Epidemic processes in complex networks

Romualdo Pastor-Satorras; Claudio Castellano; Piet Van Mieghem; Alessandro Vespignani

Source-linked AI summary

Epidemic processes in complex networks

Romualdo Pastor-Satorras, Claudio Castellano, Piet Van Mieghem, Alessandro Vespignani

arXiv:1408.2701v2Aug 12, 2014physics.soc-phcond-mat.stat-mechcs.SIq-bio.PE

TL;DR

Epidemic modeling in complex networks needs frameworks that address heterogeneous and evolving connectivity. This paper reviews these approaches, clarifies their assumptions and applicability, and synthesizes results across epidemic, social contagion, and time-varying network models.

Problem
Understanding contagion in heterogeneous and coevolving networks remains challenging because network evolution and agents’ responses to contagion are difficult to disentangle.
Method
The paper reviews epidemic models and analytical methodologies, stating their assumptions, applicability, mitigation approaches, and treatment of coevolving, coupled, and time-varying networks.
Results
The review synthesizes theoretical approaches for classical, heterogeneous, adaptive, and time-varying epidemic processes while identifying their assumptions and ranges of applicability.
Takeaways & Limitations
The synthesis indicates that coevolution between network structure and contagion dynamics is a key element of many social networks.
Takeaways & Limitations
Exact Markov approaches are limited to small graphs and often provide limited general insight because their equation sets and generators become complex.

Abstract

from arXiv · show

In recent years the research community has accumulated overwhelming evidence for the emergence of complex and heterogeneous connectivity patterns in a wide range of biological and sociotechnical systems. The complex properties of real-world networks have a profound impact on the behavior of equilibrium and nonequilibrium phenomena occurring in various systems, and the study of epidemic spreading is central to our understanding of the unfolding of dynamical processes in complex networks. The theoretical analysis of epidemic spreading in heterogeneous networks requires the development of novel analytical frameworks, and it has produced results of conceptual and practical relevance. A coherent and comprehensive review of the vast research activity concerning epidemic processes is presented, detailing the successful theoretical approaches as well as making their limits and assumptions clear. Physicists, mathematicians, epidemiologists, computer, and social scientists share a common interest in studying epidemic spreading and rely on similar models for the description of the diffusion of pathogens, knowledge, and innovation. For this reason, while focusing on the main results and the paradigmatic models in infectious disease modeling, the major results concerning generalized social contagion processes are also presented. Finally, the research activity at the forefront in the study of epidemic spreading in coevolving, coupled, and time-varying networks is reported.

I. INTRODUCTION · II. THE MATHEMATICAL APPROACH TO EPIDEMIC SPREADING · A. Classical models of epidemic spreading

The paper frames epidemic modeling as essential for understanding and managing infectious diseases, while emphasizing that complex, heterogeneous networks require unified mathematical approaches. It introduces classical compartmental models and their extensions as foundations for analyzing epidemic dynamics.

I. INTRODUCTION: Mathematical epidemic models support theory testing, uncertainty assessment, and intervention planning because in-vivo experimentation is not viable.The paper traces this modeling tradition from Bernoulli’s 1760 work through Kermack and McKendrick’s 1927 formulation.
I. INTRODUCTION: Large-scale datasets and individual-level population simulations have improved epidemiological model accuracy and enabled quantitative policy analysis.Researchers increasingly advocate using these models as real-time predictive tools.
I. INTRODUCTION: Human interactions, mobility, and contact patterns are naturally represented as networks, whose complex features include fluctuations, clustering, communities, and heavy-tailed distributions.The growing availability of social-web, mobile, and wireless data has strengthened the case for studying epidemic processes on networks.
II. THE MATHEMATICAL APPROACH TO EPIDEMIC SPREADING: Because exact solutions for network dynamical systems are often unattainable, the review unifies fragmented results and clarifies each methodology’s assumptions and applicability.It focuses on influential work across disciplines and compares approximations for static, coevolving, coupled, and time-varying networks.
A. Classical models of epidemic spreading: Classical epidemic models divide populations into disease-state compartments, commonly susceptible (S), infectious (I), and recovered (R).Additional compartments can represent other disease states.
A. Classical models of epidemic spreading: The SIS model permits infection and recovery transitions and can reach an endemic state, whereas infected individuals in the SIR model ultimately decline to zero.SIS persistence requires sufficiently large β or sufficiently small µ; SIR asymptotically dies for any β and µ.
A. Classical models of epidemic spreading: In classical models, infection depends on susceptible–infectious contacts, while recovery or removal occurs after individuals spend time fighting disease or receiving treatment.Homogeneous mixing approximates interactions as random when detailed contact data are unavailable, and transition probabilities are often assumed constant.
A. Classical models of epidemic spreading: Classical frameworks extend to SI, SIRS, and SEIR models, as well as demographic effects, age structure, and asymptomatic compartments.SEIR adds exposed individuals who are infected but not yet able to transmit, while SIRS incorporates temporary immunity.

B. Basic results from classical epidemiology … B. Network metrics

The paper moves from classical deterministic epidemic models and their threshold concept to the limitations imposed by stochasticity, non-Poissonian processes, and homogeneous mixing. It then connects epidemic dynamics with nonequilibrium statistical physics and introduces network theory, graph structures, and topology metrics relevant to spreading.

