Szczegóły publikacji
Opis bibliograficzny
Neural emulation of spreading processes on complex networks: from the zealot voter model to epidemics and bounded-confidence opinion dynamics / Vahid MOEINIFAR // Physica . A, Statistical Mechanics and its Applications ; ISSN 0378-4371. — 2026 — vol. 701 art. no. 132014, s. 1–16. — Bibliogr. s. 15–16, Abstr. — Publikacja dostępna online od: 2026-09-11
Autor
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 170239 |
|---|---|
| Data dodania do BaDAP | 2026-09-24 |
| Tekst źródłowy | URL |
| DOI | 10.1016/j.physa.2026.132014 |
| Rok publikacji | 2026 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Creative Commons | |
| Czasopismo/seria | Physica, A, Statistical Mechanics and Its Applications |
Abstract
Monte Carlo simulation of processes spreading on complex networks becomes expensive as the system size increases. It runs in the order of O(NTR) for R simulations over time T with N nodes, while mean-field reductions lose the structural detail that governs how a few fixed influential agents reshape the dynamics. We address this by benchmarking five successive neural architectures, culminating in the ZealotTransformer (ZT). Node-level structural features are embedded, passed through a multi-head self-attention module, and then separately pooled for zealot and non-zealot nodes; an LSTM decoder then generates the trajectory autoregressively. In the zealot Voter Model, ZT achieves a root-mean-square error (RMSE) below 0.05 in the Barabási–Albert, Erdős–Rényi and Watts–Strogatz ensembles under placements of hub, random and bridge. It extrapolates to network sizes well beyond the training range (N = 4096, 8192) without retraining, while descriptor-based emulators degrade sharply. These three ensembles are all represented in training, so this measures interpolation across the training families rather than transfer to an unseen graph family. The same architecture, retrained from scratch on independently generated data, also reproduces the infected fraction in the susceptible–infected–susceptible (SIS) epidemic model (RMSE below 0.02), and the opinion-cluster count of the Deffuant–Weisbuch bounded-confidence model (to within a fraction of a cluster). In both cases, it significantly outperforms mean-field and persistence baselines. Accuracy degrades on spatially embedded Random Geometric Graphs, a training-absent ensemble, setting the structural regime in which the surrogate remains valid.
