Learning graph topology from metapopulation epidemic encoder-decoder

Xin Li; Jonathan Cohen; Shai Pilosof; Rami Puzis

cs.LGMar 2, 2026

Learning graph topology from metapopulation epidemic encoder-decoder

Authors: Xin Li, Jonathan Cohen, Shai Pilosof, Rami Puzis

Organizations: Faculty of Computer and Information Science, Ben-Gurion University of the Negev, Be’er Sheva, South District, Israel · Department of Life Sciences, Ben-Gurion University of the Negev, Be’er Sheva, South District, Israel · The Goldman Sonnenfeldt School of Sustainability and Climate Change, Ben-Gurion University of the Negev, Be’er Sheva, South District, Israel

Abstract

Metapopulation epidemic models are a valuable tool for studying large-scale outbreaks. With the limited availability of epidemic tracing data, it is challenging to infer the essential constituents of these models, namely, the epidemic parameters and the relevant mobility network between subpopulations. Either one of these constituents can be estimated while assuming the other; however, the problem of their joint inference has not yet been solved. Here, we propose two encoder-decoder deep learning architectures that infer metapopulation mobility graphs from time-series data, with and without the assumption of epidemic model parameters. Evaluation across diverse random and empirical mobility networks shows that the proposed approach outperforms the state-of-the-art topology inference. Further, we show that topology inference improves dramatically with data on additional pathogens. Our study establishes a robust framework for simultaneously inferring epidemic parameters and topology, addressing a persistent gap in modeling disease propagation.

Figures & tables

Fig 1: Representative graph topologies used in this study. Top row: synthetic random graphs - er ( er ), ba ( ba ), ws ( ws ), rgg ( rgg ).

Panel of Figure 2 — Fig 2: Left: A metapopulation epidemic process showing the infection both within and among the populations. The blue blocks indicate the state evolution of subpopulation 1. Right: An example of one pathogen metapopulation epidemic time-series data $\Delta\bar{S}$ in typical ba and rgg graphs with 100 nodes, as well as empirical contiguous US [ 35 ] , where the nodes are geometric states (two nodes establish a link if they share a border), and global air transportation graphs, which represent the commute between two airports. The x-axis represents the number of days since the start of the epidemic. The nodes are ordered along the y-axis according to the distance from the seed node. The color indicates the fraction of daily new cases $\Delta\bar{S}$ .

Fig 3: Architecture of the self-supervised DTEF model. The input is the daily new infection data $\Delta\bar{S}$ . The encoder extracts node–pathogen embeddings and infers the infection matrix $\hat{Z}$ , and the decoder uses $\hat{Z}$ to reconstruct the predicted daily new infections $\Delta\hat{{S}}$ . Green ovals and rectangles denote modules with learnable parameters, blue ovals denote parameter-free operations, the black box in $\sigma$ is the scalar output of $F$ transform operation, $\sigma$ is the sigmoid activation enforcing $\hat{Z}_{ij}\in[0,1]$ (red cell), arrows indicate input sequences, and the highlighted red cell shows a specific reconstructed entry of $\Delta\hat{S}$ .

Name	Possible values
$Graph$	er , ba , ws , rgg , or real graphs
$RandomSeed$	Integer
$n$	Positive integer
$Number\ of\ Pathogen$	Positive integer
$DLModel$	dtef , ftef , Infer2018
$Dense$	Positive integer

Table 1: Configurable parameter settings for training.

Fig 4: The loss (Eq. ( 18 )) as a function of the number of optimization epochs (left); four evaluation metrics as a function of the number of optimization epochs (right).

	Spectral similarity				Pearson correlation
	DTEF	FTEF	Infer2018	RND	DTEF	FTEF	Infer2018	RND
ER	0.17	0.26	0.11	0.03	0.62	0.02	0.03	0
BA	0.22	0.31	0.14	0.05	0.72	0.03	0.03	0
WS	0.74	0.55	0.1	0.1	0.88	0.14	0.12	0
RGG	0.82	0.62	0.44	0.09	0.95	0.18	0.22	0
Contiguous US	0.8	0.55	0.4	0.01	0.98	0.14	0.17	0.01

Table 2: Topology reconstruction benchmarks on representative random and real graphs. This experiment is performed using four pathogens. We show the performance of the five fitting models and a random Baseline (RND), based on four evaluation metrics as described in the text. The absence of PR-AUC values for Mobility US is due to them being weighted graphs. The bottom-line win rate represents the percentage of tasks where the model achieves a higher accuracy score than the other models.