B. Basic results from classical epidemiology: Classical SIR and SIS equations use the law of mass action, with interaction-driven changes determined by infection force times compartment density.The infection force is defined as α = βρI, with χ = µ for SIS and χ = 0 for SIR.
B. Basic results from classical epidemiology: At the epidemic’s early stage, linearization shows that infectious individuals can grow exponentially when the reproduction condition is satisfied.The approximation assumes ρI ≃0.
B. Basic results from classical epidemiology: R0 is the average number of secondary infections caused by one primary case in a fully susceptible population, and outbreaks require R0 > 1.Above this threshold, SIR-like models can produce finite-size outbreaks or SIS-like models can reach a finite endemic steady state.
B. Basic results from classical epidemiology: The classical deterministic approach neglects spatial diffusion, treats epidemic spreading probabilistically only approximately, and assumes exponentially distributed infection or recovery processes.It also assumes random homogeneous mixing, although individuals’ social contact networks and connectivity patterns differ.
B. Basic results from classical epidemiology: The review focuses on how heterogeneous connectivity patterns and underlying network topology affect epidemic behavior.This focus addresses limitations of classical homogeneous assumptions in real-world contact structures.
C. Connections with other statistical physics models: Epidemic thresholds correspond to phase transitions in nonequilibrium statistical physics, while SIS dynamics belongs to models with absorbing states.In SIS, the infected state is the active phase and the infection-free state is absorbing.
C. Connections with other statistical physics models: SIR dynamics resembles percolation because disease outbreaks transition between affecting a finite population fraction and affecting only a limited number of individuals.Percolation similarly separates macroscopic connected clusters from finite clusters at a critical value pc.
III. NETWORK MEASURES AND MODELS: Network theory supplies graph-based definitions and metrics for analyzing heterogeneous interactions, components, giant components, and topology relevant to epidemic spreading.Graphs consist of vertices joined by edges, and a giant component can support infection of a macroscopic fraction of the graph.

1. Shortest path length and network diameter … 3. Heavy-tailed networks

The paper introduces core structural measures for networks, then distinguishes random, small-world, and heavy-tailed network classes through their connectivity, distances, clustering, correlations, and generative mechanisms.

1. Shortest path length and network diameter: Shortest path distance ℓij is the length of the shortest path between nodes i and j, while network diameter is the maximum pairwise shortest path length.The average shortest path length ⟨ℓ⟩ averages ℓij over all vertex pairs.
2. Degree and degree distribution: Degree counts incident edges, and the degree distribution P(k) describes the probability or fraction of vertices having degree k; its first moment measures network density.For directed networks, in-degree and out-degree have separate distributions, with ⟨kin⟩=⟨kout⟩.
3. Degree correlations: Degree correlations are measured by P(k′|k) or ¯knn(k), with increasing ¯knn(k) indicating assortativity and decreasing ¯knn(k) indicating disassortativity.Uncorrelated networks have ¯knn(k)=⟨k^2⟩/⟨k⟩ and r=0, whereas assortative and disassortative networks have r>0 and r<0, respectively.
4. Clustering coefficient and clustering spectrum: Clustering measures network transitivity through the relative propensity of two nodes sharing a neighbor to connect, using the global coefficient C and local measures ci.The global coefficient C is the ratio of triangles to connected triples.
5. Centrality and structure in networks: Centrality captures relative node importance through measures including degree, betweenness, and K-core decomposition, while community structure identifies densely internally connected modules.Betweenness assumes information follows shortest paths, and K-core decomposition classifies nested connectivity levels.
C. Generalizations of simple graphs: Bipartite graphs connect two different node types, whereas weighted networks assign edge weights ωij and generalize degree through node strength si.Weighted networks represent systems such as transportation networks, where edge weights can measure fractions of people or goods transported.
D. Network classes and basic network models: Empirical network data reveal diverse structural classes, motivating generative models that produce synthetic networks with controlled topological properties; the Erdős–Rényi model yields homogeneous Poisson-like degrees.For ⟨k⟩>1, the random graph has the small diameter observed in many real-world networks, with ⟨ℓ⟩≃log N/log⟨k⟩.

E. Static versus dynamic networks … B. Degree-based mean-field approach

The review contrasts epidemic modeling on static, dynamic, and coevolving networks, then surveys exact Markov and mean-field approaches with their assumptions and applicability limits. It emphasizes IBMF and DBMF, including IBMF’s adjacency-matrix dependence and DBMF’s degree-class and annealed-network approximations.

