, while heavy-tailed Mittag–Leffler-type laws model trap nodes, producing algebraic first-passage tails (Tupikina et al., 2018). The framework defines first-passage-based network measures including MFPT, most probable FPT, quantiles, closeness-FPT, efficiency-FPT, and variability centrality (Tupikina et al., 2018).

A different reductionist approach treats node identity as a heterogeneous parameter and imports tools from uncertainty quantification. Each node $V=\bigcup_{k=1}^K V_k$15 carries a feature vector $V=\bigcup_{k=1}^K V_k$16, the empirical distribution $V=\bigcup_{k=1}^K V_k$17 is used as a weight function, and the mean state as a function of identity $V=\bigcup_{k=1}^K V_k$18 is expanded in orthogonal polynomials $V=\bigcup_{k=1}^K V_k$19 (Rajendran et al., 2015). In the univariate case,

$V=\bigcup_{k=1}^K V_k$20

with coefficients obtained by projection under the $V=\bigcup_{k=1}^K V_k$21-weighted inner product (Rajendran et al., 2015). A coarse time-stepper $V=\bigcup_{k=1}^K V_k$22 is then used for coarse projective integration and matrix-free fixed-point computation (Rajendran et al., 2015). This suggests an alternative view of heterogeneous network modelling: rather than encoding types as graph structure, one may encode them as distributions over slow variables that support reduced-order dynamics.

5. Learning, inference, and algorithmic frameworks

Recent work treats heterogeneous networks as a substrate for end-to-end learning. HINA introduces a three-level workflow on bipartite or multipartite Heterogeneous Interaction Networks: node-level summaries, dyadic-level significance testing, and meso-level nonparametric clustering (Feng et al., 11 Jan 2026). For a bipartite HIN $V=\bigcup_{k=1}^K V_k$23, the total network weight is

$V=\bigcup_{k=1}^K V_k$24

the quantity measure is

$V=\bigcup_{k=1}^K V_k$25

and the diversity measure is the normalized Shannon entropy

$V=\bigcup_{k=1}^K V_k$26

with $V=\bigcup_{k=1}^K V_k$27 (Feng et al., 11 Jan 2026). Dyadic pruning retains only edges whose weight exceeds a null-model threshold, and meso-level clustering minimizes an MDL-based description length $V=\bigcup_{k=1}^K V_k$28 over partitions $V=\bigcup_{k=1}^K V_k$29 of one node set (Feng et al., 11 Jan 2026). The framework’s emphasis is not merely on representation but on multi-scale statistical analysis.

HMill provides a distinct learning architecture for heterogeneous, hierarchical, and graph-structured data through three node types: array nodes, bag nodes, and product nodes (Mandlik et al., 2021). Array nodes $V=\bigcup_{k=1}^K V_k$30 carry information fragments mapped into Euclidean vectors by $V=\bigcup_{k=1}^K V_k$31, bag nodes $V=\bigcup_{k=1}^K V_k$32 represent unordered multisets, and product nodes $V=\bigcup_{k=1}^K V_k$33 assemble heterogeneous pieces according to schema (Mandlik et al., 2021). Model trees mirror sample trees through array models, bag models $V=\bigcup_{k=1}^K V_k$34, and product models $V=\bigcup_{k=1}^K V_k$35, with permutation-invariant aggregations such as mean, max, log-sum-exp, and $V=\bigcup_{k=1}^K V_k$36-norm (Mandlik et al., 2021). The framework extends the universal approximation theorem from Euclidean domains to bags of measures and Cartesian products of measures, and supports message-passing inference by treating each vertex as a sample with bags of neighbors under different graphs (Mandlik et al., 2021).

Synthetic data generation is another algorithmic concern. The HLN paper provides a parameterized generation algorithm in which one fixes layers $V=\bigcup_{k=1}^K V_k$37, allowed node-types per layer, minimum intra-layer degrees $V=\bigcup_{k=1}^K V_k$38, minimum inter-layer links $V=\bigcup_{k=1}^K V_k$39, and weights $V=\bigcup_{k=1}^K V_k$40 that control “self-degree” versus “neighbor-degree” in preferential attachment (Chatterjee et al., 2023). Inserting one node has complexity $V=\bigcup_{k=1}^K V_k$41 for intra-layer edges plus $V=\bigcup_{k=1}^K V_k$42 for inter-layer edges (Chatterjee et al., 2023). The generated networks are reported to be more consistent in modelling the layer-wise degree distribution of a real-world Twitter network than existing models (Chatterjee et al., 2023).