	Spectral similarity				Pearson correlation
	DTEF	FTEF	Infer2018	RND	DTEF	FTEF	Infer2018	RND
ER	0.16	0.28	0.09	0.03	0.61	0.65	0.57	0
BA	0.14	0.06	0.41	0.05	0.75	0.77	0.66	0
WS	0.8	0.71	0.75	0.1	0.92	0.94	0.91	0
RGG	0.8	0.83	0.76	0.09	0.97	0.93	0.85	0
Contiguous US	0.97	0.9	0.83	0.01	0.99	1.0	0.98	0.01

Table 3: Topology reconstruction benchmarks on representative random and real graphs with epidemic parameters set to their ground truth values. Other settings are the same as table 2 .

	RMSE( $\beta$ )					RMSE(1/ $\gamma$ )
—	DTEF	FTEF	Infer2018	Bayes	NLS	DTEF	FTEF	Infer2018	Bayes	NLS
ER	0	1.05	0.69	0.02	0.01	0.01	1.79	3.01	1.49	0
BA	0	1.04	0.55	0.02	0.01	0.02	1.84	2.76	1.30	0
WS	0	1.01	0.62	0.02	0.03	0.01	0.4	2.57	1.79	0
RGG	0	0.97	0.57	0.02	0.01	0.02	0.5	6.84	1.30	0
Contiguous US	0	0.16	0.21	0	0.01	0	0.12	6.93	0.47	0

Table 4: Epidemic parameter estimations of

\hat{\beta}

(upper Table) and

\hat{\gamma}

(lower Table) on both synthetic random and real-world graphs. The performance of our different model configurations is evaluated based on the rmse metric. In addition, the results of Infer2018, Bayes and NLS are shown for comparison (see text).

Fig 5: DTEF model performance PR curves for various random graph models, with 1-4 pathogens and a random baseline.

Fig 6: Evolution of the DTEF model’s topology inference with different numbers of pathogens. Each row contains four pairs of graphs: the spring layout of the graph titled with the Jaccard similarity where the blue color represents the ground-truth graph, while the orange color is the inferred graph (left); the degree distribution histogram comparison with the x-label degree and y-label probability (right).

Fig 7: Effect of different mobility rates on the topology inference accuracy of the DTEF model.

	Nodes	Links	Weighted
Contiguous US	49	107	no
Contiguous China	33	69	no
Contiguous EU	24	38	no
Contiguous Africa	47	103	no
Mobility German	52	-	yes
Mobility US	401	-	yes

Table S1: Real graphs’ attributions, including the node number, link number, and whether the graph is weighted or not.

Fig S1: Adjacency matrix comparison between ground-truth and inferred rgg graphs containing 100 nodes with four pathogens: the real binary graph (left); the inferred adjacency topology according to the DTEF model with four pathogens of data (right).

Fig S2: Sparsity index of the mobility matrix as a function of the number of optimization epochs.

	Spectral similarity					Pearson correlation					Jaccard similarity
Avg degree	4	5	6	7	8	4	5	6	7	8	4	5	6	7	8
er	0.2	0.22	0.32	0.19	0.33	0.61	0.56	0.52	0.51	0.49	0.25	0.23	0.21	0.22	0.22
ba	0.21	0.38	0.43	0.36	0.48	0.72	0.63	0.64	0.6	0.59	0.34	0.28	0.29	0.28	0.27
ws	0.78	0.75	0.83	0.83	0.83	0.88	0.87	0.84	0.82	0.79	0.59	0.57	0.54	0.52	0.5
rgg	0.83	0.83	0.84	0.88	0.9	0.95	0.96	0.95	0.95	0.94	0.75	0.76	0.76	0.75	0.73
	PR-AUC

Table S2: Effect of different average degrees on the topology inference accuracy of the DTEF model.

Fig S3: Effect of different average degrees on the topology inference accuracy of the DTEF model.

