Estimating the parameters of epidemic spread on two-layer random graphs: a classical and a neural network approach

In this paper, we study the spread of a classical SIR process on a two-layer random network, where the first layer represents the households, while the second layer models the contacts outside the households by a random scale-free graph. We build a three-parameter graph, called polynomial model, whe...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Backhausz, Ágnes, Bognár, Edit, Csiszár, Villő, Tárkányi, Damján, Zempléni, András
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In this paper, we study the spread of a classical SIR process on a two-layer random network, where the first layer represents the households, while the second layer models the contacts outside the households by a random scale-free graph. We build a three-parameter graph, called polynomial model, where the new vertices are connected to the existing ones either uniformly, or preferentially, or by forming random triangles. We examine the effect of the graph's properties on the goodness of the estimation of the infection rate $\tau$, which is the most important parameter, determining the reproduction rate of the epidemic. In the classical maximum likelihood approach, to estimate $\tau$ one needs to approximate the number of SI edges between households, since the graph itself is supposed to be unobservable. Our simulation study reveals that the estimation is poorer at the beginning of the epidemic, for larger preferential attachment parameter of the graph, and for larger $\tau$. We present two heuristic improvement algorithms and establish our method to be robust to changes in average clustering of the graph model. We also extend a graph neural network (GNN) approach for estimating contagion dynamics for our two-layered graphs. We find that dense networks offer better training datasets. Moreover, GNN perfomance is measured better using the $l_2$ loss function rather than cross-entropy.
DOI:10.48550/arxiv.2303.02195