E. Static versus dynamic networks: Static-network approximations are appropriate when network evolution is slow, whereas comparable spreading and connectivity timescales require modeling their concurrent dynamics.Some networks create, destroy, or rewire links on intrinsic timescales comparable to epidemic dynamics.
E. Static versus dynamic networks: Co-evolution occurs when epidemic-driven behavioral changes alter network topology, which then feeds back on spreading dynamics.Avoiding contacts can effectively delete links and modify the underlying social network.
IV. THEORETICAL APPROACHES FOR EPIDEMIC MODELING ON NETWORKS: Continuous-time compartmental epidemics with constant transition rates on a graph can be represented by a Markov chain with independent Poisson transitions.The state of each node is specified by its compartment-valued random variable.
IV. THEORETICAL APPROACHES FOR EPIDEMIC MODELING ON NETWORKS: Exact Markov analysis is limited because solving q^N × q^N equations restricts applications to very small graphs and the generator structure obscures general insights.Consequently, few exact results exist for epidemic spreading in networks.
A. Individual-based mean-field approach: IBMF simplifies the exact description by assuming node states are statistically independent of nearest-neighbor states while retaining the static network’s full adjacency matrix.The approach factorizes expected variable pairs as E[XiXj] = E[Xi]E[Xj].
A. Individual-based mean-field approach: IBMF predictions depend on the largest adjacency-matrix eigenvalue Λ1 and generally agree with simulations on static networks, but agreement decreases as densities approach zero.The deterioration occurs when the independence assumption breaks down.
B. Degree-based mean-field approach: DBMF treats all nodes of degree k as statistically equivalent and models epidemic partial densities by degree class rather than individual adjacency relations.This replaces the adjacency matrix with an ensemble average preserving degree and two-vertex correlations; for uncorrelated networks, ¯aij = kikj/(N⟨k⟩).
B. Degree-based mean-field approach: DBMF is strong for static networks but suitable when interactions rewire faster than spreading, yielding an annealed mean-field network that preserves P(k) and P(k′|k).Its solutions depend on network topological statistics and, for uncorrelated networks, degree-distribution moments.

C. Generating function approach … 1. Degree-based mean-field theory

Generating functions map SIR outbreak properties to bond percolation and characterize giant-component critical behavior in heterogeneous networks. For SIS dynamics, degree-based mean-field theory relates infection probabilities to degree and derives thresholds and prevalence scaling from network connectivity.

C. Generating function approach: For SIR-like models without steady states, long-time outbreak properties map to bond percolation, with bond occupation probability p representing disease transmission probability.The generating-function framework analyzes whether a giant component exists under this mapping.
C. Generating function approach: The generating-function approach yields PG(p) ∼ (p − pc)^βperc, with βperc = 1 for homogeneous networks.For heterogeneous networks with P(k) ∼ k−γ, the percolation threshold tends to zero when γ < 3 as N →∞.
C. Generating function approach: For P(k) ∼ k−γ, βperc equals 1/(3 −γ) for γ < 3, 1/(γ −3) for 3 < γ ≤4, and 1 for γ ≥4.At γ = 3, PG(p) follows a stretched exponential form, PG(p) ∼ e1/p.
A. Susceptible-Infected-Susceptible model: The SIS model reaches a stationary state, enabling theoretical analyses ranging from approximate mean-field theories to exact methods.The review presents SIS results as a major body of work on epidemics in complex networks.
1. Degree-based mean-field theory: Degree-based mean-field theory assumes statistical equivalence among nodes of degree k and describes SIS dynamics through the infection probability ρI_k(t).Its dynamical equation uses mass action, recovery, susceptible-node infection, conditional neighbor degree P(k′|k), and spreading rate λ = β/µ.
1. Degree-based mean-field theory: The endemic phase begins when −1+λΛM > 0, where ΛM is the largest eigenvalue of the connectivity matrix.For general degree correlations, the DBMF equations lack a closed-form solution, but linear stability analysis gives the epidemic threshold.
1. Degree-based mean-field theory: In uncorrelated networks, the threshold condition depends on ⟨k2⟩/⟨k⟩, and for power-law networks with 2 < γ ≤3 it tends asymptotically to zero as network size becomes infinite.Higher-degree nodes have higher infection probability, showing how degree heterogeneity affects epidemic spreading.
1. Degree-based mean-field theory: For SIS prevalence, the critical exponent is 1/(3 −γ) for γ < 3, 1/(γ −3) for 3 < γ ≤4, and 1 for γ ≥4.At γ = 3, prevalence follows ρI(λ) ∼e−1/(mλ); for 2 < γ ≤3, prevalence grows very slowly above the vanishing threshold.

2. Individual-based mean-field theory

Individual-based mean-field theory closes the SIS dynamics by assuming neighboring-node independence, yielding a tractable node-level description and an eigenvalue-based epidemic threshold. Its predictions include threshold scaling and near-threshold prevalence, but their endemic interpretation depends on principal-eigenvector localization and the approximation is not exact.

Individual-based mean-field theory: The exact SIS dynamics track each node’s infection indicator, but the resulting equations depend on two-node expectations and lack an explicit closed solution.The state uses Bernoulli variables Xi(t), with E[Xi]=Pr[Xi=1]≡ρ_i^I; closure requires joint probabilities.
Individual-based mean-field theory: IBMF, also called QMF or NIMFA, closes the SIS equations by assuming that neighboring nodes are statistically independent.Under this approximation, E[Xi(t)Xj(t)] is replaced by E[Xi(t)]E[Xj(t)].
Threshold prediction: The IBMF epidemic threshold is determined by the largest eigenvalue Λ1 of the adjacency matrix through the linear stability of the disease-free state.Linearization gives Jij=−δij+λaij, and an endemic state appears when the largest eigenvalue of J is positive.
Threshold prediction: For power-law networks, the predicted threshold scales as 1/√kmax for γ>5/2 and as ⟨k⟩/⟨k2⟩ for 2<γ<5/2.Because kmax grows with network size in essentially all random, non-regular networks, IBMF predicts a vanishing thermodynamic-limit threshold, though the expression is not exact.
Near-threshold prevalence: Above threshold, IBMF predicts stationary prevalence through the principal eigenvector, but the transition is endemic only for γ<5/2 when that eigenvector is delocalized.For γ>5/2, localization on a few nodes means λc^IBMF marks an active-state transition without a finite infected fraction in the large-network limit.