	Spectral similarity				Pearson correlation				Jaccard similarity				PR-AUC
Nodes	50	100	200	400	50	100	200	400	50	100	200	400	50	100	200	400
er	0.18	0.2	0.12	0.21	0.85	0.61	0.46	0.35	0.52	0.25	0.14	0.07	0.82	0.56	0.37	0.22
ba	0.45	0.21	0.12	0.22	0.87	0.72	0.55	0.49	0.56	0.34	0.19	0.13	0.82	0.69	0.48	0.38
ws	0.76	0.78	0.76	0.68	0.97	0.88	0.78	0.73	0.81	0.59	0.41	0.34	0.87	0.86	0.78	0.71
rgg	0.9	0.83	0.77	0.78	0.98	0.95	0.94	0.92	0.85	0.75	0.7	0.65	0.88	0.89	0.9	0.89

Table S3: Effect of different numbers of nodes on the topology inference accuracy of the DTEF model.

Fig S4: Effect of different numbers of nodes on the topology inference accuracy of the DTEF model.

	Spectral similarity									Pearson correlation
$M_{R}$	2e-4	5e-4	1e-3	2e-3	5e-3	1e-2	2e-2	5e-2	1e-1	2e-4	5e-4	1e-3	2e-3	5e-3	1e-2	2e-2	5e-2	1e-1
er	0.22	0.25	0.17	0.2	0.19	0.2	0.24	0.19	0.24	0.53	0.57	0.62	0.63	0.63	0.61	0.6	0.57	0.23
ba	0.34	0.3	0.3	0.2	0.3	0.21	0.3	0.21	0.25	0.58	0.67	0.68	0.71	0.71	0.72	0.71	0.52	0.24
ws	0.72	0.76	0.76	0.79	0.75	0.78	0.76	0.76	0.77	0.8	0.86	0.9	0.9	0.89	0.88	0.87	0.86	0.73
rgg	0.78	0.82	0.8	0.82	0.84	0.83	0.84	0.85	0.59	0.87	0.93	0.93	0.95	0.94	0.95	0.95	0.94	0.36
	Jaccard similarity									PR-AUC

Table S4: Effect of different mobility rates

M_{R}

on the topology inference accuracy of the DTEF model.

	Spectral similarity				Pearson correlation				Jaccard similarity				PR-AUC
$Pathogen$	1	2	3	4	1	2	3	4	1	2	3	4	1	2	3	4
contiguous_US	0.46	0.78	0.82	0.84	0.58	0.93	0.98	0.98	0.26	0.69	0.86	0.83	0.53	0.82	0.83	0.83
contiguous_China	0.55	0.81	0.71	0.98	0.64	0.93	0.99	1	0.3	0.69	0.9	0.96	0.57	0.82	0.83	0.83
contiguous_EU	0.17	0.35	0.63	0.58	0.7	0.96	0.99	1	0.38	0.78	0.92	0.96	0.66	0.8	0.8	0.8
contiguous_Africa	0.36	0.73	0.81	0.85	0.53	0.9	0.96	0.99	0.22	0.63	0.77	0.89	0.44	0.81	0.83	0.83
mobility_german	0.21	0.13	0.08	0.04	0.05	0.32	0.4	0.54	0	0.12	0.17	0.22	0.31	0.44	0.51	0.52

Table S5: Real-world graph evaluation results for pathogens one to four, as inferred by the DTEF model.

Fig S5: The inference accuracy of the DTEF model on real contiguous region graphs: spring layout topology (left) and degree distribution (right).

Fig S6: The inference accuracy of the DTEF model on real global plane graphs: spring layout topology (left) and degree distribution (right).

Fig S7: The inference accuracy of the DTEF model on real Spanish graphs: spring layout topology (upper Figures) and degree distribution (lower Figures).

Graph Structure Learning

Graph structure learning infers or adapts network connections from data for statistical discovery and downstream prediction.

