Block-Fitness Modeling of the Global Air Mobility Network

TL;DR

The paper tackles the lack of high-resolution, near-real-time WAN data by introducing a block-fitness maximum-entropy model that uses airport strengths and geographic community structure to generate scalable, sparse surrogates of the WAN. It formulates a probabilistic, two-step generative process where inter-node connections are drawn with probabilities $p_{ij}$ depending on block-specific fitness and, optionally, distance, and then weighted to reproduce node strengths, yielding ensemble networks with preserved flows. Among several variants, the block-based model (B) provides the best reconstruction of topology, weights, and inter-regional structure, and produces spreading dynamics in metapopulation simulations that closely match those on the empirical WAN, including spatial heterogeneity and timing. The framework is interpretable and scalable, offering a data-efficient tool for mobility forecasting and policy analysis with applications to air, sea, and trade networks when fine-grained data are scarce.

Abstract

Accurate representations of the World Air Transportation Network (WAN) are fundamental inputs to models of global mobility, epidemic risk, and infrastructure planning. However, high-resolution, real-time data on the WAN are largely commercial and proprietary, therefore often inaccessible to the research community. Here we introduce a generative model of the WAN that treats air travel as a stochastic process within a maximum-entropy framework. The model uses airport-level passenger flows to probabilistically generate connections while preserving traffic volumes across geographic regions. The resulting reconstructed networks reproduce key structural properties of the WAN and enable simulations of dynamic spreading that closely match those obtained using the real network. Our approach provides a scalable, interpretable, and computationally efficient framework for forecasting and policy design in global mobility systems.

Figures (14)

• Figure 1: Topological features of the WAN. (A) Degree distribution with power-law fit $P(k)\sim k^{-\gamma_k}$ and exponent $\gamma_k=1.8(3)$, and strength distribution with power-law fit $P(s)\sim s^{-\gamma_s}$ and exponent $\gamma_s=1.(7)$. (B) Degree-strength relation, with OLS fit $s\sim k^\beta$ and exponent $\beta=1.2(2)$. (C) Link probability as a function of geographic distance. The curve is fitted with a linear function $l(d) = m\cdot d+q$ up to a characteristic distance $d^*=50km$ and then with an exponential $l(d) = \exp[-a(d-d_0)]+b$. Best fit parameters are $m=0.003$, $q=-0.0013$ for the linear term and $d_0=-2.(9)\cdot10^3$, $a=7.(1)\cdot10^{-4}$, $b=-2.(4)\cdot10^{-3}$ for the exponential decay ($R^2=0.99$). (D) World partition into regions: Africa, Asia, Central America, East Asia, Europe, Europe/Asia, Middle East, North America, Oceania, South America, Southeast Asia. (E) Community structure in the WAN adjacency matrix induced by the world partition, with modularity $M=0.67$.
• Figure 2: Reconstruction of the topological properties of the WAN. (A): Model degrees $\langle k_i\rangle = \sum_{j\neq i} p_{ij}$ versus empirical degrees $k_i$, with Root Mean Squared Percentage Error (RMSRE) equal to 0.18. (B): Empirical link probability (given by the fraction $f_{ij}$ of $ij$ pairs that are actually connected in each bin of $p_{ij}$) as a function of the model link probability $p_{ij}$, with RMSRE = 0.016. (C): Average model weights $\langle w\rangle_{ij}$ versus empirical weights $w_{ij}$, with RMSRE = 0.063. For this exercise, we limit the test set to links $w_{ij}>0$ that are actually realized. For (A,B,C), the dashed identity line marks the perfect agreement between empirical and model values, while the shaded regions represent the inter-quartile range and the whiskers of the set of model values. (D) Degrees vs Strengths relation, with OLS fit exponent $\beta=1.2(2)$ on empirical data and $\beta=1.17(0)$ on model data. (E) Link probability as a function of distance, with Mean Squared Error (MSE) between empirical and model curve of 2.21. (F) Block structure of a sample reconstructed network. Modularity is 0.63 on average over the model ensemble (to be compared with $M=0.67$ for real data).
• Figure 3: Epidemic simulations starting from London. Prevalence map at simulation step $t=80$ for the empirical (A) and model (B) WAN. Time evolution of the prevalence (total fraction of infected individuals) (C) and of the entropy of the prevalence distribution (D), for one real simulation and 20 model simulations.
• Figure S4: For models B and C, the parameter $z_{g_ig_j}$ for each pair of communities ($g_ig_j$) is determined by equating the total number of links of the model network, namely $\sum_{i\in g_i}\sum_{j(\neq i)\in g_j}p_{i\to j}$, with the number of links of the real network $L_{g_ig_j}={i\in g_i}\sum_{j(\neq i)\in g_j}a_{i\to j}$. Here we report the specific equation for each model and the heat maps with the corresponding values of $z$ for each pair of regions.
• Figure S5: We obtain the optimal $\alpha$ exponent for models S and C by minimizing the RMSD between the real and reconstructed degree sequence as a function of $\alpha$. Basins dataset (upper plots): for S (left plot) we fitted the curve with a 3rd degree polynomial, yielding the minimum value $\alpha=0.80$; for C (right plot) we used a 10th degree polynomial, yielding $\alpha=1.56$. Airport dataset (lower plots): for S (left plot), $\alpha=0.72$; for C (right plot), $\alpha=1.16$.