3. Extensions of degree-based and individual-based mean-field approaches

Extensions of degree-based and individual-based mean-field theories incorporate dynamical correlations through higher-order node-state variables, correlation hierarchies, or reinfection mechanisms. These extensions improve prevalence and threshold estimates but increase computational complexity and require closure or approximation.

Degree-based approaches: Extended degree-based theories represent pairs and triples of neighboring nodes to account for dynamical correlations neglected by standard mean-field approaches.Approximating triples using pairs and single-node averages reduces the dynamics to O(k_max^2) nonlinear ordinary differential equations.
Degree-based approaches: Gleeson’s binary-state theory explicitly models correlations between adjacent nodes and can determine prevalence and epidemic thresholds.It describes prevalence evolution well and gives good threshold estimates for random regular lattices, but complex networks require numerically solving large coupled equation sets.
Individual-based approaches: Individual-based correlation expansions produce an exact hierarchy in which order-n correlations depend on order-(n+1) correlations, yielding 2^N − 1 SIS equations.Practical computation therefore requires truncating the hierarchy with a closure condition.
Individual-based approaches: Higher-order closures retain more dynamical correlations and improve system descriptions, while the simplest factorized closure recovers the individual-based mean-field approximation.Closure choices trade accuracy against tractability; one particular closure yields an epidemic-threshold expression based on the largest eigenvalue of a new Jacobian matrix.
Alternative threshold approach: A separate method approximately maps fixed-infection-time SIS dynamics to percolation while incorporating the reinfection probability π.The mapping follows the strategy used for SIR dynamics but is approximate for SIS processes.

4. Exact results · 5. Numerical simulations of the SIS model on networks

Exact analyses provide rigorous SIS threshold bounds, extinction and survival-time results, and critical exponents, while simulations compare approximations and estimate thresholds using quasi-stationary methods.

4. Exact results: The epidemic dies out exponentially when λΛ1 −1 ≤0, yielding a rigorous threshold lower bound that coincides with the IBMF result.The fastest-growing mode is determined by the largest eigenvalue Λ1 of the non-negative adjacency matrix A.
4. Exact results: When λ < 1 Λ1, the average absorption time is bounded, whereas above threshold it scales as E[T] = O(ecN) for some c > 0.For power-law graphs, E[T] > O(ebN 1−δ) for any δ > 0, with δ > 0.
4. Exact results: The complete graph has an exact SIS survival time for all λ and N, and its maximum lifetime bounds the survival time of epidemics on any N-node network.The complete graph permits the longest infection survival among networks with N nodes.
4. Exact results: For power-law graphs, the critical exponent obeys γ −1 ≤βSIS ≤2γ −3, ruling out βSIS = 1 for any γ > 2.The same bounds imply failure of the DBMF prediction in Eq. (29).
5. Numerical simulations of the SIS model on networks: For heavy-tailed networks, IBMF and DBMF agree for γ < 5/2 but differ for γ > 5/2, especially γ > 3, where DBMF predicts a finite threshold.The supplied passage states that IBMF indicates a vanishing threshold in this region, but the sentence is truncated.
5. Numerical simulations of the SIS model on networks: Numerical studies determine SIS thresholds on power-law networks to assess competing theoretical approaches, especially where IBMF and DBMF predictions diverge.The computational literature focuses mainly on numerical epidemic-threshold determination.
5. Numerical simulations of the SIS model on networks: The surviving-runs method averages infection density only over active realizations, but becomes inefficient near threshold because long-lived surviving configurations are rare.The quasi-stationary method instead replaces attempted absorption with an active configuration sampled from the simulation history.
5. Numerical simulations of the SIS model on networks: Susceptibility peaks at λp(N), with λp(N) −λc(∞) ∼ N −1/¯ν, while finite-realization lifetimes provide another susceptibility-like peak for threshold estimation.Large-scale quasi-stationary simulations use these peaks to locate the transition in finite systems.

6. Finite size effects and the epidemic threshold · B. Susceptible-Infected-Removed model

Finite network size and connectivity limits restore a nonzero epidemic threshold in realistic heterogeneous networks, although homogeneous approaches can substantially overestimate it. For the SIR model, degree-based mean-field analyses relate the threshold to connectivity eigenvalues and degree moments, while also describing prevalence and early outbreak growth.

