Generating fine-grained surrogate temporal networks
Fondazione Bruno Kessler · University of Trento · Aix-Marseille Université, Université de Toulon & CNRS · Technical University of Denmark · Copenhagen Center for Social Data Science
Communications Physics · Volume 7 · Article 22
In one sentence. ETN-gen learns how short egocentric neighborhoods evolve and recombines those local patterns into realistic temporal networks of arbitrary duration and size.
Problem
Fine-grained temporal networks are costly to collect, difficult to share, and often too small or short for studying contagion, mobility, or social dynamics. Existing generators may preserve topology without temporal correlations, model individual activity without realistic structure, or become too expensive when motifs and deep models are applied to many sparse snapshots.
Main contributions
- ETN-gen represents recent node activity through compact Egocentric Temporal Neighborhood signatures.
- A learned distribution of signature continuations generates each new network layer without storing node identities.
- The method scales linearly with nodes and timestamps and can enlarge a dataset in time, node count, or both.
- Structural and dynamical tests compare ETN-gen with Dymond, STM, and TagGen.
Method
For every node and sliding time window, ETN-gen removes links among neighbors and encodes the remaining star-like temporal pattern as a binary signature. Maximum-likelihood transition dictionaries estimate how each signature can extend into the next snapshot. Sampled node requests are reconciled into undirected edges, while separate dictionaries for times of day or weekdays preserve periodic activity.
Experimental setting
- Main datasets
- Hospital (75 nodes), workplace (92 nodes), and high school (126 nodes) face-to-face networks.
- Resolution
- Five-minute snapshots and temporal horizon k = 2; larger social and communication datasets appear in supplementary tests.
- Structure
- Seventeen time-dependent and aggregated measures compared through Kolmogorov–Smirnov distance.
- Dynamics
- Temporal random walks and Susceptible–Infectious–Recovered spreading simulations.
- Baselines
- Dymond, Structural Temporal Modeling (STM), and TagGen; ten stochastic network realizations.
Key results
ETN-gen is the only compared method that simultaneously preserves node count, interaction scale, and the original periodicity. It is strongest on fine-grained temporal properties and can extrapolate a second workplace week from the first. It also reconstructs high-school activity after retaining only part of the original nodes.
| Evaluation | Reported outcome |
|---|---|
| Temporal structure | High fidelity for density, active individuals, new conversations, connected components, and hub-like structure. |
| Random walks | Mean first-passage times remain similar; coverage is competitive and depends on dataset and starting time. |
| Temporal extension | A model learned from week one recreates the workplace interaction pattern in week two. |
| Size expansion | Surrogates recover high-school activity from reduced networks containing 30% or 70% of the nodes. |
When this work is relevant
ETN-gen is relevant for temporal-network augmentation, scalable synthetic contact data, privacy-conscious data sharing, epidemic simulations, high-resolution human interactions, and interpretable alternatives to motif-heavy or neural temporal graph generators.
Limitations
The egocentric representation discards links among neighbors, reducing fidelity for clustering, assortativity, shortest paths, and communities. Its short memory cannot reproduce long-term relationship reinforcement or daily and weekly recurrences unless periodicity is provided explicitly. The paper argues that identities are obscured, but does not formally prove resistance to de-anonymization.
Citation
Persistent identifier: https://doi.org/10.1038/s42005-023-01517-1