An inquiry into diffusion processes over interaction networks

Abstract

This thesis aims to develop a comprehensive framework for modelling and controlling diffusion processes over interaction networks, striving to inform and improve public health policies against viral epidemics. Our work introduces four main contributions: (1) a new modelling technique that captures the heterogeneity and uncertainty of contact patterns and evaluates the impact of different testing and tracing strategies, which can be utilized in conjunction with any compartmental formulation to study complex spreading dynamics. Using this technique, we introduce and simulate a novel epidemiological model, SEIR-T, showing that contact tracing in a COVID-19 epidemic can be effective despite suboptimal digital uptakes or pervasive interview inefficiencies; (2) a versatile and cost-effective approach to optimizing the allocation of testing, tracing and vaccination resources based on the network structure and epidemic dynamics, which ranks individuals based on their role in the network and the epidemic state, being adaptable to the budget and risk preferences of regional policy makers, while still breaking high-risk transmission chains; (3) a reinforcement learning-based agent, underpinned by a highly transferable graph neural architecture, that can find optimal epidemic control policies from simulation data, outperforming standard heuristic approaches by up to 15% in the containment rate, while far surpassing more standard random samplers by margins of 50% or more; and (4) a range of visualization tools that can aid in understanding and communicating the effects of public health interventions to policy makers and the populace, which
include prediction explanation and state visualization techniques for scrutinizing the learning-based policies introduced, and other tools the authorities can use to assess the cost-benefit trade-off of enacting different combinations of interventions. The simulation-control framework we introduce is particularly flexible and can effectually model the spread of various pathogens or analogous diffusion processes, such as information dissemination. Similarly, the learned epidemic policies are versatile and easily transferable to a wide range of diffusion scenarios and network structures.

More information

Identifiers

ORCID for Andrei C. Rusu: orcid.org/0000-0002-6053-1685

ORCID for Kate Farrahi: orcid.org/0000-0001-6775-127X

ORCID for Mahesan Niranjan: orcid.org/0000-0001-7021-140X

Catalogue record

Export record

Altmetrics