6. Finite size effects and the epidemic threshold: In finite real-world networks, the network size N limits the very high degrees that can make the infinite-network epidemic threshold vanish.Heavy-tailed degree distributions may predict a vanishing threshold in the infinite-size limit, but real networks contain a finite number of nodes.
6. Finite size effects and the epidemic threshold: An exponential cutoff, P(k) ≃ k^-γ exp(−k/k_c), models the intrinsic limit on node connectivity in heavy-tailed networks.The cutoff makes nodes with degree much larger than k_c extremely unlikely.
6. Finite size effects and the epidemic threshold: Homogeneous approaches can overestimate the actual epidemic threshold in heterogeneous networks by one or more orders of magnitude.Finite-size effects do not justify neglecting network heterogeneity.
B. Susceptible-Infected-Removed model: The SIR model describes diseases that confer host immunity and is also used for knowledge and information diffusion, but lacks a stationary state.Its theoretical analysis commonly uses degree-based mean-field theory and mappings to related processes.
B. Susceptible-Infected-Removed model: For correlated networks, the SIR epidemic threshold is the inverse of the largest connectivity-matrix eigenvalue, while uncorrelated networks give λ_c = ⟨k⟩/⟨k^2⟩.The degree-based mean-field solution also recovers λ_c = 1/⟨k⟩ for regular networks.
B. Susceptible-Infected-Removed model: Near the threshold, SIR prevalence scales as ρ_R∞ ∼ (λ − λ_c)^β_SIR, with β_SIR equal to the bond-percolation exponent.These results are exact for annealed networks whose topology changes rapidly while preserving P(k).
B. Susceptible-Infected-Removed model: Accounting for the non-susceptible source neighbor yields λ_c = 1/˜Λ_M and distinguishes vanishing thresholds in scale-free networks from finite thresholds for γ > 3.For uncorrelated networks, the modified connectivity matrix has largest eigenvalue ˜Λ_M = ⟨k^2⟩/⟨k⟩ − 1; the approximation captures the correct qualitative behavior.
B. Susceptible-Infected-Removed model: At very short times, infected density grows exponentially as ρ_I(t) ∼ e^(t/τ), with τ = (β˜Λ_M)^−1; growth becomes faster as ⟨k^2⟩ increases.When ⟨k^2⟩ diverges in scale-free networks, both the threshold and the time to establish infection tend to vanish in the thermodynamic limit.

2. Individual and pair-based mean-field approaches … A. Efficient immunization protocols

The paper develops mean-field, alternative, and percolation-based frameworks for SIR epidemics, then uses network structure to design immunization strategies that raise epidemic thresholds and reduce prevalence.

2. Individual and pair-based mean-field approaches: The SIR master equation yields 2N microscopic-state probability equations, closed either by assuming neighbor independence or by using pair-based dynamics.Pair-based closure introduces neighbor-pair probabilities whose evolution depends on triples.
3. Other approaches: Extended degree-based approaches provide closed ODEs for SIR dynamics, while Volz’s formulation uses only 3 coupled nonlinear ODEs and is exact in the thermodynamic limit.The Volz solution agrees excellently with numerical simulations.
3. Other approaches: Message passing exactly describes SIR dynamics on trees through closed integro-differential equations, while providing rigorous bounds on outbreak size when loops are present.It calculates node probabilities of being susceptible, infected, or removed at any time.
4. Mapping the SIR model to a percolation process: SIR outbreaks with uniform transmissibility map exactly to bond-percolation clusters, so late-time epidemic properties and tree-like-network thresholds follow from percolation geometry.The correspondence identifies removed nodes with the percolation cluster containing the initial seed.
4. Mapping the SIR model to a percolation process: For scale-free networks with γ < 3, the epidemic threshold vanishes because the degree-distribution second moment diverges, making such networks highly vulnerable to disease spreading.This contrasts qualitatively with scale-rich networks having γ > 3.
A. Efficient immunization protocols: Random immunization raises epidemic thresholds too slowly in heterogeneous networks; when ⟨k2⟩ diverges, gc(λ) tends to 1 as N →∞.Thus, suppressing disease requires immunizing almost the whole network under random selection.
A. Efficient immunization protocols: Targeted immunization is substantially more efficient: for scale-free networks with γ = 3, gc(λ) ≃exp[−2/(mλ)], whereas proportional immunization gives gc(λ) ≃(mλ)2/3.The targeted threshold is exponentially small over a large range of spreading rate λ.
A. Efficient immunization protocols: When global network information is unavailable, acquaintance and other local strategies select neighbors, high-degree nodes, or infected contacts to reduce prevalence and increase epidemic thresholds.Percolation-based vaccination further targets GSCC nodes to reduce both major-outbreak probability and outbreak size.

B. Relevant spreaders and activation mechanisms … 1. Non-Markovian epidemics on networks

The section examines how network structure and epidemic timing shape spreading, identifying influential nodes and activation mechanisms while extending Markovian models to realistic non-Poissonian infection and recovery processes.

B. Relevant spreaders and activation mechanisms: Choosing an initial seed to maximize the eventual outbreak size motivates the search for influential or effective spreaders.
B. Relevant spreaders and activation mechanisms: The K-core index can predict final SIR outbreak size better than degree in several real networks, challenging degree-based superspreader identification.
B. Relevant spreaders and activation mechanisms: No unique best centrality measure has emerged because spreading influence varies across epidemic phases, models, and network types.
B. Relevant spreaders and activation mechanisms: Efficient spreaders are not necessarily good immunization targets: removing high-K-core nodes can have limited effects because multiple paths connect central-core nodes.
B. Relevant spreaders and activation mechanisms: Friends of randomly selected individuals can identify early epidemic sensors without global network knowledge, selecting people with high degree, high betweenness, small clustering, and high K-core index.
B. Relevant spreaders and activation mechanisms: For SIS, the IBMF epidemic threshold scales as the inverse of the largest adjacency-matrix eigenvalue Λ1, linking global activation to network subsets.
A. Realistic models: Realistic epidemic models add compartments or transitions, while non-Poissonian formulations replace memoryless constant-rate dynamics with distributed infectious and transmission times.
1. Non-Markovian epidemics on networks: Message passing gives exact integro-differential equations on trees and locally tree-like networks, while renewal theory preserves the SIS prevalence form after replacing λ = β/µ with average infection attempts during recovery.

