Community aware temporal network generation
Fondazione Bruno Kessler · University of Trento · Université de Toulon, CNRS & Aix-Marseille Université
Applied Network Science · Volume 10 · Article 43
In one sentence. LETN-gen conditions local temporal interaction patterns on node labels, allowing synthetic networks to retain communities and the distinct behavior of social roles.
Problem
ETN-gen efficiently learns local temporal behavior, but treats every node as if it drew from the same pattern distribution. Real social networks are modular: students, teachers, clinicians, patients, and workplace groups interact with different partners and durations. Ignoring those roles produces plausible activity while losing the community organization that shapes diffusion and other network dynamics.
Main contributions
- Labeled Egocentric Temporal Neighborhoods (LETN) encode both recent local structure and neighbors’ categories.
- CLETN and DLETN infer labels through aggregated or time-local community detection when metadata are unavailable.
- The generator remains effectively linear in nodes and snapshots when temporal neighborhoods stay bounded.
- Seven SocioPatterns datasets test community structure, role-dependent behavior, and interaction duration.
Method
LETN augments each binary egocentric temporal signature with the labels of neighboring nodes. Counts of signature continuations form conditional probability dictionaries, which are sampled to build future layers. Reciprocal requests are accepted first; remaining stubs are matched to maximize agreement. Labels can come from metadata, Louvain communities on the aggregated graph, or partitions recomputed over local time splits.
Experimental setting
- Datasets
- Primary school; high schools from 2011, 2012, and 2013; hospital; workplaces from 2013 and 2015.
- Scale
- 75–327 participants, 4–12 labels, and 9,827–188,508 recorded interactions.
- Protocol
- Ten generated networks per dataset, five-minute snapshots, and temporal horizon k = 2.
- Comparisons
- LETN against the original ETN generator and Motif Transition Model (MTM).
- Measures
- Modularity, label assortativity, clustering, cross-community interaction counts, and interaction duration.
Key results
LETN reproduces the temporal distributions of modularity and label assortativity across the seven datasets, while ETN and MTM largely lose the partition structure. It also captures which groups interact and for how long. CLETN and DLETN remain close to metadata-based LETN, showing that detected communities can substitute for missing labels.
| Dataset / measure | Original | LETN | ETN |
|---|---|---|---|
| High school 2013 / modularity | 0.76 ± 0.10 | 0.80 ± 0.04 | −0.01 ± 0.06 |
| High school 2013 / label assortativity | 0.92 ± 0.11 | 0.93 ± 0.05 | −0.01 ± 0.07 |
| Primary school / modularity | 0.56 ± 0.14 | 0.64 ± 0.11 | 0.00 ± 0.06 |
| Workplace 2015 / label assortativity | 0.71 ± 0.26 | 0.70 ± 0.10 | −0.13 ± 0.14 |
When this work is relevant
The method is useful for privacy-conscious synthetic interaction data, epidemic and diffusion simulations, temporal social networks, role-aware human-contact modeling, and enlarging short or small face-to-face datasets without discarding their mesoscopic organization.
Limitations
Validation focuses on sparse face-to-face social networks and one main external competitor. Generated layers require seed snapshots, while inferred labels depend on a chosen community-detection procedure. Matching modularity and assortativity supports structural fidelity, but does not by itself establish fidelity for every downstream dynamical process or non-social temporal domain.
Citation
Persistent identifier: https://doi.org/10.1007/s41109-025-00731-w