Bayesian inference remains prominent when heterogeneous effects are latent rather than typed. In the multidimensional latent-space model, a Metropolis-within-Gibbs sampler updates latent positions, sender and receiver effects, and layer parameters; initialization is based on classical MDS and logistic fits; and convergence is assessed with trace-plots, Geweke-, and Heidelberger–Welch diagnostics (D'Angelo et al., 2018). Simulations for settings $V=\bigcup_{k=1}^K V_k$43, $V=\bigcup_{k=1}^K V_k$44 and $V=\bigcup_{k=1}^K V_k$45 report distance-correlations always $V=\bigcup_{k=1}^K V_k$46, typically $V=\bigcup_{k=1}^K V_k$47–$V=\bigcup_{k=1}^K V_k$48, Procrustes correlations $V=\bigcup_{k=1}^K V_k$49, and Spearman’s $V=\bigcup_{k=1}^K V_k$50 for $V=\bigcup_{k=1}^K V_k$51 often $V=\bigcup_{k=1}^K V_k$52 (D'Angelo et al., 2018).

Taken together, these frameworks show that heterogeneous network modelling now includes representation learning, statistical validation, nonparametric clustering, universal approximation over structured inputs, and synthetic benchmark generation, in addition to classical analytic models.

6. Domain-specific formulations and inverse design

Heterogeneous network modelling is strongly domain-dependent, and the notion of heterogeneity changes with the scientific object. In optical disordered materials, a two-phase material is modelled as a weighted graph on particles, with node set $V=\bigcup_{k=1}^K V_k$53, phase partition $V=\bigcup_{k=1}^K V_k$54, and a fully connected symmetric weighted adjacency matrix $V=\bigcup_{k=1}^K V_k$55 derived from scattering amplitudes under the first-Born approximation (Youn et al., 30 Jul 2025). The structure factor averaged over a reciprocal-space region $V=\bigcup_{k=1}^K V_k$56 is expressed through network weights as

$V=\bigcup_{k=1}^K V_k$57

and decomposes into $V=\bigcup_{k=1}^K V_k$58, $V=\bigcup_{k=1}^K V_k$59, and $V=\bigcup_{k=1}^K V_k$60 subnetworks (Youn et al., 30 Jul 2025). A phase-sensitive microstructure design algorithm then adds particles sequentially by minimizing

$V=\bigcup_{k=1}^K V_k$61

where $V=\bigcup_{k=1}^K V_k$62 tune the relative importance of intra- and inter-phase links (Youn et al., 30 Jul 2025). For quasi-isoscattering stealthy hyperuniform materials, all designed samples achieve near-identical suppression of $V=\bigcup_{k=1}^K V_k$63 in $V=\bigcup_{k=1}^K V_k$64, while node-degree distributions and real-space microstatistics differ by phase as $V=\bigcup_{k=1}^K V_k$65 increases (Youn et al., 30 Jul 2025).

In traffic systems, heterogeneity is infrastructural. An urban network $V=\bigcup_{k=1}^K V_k$66 consists of directed links $V=\bigcup_{k=1}^K V_k$67 and intersections $V=\bigcup_{k=1}^K V_k$68, where $V=\bigcup_{k=1}^K V_k$69 are signalized intersections and $V=\bigcup_{k=1}^K V_k$70 are non-signalized intersections operating under FCFS gap-acceptance rules (Zhang et al., 2017). Link volumes satisfy

$V=\bigcup_{k=1}^K V_k$71

signalized junctions use stage indicators $V=\bigcup_{k=1}^K V_k$72, speed-level selectors $V=\bigcup_{k=1}^K V_k$73, and discharge flows

$V=\bigcup_{k=1}^K V_k$74

while non-signalized junctions are handled by a cell-transmission-plus-FCFS submodel (Zhang et al., 2017). The signal control problem becomes a mixed integer programming problem with objective of minimizing total network delay, and a Lagrangian-multiplier-based hierarchical distributed solution is constructed by relaxing coupling constraints and updating multipliers through subgradient steps (Zhang et al., 2017). On a $V=\bigcup_{k=1}^K V_k$75 grid over a 180 s horizon, the reported comparison between fully uncontrolled, partially controlled, and fully controlled designs shows that at low demand, non-signalized control suffices, while at moderate to high demands, a mixed heterogeneous design outperforms full signalization (Zhang et al., 2017).