2. The SIRS model … 3. Weighted networks

The reviewed models show how epidemic dynamics depend on recovery and exposure stages, realistic network topology, clustering, and heterogeneous contact weights. Analytical mappings and spectral or percolation methods clarify thresholds, while numerical studies expose synchronization, outbreak-size, and spreading-rate effects.

2. The SIRS model: Within DBMF, SIRS has the same critical properties as SIS, with immunity decay η only rescaling infected density.This follows from an exact mapping between the steady-state SIRS and SIS solutions.
2. The SIRS model: SIRS extensions incorporate death, birth, simple recovery, and susceptible death, recovering a threshold inversely proportional to the relevant network quantity.Liu et al. apply a DBMF formalism to transitions I → R, R → S, I → S with rate γ, and S → R with rate α.
3. The SEIR model: SEIR models influenza-like and respiratory infections, including SARS studies using deterministic and stochastic network models calibrated with empirical spreading data.Edge-based compartmental modelling also accommodates multiple infectious stages, with SEIR as a particular case.
B. Realistic static networks: Degree correlations complicate epidemic analysis, but for SIS the IBMF threshold remains the inverse of the adjacency matrix’s largest eigenvalue Λ1.Assortativity increases Λ1 and reduces the threshold, whereas disassortativity decreases Λ1.
B. Realistic static networks: For SIR, assortative correlations can produce a vanishing threshold even with finite second degree moment, while branching-matrix analysis extends to SEIR.Degree correlations do not change critical exponents when the stated branching-matrix eigenvalue and eigenvector conditions hold.
2. Effects of clustering: Increasing clustering decreases the SIS epidemic threshold, while clustered-network outbreak sizes decrease with clustering and saturate at large C as transmissibility rises.These conclusions derive from adjacency-matrix bounds and percolation-based analyses of clustered networks.
3. Weighted networks: In weighted networks, IBMF replaces the adjacency matrix with Ωij = ωijaij, making the SIS threshold the inverse of Ω’s largest eigenvalue.Weight distributions can concentrate infectiousness on fewer links and increase the epidemic threshold; weighted spreading can also slow and shift from exponential to slow power-law behavior.

4. Directed networks · 5. Bipartite networks

Directed-network epidemics depend on in-/out-degree structure, component topology, and directionality, while bipartite-network thresholds couple transmission across the two node types. These frameworks show that directional or cross-type constraints can substantially alter outbreak probabilities and endemicity.

4. Directed networks: Directed transmission models capture epidemics in systems where propagation has intrinsic directionality, such as blood transfusions or needle sharing.A node’s network component can restrict or enhance its spreading capabilities.
4. Directed networks: In directed SIR networks, percolation analysis depends on the joint in-degree/out-degree distribution P(k_in, k_out), including correlations between the two degrees.These correlations influence epidemic behavior in the giant weakly connected component.
4. Directed networks: The supercritical SIR outbreak exponent uses the effective exponent γ*=γ_out+(γ_in−γ_out)/(γ_in−1).This result applies with the effective degree distribution described for epidemics in the giant weakly connected component.
4. Directed networks: In directed networks, outbreak probability and expected affected fraction differ, and either can be larger depending on topology.For SIS dynamics, the individual-based mean-field result matches undirected networks because adjacency asymmetry does not explicitly enter the theory.
4. Directed networks: For directed SIR dynamics, the general threshold depends on the largest eigenvalue of an extended connectivity matrix and reduces to Eq. (73) without degree-degree correlations.The degree classes are defined by (k_in, k_out), with conditional neighbor-degree probabilities describing outgoing connections.
5. Bipartite networks: Bipartite networks model transmission between two distinct node types, including sexually transmitted and vector-borne diseases.Examples include male-female contacts and vector-host transmission.
5. Bipartite networks: For bipartite SIR and SIS models, epidemic thresholds form hyperbolas coupling the two types’ transmissibilities or spreading rates.The SIR formulation uses partial degree distributions P_m(k) and P_f(k), while the SIS threshold is expressed through λ_m and λ_f.
5. Bipartite networks: In bipartite SIR dynamics, when λ_f diverges, the threshold value for λ_m remains finite, whereas in SIS dynamics its asymptotic value is λ_m=0.Thus, reducing the spreading rate of one node type can rule out an endemic outbreak.

6. Effect of other topological features … VIII. EPIDEMIC PROCESSES IN TEMPORAL NETWORKS

The paper shows that modularity, adaptive rewiring, competing pathogens, and temporal contact patterns substantially alter epidemic thresholds, propagation, phase diagrams, and immunization strategies. These effects depend on network structure, coevolution, cross-immunity, temporal ordering, burstiness, and the relative time scales of contacts and contagion.