See all 134 papers

Oct 15, 2026

Learning and Forecasting Multiscale Spatiotemporal Epidemic Dynamics with Lagrangian-Eulerian Graph Neural ODEs

Accurate medium- to long-range epidemic forecasting requires extrapolating high-dimensional, heterogeneous and nonstationary dynamics from partial observations. Yet hybrid models often leave implicit which observations inform which epidemic processes and how learned graph interactions modify system state. We introduce LE-EpiGNODE, which replaces generic message passing with process-specific, mechanism-guided neural operators: age-contact information drives Lagrangian infection generation, whereas mobility drives conservative Eulerian population transport within a structure-preserving universal differential equation. We evaluated LE-EpiGNODE across 14 controlled synthetic mechanisms spanning time-modulated, state-dependent, nonlinear and coupled dynamics, three complementary COVID-19 settings, 21 comparator models and five targeted ablations. Forecasting gains tended to become more pronounced at longer horizons; at the longest horizon, LE-EpiGNODE achieved the lowest WMAPE in all three real-world datasets, reducing error by 17.7-40.0% relative to the best baseline. Controlled ablations linked these gains to the Lagrangian-Eulerian decomposition and process-aligned priors, while synthetic ground truth showed that learned operators and induced state fluxes track known generating dynamics. These results support mechanism-organized graph neural dynamics as a route to reliable long-horizon extrapolation with directly testable, process-level mechanistic interpretation beyond trajectory accuracy.

Jinyi Duan, Qian Li, Biao Tang

School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an, 710049, People’s Republic of China . · School of Mathematics and Statistics, Xi’an University of Technology, Xi’an, 710054, People’s Republic of China . · Center for Intersection of Mathematics and Life Sciences, Xi’an Jiaotong University, Xi’an, 710049, People’s Republic of China .

Epidemic Forecasting Spatial Epidemic Modeling Spatiotemporal GNNs Long-Term Time Series Forecasting Neural ODEs Explainable Time Series Forecasting

Ask PDF

Sep 1, 2026

A Network Science Perspective on Evaluating Deep Graph Generative Models

Traditional network models from network science, such as the Erdos-Renyi and configuration models, generate random networks that reproduce few selected topological properties observed in real-world networks. Deep graph generative models emerge as a data-driven approach, leveraging deep neural network architectures to learn complex structural distributions directly from real-world networks to generate more realistic synthetic networks. Because real social contact networks cannot be shared due to privacy risks, synthetic networks serve as an alternative for developing and evaluating epidemic mitigation strategies. In this work, we evaluate deep graph generative models as well as the configuration from a network science perspective by assessing both the topological similarity between generated and real-world networks and their utility in identifying effective node immunization strategies to sup- press epidemic/misinformation spreading. It is found that two deep graph generative models produce synthetic networks that closely resemble the structural properties of real-world networks, enabling them to identify effective immunization strategies.

Tianrui Mao, Abele Malan, Megha Khosla +2

Delft University of Technology, the Netherlands · University of Neuchâtel, Switzerland

Synthetic Data Generation Generative Modeling Synthetic Data Evaluation

Ask PDF

Sep 28, 2026

Graph World Models for Constrained Epidemic Policy Planning

Epidemic policy planning often requires coordination between geographical regions, taking into account mobility-driven spillovers and how to make use of limited resources. Existing methods either lack action-conditioned models of coupled dynamics or cannot guarantee per-period feasibility. We present EpiMind, a graph world model framework for constrained epidemic policy planning across regions. A graph-factored recurrent state-space model generates joint policy-conditioned rollouts from regional latent beliefs, while graph-temporal ADMM optimizes regional interventions, enforces shared-resource feasibility through projection, and evaluates temporal specifications under the learned model. EpiMind reduces admission RMSE by 29% relative to graph-free dynamics modeling, plans within 1-5% of the best feasible constant policy with guaranteed shared-budget feasibility, and outperforms all deployable baselines across three resource budgets in real-context evaluation. These results demonstrate that graph-structured policy imagination with explicit constrained coordination supports effective epidemic interventions from learned dynamics.

Yiqi Su, Rashed Shelim, Lingyi Wang +2

Department of Computer Science, Virginia Tech, Alexandria, VA 22305, USA · Department of Electrical and Computer Engineering, Virginia Tech, Alexandria, VA 22305, USA

Action-Conditioned World Models Graph-Based Planning Spatial Epidemic Modeling

Ask PDF

Abstract

Graph Structure Learning

Explore similar work

Learning and Forecasting Multiscale Spatiotemporal Epidemic Dynamics with Lagrangian-Eulerian Graph Neural ODEs

A Network Science Perspective on Evaluating Deep Graph Generative Models

Graph World Models for Constrained Epidemic Policy Planning