In network formation theory, heterogeneity is cost-based. With agents $V=\bigcup_{k=1}^K V_k$76, adjacency matrix $V=\bigcup_{k=1}^K V_k$77, degree $V=\bigcup_{k=1}^K V_k$78, and separable heterogeneous connection costs $V=\bigcup_{k=1}^K V_k$79, utility is

$V=\bigcup_{k=1}^K V_k$80

and social welfare is $V=\bigcup_{k=1}^K V_k$81 (Heydari et al., 2015). The efficient network has a generalized star structure when connected, a connected core of nodes $V=\bigcup_{k=1}^K V_k$82, and a core-periphery organization characterized by cost thresholds involving $V=\bigcup_{k=1}^K V_k$83 and $V=\bigcup_{k=1}^K V_k$84 (Heydari et al., 2015). The paper further gives a lower bound for the clustering coefficient of efficient networks at fixed density and shows pairwise stability under the stated inequalities (Heydari et al., 2015).

These cases illustrate that heterogeneous network modelling is not restricted to observational data analysis. It is also used as an inverse-design formalism, a control-theoretic substrate, and a normative model for efficient structure formation.

7. Recurrent themes, limitations, and interpretive cautions

Several recurrent themes cut across these literatures. First, homogeneous surrogates are often analytically attractive but structurally lossy. Wireless papers repeatedly note that PPP assumptions, independent tier placement, and interference independence simplify derivations yet may miss dependence between layers, exclusion effects, traffic clustering, or correlation across spectrum, space, and time (Chun et al., 2015). Learning and graph-representation papers make a parallel argument: homogeneous relationships, single-type co-occurrence networks, or fixed-vector input assumptions are limited in capturing distributed, multi-faceted interactions (Feng et al., 11 Jan 2026).

Second, there is a persistent trade-off between tractability and realism. The dependent HetNet model explicitly states that it does not derive closed-form SINR-distribution or coverage-probability expressions and instead focuses on model definition and simulation (Haenggi, 2013). The queueing-plus-Poisson framework for large cellular networks is described as semi-analytic, with temporal evolution handled by queuing-theoretic results and static geometry handled by simulation (Blaszczyszyn et al., 2014). HCTRW retains arbitrary edge-by-edge waiting-time laws, but the price is Laplace-domain analysis and greater inferential complexity (Tupikina et al., 2018).

Third, heterogeneity is often measurable only through carefully chosen summaries. In HINA, the key summaries are quantity $V=\bigcup_{k=1}^K V_k$85, diversity $V=\bigcup_{k=1}^K V_k$86, statistically validated backbones, and MDL-selected meso-level clusters (Feng et al., 11 Jan 2026). In HetHetNets, the central descriptors are the coefficient of variation $V=\bigcup_{k=1}^K V_k$87 and the correlation coefficient $V=\bigcup_{k=1}^K V_k$88 (Mirahsan et al., 2015). In uncertainty-quantification-based reductions, the operative summaries are the polynomial-chaos coefficients $V=\bigcup_{k=1}^K V_k$89 (Rajendran et al., 2015). This suggests that a heterogeneous model is not defined solely by richer state space; it also requires observables that remain interpretable under that richness.

Fourth, equivalence results caution against naive model comparison. The isotropic representation theorem for marked Poisson networks shows that seemingly different heterogeneous networks can be identical in law from the point of view of the typical user once propagation processes are matched (Blaszczyszyn et al., 2013). A plausible implication is that some empirical distinctions between heterogeneous deployments are not operationally meaningful for user-centric metrics, whereas others, such as interference correlation or traffic–infrastructure dependence, remain irreducible because they alter joint distributions rather than only marginal propagation laws.

Finally, domain transfer is possible but not automatic. The term “heterogeneous network” spans cellular infrastructure, learning interactions, latent relational data, opinion dynamics, transport processes, urban junction systems, and wave-scattering materials. The overlap lies in the methodological insistence on typed, layered, parameter-varying, or temporally non-uniform relations. The specific state variables, objective functions, and analytical tools differ sharply.

In that sense, heterogeneous network modelling is best understood as a general modelling doctrine: encode the non-uniform structure that materially affects dynamics, inference, control, or design, and choose the mathematical formalism—typed graph, layered network, point process, latent model, queueing system, or coarse-grained dynamical expansion—that preserves that structure at the scale of the question being asked.