6. Effect of other topological features: Community structure can slow SI infection growth by confining outbreaks near their initial seeds and hindering transmission between communities.This effect is further enhanced in weighted social networks through correlations between topology and weights.
6. Effect of other topological features: SIS dynamics on complex networks can exhibit Griffiths phases, with absorbing stationary states reached through anomalously long, nonuniversal relaxation linked to rare-region effects.On loopless weighted networks, these effects have been related to localization properties of the largest adjacency-matrix eigenvalue.
7. Epidemics in adaptive networks: Adaptive rewiring hinders disease propagation by increasing the epidemic threshold and can produce bistability, discontinuous transitions, and hysteresis.Without rewiring, only a single continuous transition occurs for pc ≈0.0001; rewiring instead produces richer phase diagrams.
7. Epidemics in adaptive networks: For adaptive SIR dynamics, coevolutionary effects are weaker because the absorbing state is reached in a time logarithmic in system size N, leaving the global topology weakly perturbed.The phase diagram remains qualitatively similar to the nonadaptive case, with a single epidemic transition.
C. Competing pathogens: Competing epidemic processes can coexist or interact through cross-immunity and coinfection, depending on pathogen order, network topology, and model parameters.With total cross-immunity, a second SIR pathogen spreads only among survivors of the first; competing SIS processes can also reach steady states with finite prevalence for both pathogens.
VIII. EPIDEMIC PROCESSES IN TEMPORAL NETWORKS: Temporal networks require contact-sequence representations because aggregated static projections can be inappropriate for processes unfolding on time-varying connectivity patterns.Temporal ordering limits possible epidemic paths, while heterogeneous contact patterns commonly exhibit heavy-tailed, skewed distributions and burstiness.
VIII. EPIDEMIC PROCESSES IN TEMPORAL NETWORKS: n(t) ∼t−β in sparse integrated email-contact networks, implying substantially slower epidemic growth than the exponential increase found in regular static networks.This slowing follows from heavy-tailed inter-event times and is reported for SI spreading models of email worms.
VIII. EPIDEMIC PROCESSES IN TEMPORAL NETWORKS: Temporal-network dynamics alter epidemic thresholds: long-time-scale epidemics can have vanishing thresholds on scale-free integrated networks, whereas short-time-scale epidemics can have finite thresholds.In activity-driven networks, the SIS threshold is independent of integration time, and immunization saturates when the training window covers about a 20% - 40% of the contact sequence.

IX. REACTION-DIFFUSION PROCESSES AND METAPOPULATION MODELS · A. SIS model in metapopulation networks

Reaction-diffusion and metapopulation models represent populations distributed across heterogeneous networks, combining particle mobility with local reactions. In SIS metapopulations, epidemic thresholds depend strongly on the infection law and network topology, while realistic analyses must address heterogeneous mobility.

IX. REACTION-DIFFUSION PROCESSES AND METAPOPULATION MODELS: Reaction-diffusion networks model nodes as subpopulations hosting variable numbers of particles that diffuse along edges and react locally according to specified interaction schemes.This framework represents spatially structured systems such as cities, urban areas, or geographical regions.
IX. REACTION-DIFFUSION PROCESSES AND METAPOPULATION MODELS: Degree-based mean-field analysis assumes statistical equivalence among nodes of equal degree, reducing particle diffusion to dynamical equations for degree-block populations.The approach extends heterogeneous-network mean-field methods to reaction-diffusion systems with arbitrary degree distributions.
IX. REACTION-DIFFUSION PROCESSES AND METAPOPULATION MODELS: In homogeneous diffusion on uncorrelated networks, stationary particle occupation depends on degree, showing that higher-degree nodes are more likely to be visited by diffusing particles.The diffusion rate per link is specified as r/k′, with degree correlations represented through P(k′|k).
A. SIS model in metapopulation networks: The metapopulation SIS model assigns susceptible and infectious individuals to network nodes, with infection and recovery probabilities β and µ and diffusion across neighboring subpopulations.Individuals diffuse along departing links with probability 1/k under the simplifying assumption pI = pS = 1.
A. SIS model in metapopulation networks: β/µ > 1 is required for an endemic state under the usual mass-action law, recovering the classic epidemic threshold for homogeneous systems.Here Γk = IkSk/Nk, so local epidemic activity is rescaled by the population present in each node.
A. SIS model in metapopulation networks: Under the pseudo mass-action law Γk = IkSk, an infectious stationary state requires average particle density above a topology-dependent critical threshold.This law treats each susceptible individual as potentially interacting with all infectious individuals in the same node.
A. SIS model in metapopulation networks: ¯Nc →0 as network size becomes infinite, because topological fluctuations can eliminate the epidemic threshold under pseudo mass-action dynamics.The critical density is therefore affected by network degree fluctuations rather than determined solely by β and µ.
A. SIS model in metapopulation networks: Because individual mobility is highly heterogeneous, analytical epidemic models often use Markovian movement matrices dij, with traffic-based rates satisfying dij ∼ wij/Nj.These mobility processes support large-scale infectious-disease spreading models when inter-subpopulation traffic wij is known.

B. SIR model in metapopulation networks and the global invasion threshold

The metapopulation SIR analysis introduces a global invasion threshold that captures whether infection spreads across a macroscopic fraction of subpopulations, beyond the local condition R0 > 1. Mobility and network heterogeneity jointly determine invasion, while continuous approximations may miss the discrete seeding mechanism.

Global invasion threshold: R∗ measures the number of subpopulations infected by one initially infected subpopulation and defines the conditions for macroscopic spread.It is the subpopulation-level analogue of the basic reproductive number R0.
Branching-process formulation: The spreading process is modeled as a branching process in which infected subpopulations seed neighboring subpopulations across successive generations.The number of infected degree-k subpopulations is tracked as Dn_k at generation n.
Global invasion threshold: A growing epidemic requires the derived global threshold condition, which identifies when each subpopulation can seed more than one neighboring subpopulation on average.The threshold is obtained under near-local-threshold, early-stage, and uncorrelated-network assumptions.
Mobility threshold: For R0 > 1, a minimum mobility rate is required for global invasion, and in the SIR case near R0 = 1 the approximation ᾱ ≃2(R0 −1)/R2 0 yields a critical mobility pc.Below pc, the epidemic cannot invade the metapopulation system.
Global perspective: The invasion threshold R∗ > 1 shifts assessment from local outbreaks to global spreading by emphasizing interconnectivity, mobility, and network heterogeneity.The explicit expression includes the factor ⟨k⟩2/⟨k2⟩.

C. Agent Based Models and Network Epidemiology · X. GENERALIZING EPIDEMIC MODELS AS SOCIAL CONTAGION PROCESSES · A. Threshold models

The review connects data-driven agent-based network models for forecasting infectious disease outbreaks with epidemic-inspired analyses of social contagion. It then develops threshold-model results showing how exposure dependence, network structure, and initiator density govern whether cascades remain local or become global.

C. Agent Based Models and Network Epidemiology: Agent-based, spatially structured epidemic models provide quantitative forecasts and scenario analyses at detailed spatial resolutions.These approaches have evolved into large-scale microsimulations using data-driven representations of real outbreaks.
C. Agent Based Models and Network Epidemiology: GLEAM integrates census and mobility data into a stochastic metapopulation network to simulate influenza-like illness spreading globally.It partitions populations into geographic census areas connected by mobility fluxes around transportation hubs.
C. Agent Based Models and Network Epidemiology: Data-driven computational models support detailed epidemic analysis, forecasting, and policy-making scenario evaluation.The review emphasizes that understanding reaction-diffusion processes in complex networks remains crucial for assessing these models’ reliability and predictive power.
X. GENERALIZING EPIDEMIC MODELS AS SOCIAL CONTAGION PROCESSES: Epidemic models also describe social contagion, including information, propaganda, and other social spreading phenomena enabled by abundant digital data.Social contagion data motivate theoretical approaches to measure, interpret, model, and predict these processes.
X. GENERALIZING EPIDEMIC MODELS AS SOCIAL CONTAGION PROCESSES: Social contagion differs qualitatively from pathogen spreading, while the review acknowledges limits imposed by its broad interdisciplinary scope and related processes beyond its coverage.These related processes include failure cascades and game-theoretic strategy adoption.
A. Threshold models: Threshold models replace independent exposure probabilities with adoption triggered when an individual’s accumulated neighbor influence reaches a personal threshold.Watts’ model assigns each agent a quenched random threshold and permits transitions from S to I without recovery.
A. Threshold models: The critical threshold is φc = 1/⟨k⟩: φ < φc permits global cascades, whereas φ > φc produces localized spreading.Adding links increases ⟨k⟩, lowers φc, and makes system-wide spreading more difficult in this threshold model.
A. Threshold models: Threshold-model behavior changes with exposure memory, dose variability, recovery, initiator density, clustering, and network organization.Finite initiator fractions can shift or make transitions discontinuous, while clustering reduces cascades at large and small ⟨k⟩ but increases them at intermediate ⟨k⟩.

B. Rumor spreading · C. Empirical studies · XI. OUTLOOK

The section examines rumor dynamics, empirical spreading processes, and future challenges involving adaptive, coupled, and increasingly detailed network models. It emphasizes how interaction rules, network structure, and social behavior shape contagion and motivate interdisciplinary research.

B. Rumor spreading: Rumor models modify SIR dynamics so spreaders stop transmitting through interactions with agents who already know the rumor.The Daley-Kendall and Maki-Thompson formulations differ in which interacting individuals become stiflers.
B. Rumor spreading: Reliability r∞ measures the asymptotic stifler fraction, distinguishing localized rumors from macroscopic spreading; on complete graphs, r∞ is positive for every β/α > 0.Unlike homogeneous-network SIR dynamics, rumor spreading therefore has no finite threshold in these models.
B. Rumor spreading: Network topology changes rumor outcomes: scale-free heterogeneity lowers final reliability without creating a finite threshold, while Watts-Strogatz networks exhibit a finite rewiring threshold for macroscopic spreading.High-K-core initiators also do not determine the final stifler density in the Maki-Thompson model.
C. Empirical studies: Retail email recommendations generally produce short purchase cascades because infection probability saturates quickly at a low value and success per recommendation declines with sender activity.The passage characterizes viral marketing as substantially different from epidemic-like spreading.
C. Empirical studies: Internet chain-letter cascades are deep and tree-like, whereas simple epidemic models produce wide, shallow trees; asynchronous response and additional mechanisms yield more realistic propagation.Newsletter recommendation cascades from 7154 initiators were also essentially tree-like.
C. Empirical studies: Empirical social contagion studies report spreading of obesity, smoking, and happiness, while controlled online-community experiments examine how social influence favors propagation.The cited claims are accompanied by criticism concerning the interpretation of such effects.
C. Empirical studies: Online-network data reveal that scale-free contact structure can coexist with lognormally distributed cascades, with nearly all Digg propagations reaching less than 1% of the network.Twitter meme variability can be reproduced using heterogeneous follower graphs and limited attention caused by finite memory and competition.
XI. OUTLOOK: Future epidemic-network research must explain co-evolutionary feedback, competing contagions, interdependent and multilayer networks, and nonlinear behavior in detailed large-scale population models.Progress requires combining large-scale data mining, computational methods, analytical techniques, and interdisciplinary collaboration.