NIRG - Generating Socially Realistic Synthetic Networks
Lead Research Organisation:
Durham University
Department Name: Sociology
Abstract
The ongoing COVID-19 epidemic has demonstrated the many ways in which simulations can assist policy makers in a rapidly changing world. It is well documented that the first lockdown was introduced from 23 March 2020 in response to modelling work by Neil Ferguson and his team at Imperial College. Their model showed that hundreds of thousands of deaths were likely in the UK if the epidemic was left uncontrolled.
Simulations have been used throughout the epidemic to adjust social distancing measures in a careful balance between hospital capacity and the economic impact of restricting activities. The most complete measure of COVID-19 spread and impact is hospital admissions, because not all cases are detected. However, the delay between exposure and potential admission means that even real time hospital admission data are not available until after decisions must be made. Simulations are able to guide policy because they can combine theoretical processes with current data to construct justified stories about plausible futures. Such stories are able to provide insight even if the future policy measures are different than those that generated the data.
Big data initiatives have made it relatively straightforward to include schools, transport and other infrastructure into these simulations. Similarly, census data can be used to construct synthetic households with simulated people who go to work or school or leisure activities. Efforts are ongoing to include more detailed high resolution information into epidemic models.
What is missing, however, is similar resolution data about social networks. Thanks to some big studies, we know how many people come in contact with each other and the age and gender mix of such contact patterns. But we don't know much about the broader structure of contacts. In a parallel to existing methods to construct synthetic households, we need new methods to construct synthetic social networks that are similar enough to real networks to be used in simulations. It is important, for example, to include structures that recognise mutual friends because some of the people in contact with an infected person may be already exposed.
This project will develop methods to build synthetic social networks that reproduce structural properties that we know are important in real social networks, like mutual friends.
Simulations have been used throughout the epidemic to adjust social distancing measures in a careful balance between hospital capacity and the economic impact of restricting activities. The most complete measure of COVID-19 spread and impact is hospital admissions, because not all cases are detected. However, the delay between exposure and potential admission means that even real time hospital admission data are not available until after decisions must be made. Simulations are able to guide policy because they can combine theoretical processes with current data to construct justified stories about plausible futures. Such stories are able to provide insight even if the future policy measures are different than those that generated the data.
Big data initiatives have made it relatively straightforward to include schools, transport and other infrastructure into these simulations. Similarly, census data can be used to construct synthetic households with simulated people who go to work or school or leisure activities. Efforts are ongoing to include more detailed high resolution information into epidemic models.
What is missing, however, is similar resolution data about social networks. Thanks to some big studies, we know how many people come in contact with each other and the age and gender mix of such contact patterns. But we don't know much about the broader structure of contacts. In a parallel to existing methods to construct synthetic households, we need new methods to construct synthetic social networks that are similar enough to real networks to be used in simulations. It is important, for example, to include structures that recognise mutual friends because some of the people in contact with an infected person may be already exposed.
This project will develop methods to build synthetic social networks that reproduce structural properties that we know are important in real social networks, like mutual friends.
Technical Summary
Network science or social network analysis is an interdisciplinary field developed from traditions in both sociology and mathematics. The sociological tradition is micro-analytical, empirical and specific; researchers collect information about the relationships within some real world social group of interest, such as 'friend of' within a particular school class or 'seeks advice from' within a company, and analyse that network to answer research questions such as identifying people who exert the greatest influence. The mathematical tradition is macro-analytical, theoretical and general. That is, researchers typically identify classes of networks and develop general rules about how different network properties are inter-related for some class.
Despite the differences, basic concepts are shared. The questions asked in both traditions examine the structure of network(s) that represent entities and the relationships between them. However, strong assumptions are required to make the mathematics of networks tractable, typically that edges occur randomly within some broad constraints. Consequently, networks generated for mathematical studies have little in common with real world social networks. For example, social networks have a high clustering coefficient and degree assortativity. Other properties, such as betweenness and path lengths, are also important for epidemic spread and other dynamic processes occurring over the network.
Simulation studies of dynamic processes such as epidemics require synthetic networks with structures that are similar to real world social networks if their results are to be applied to real world issues. Network generation algorithms can reproduce one or two of these real world structural features, but algorithms are not available for three or more features. This project will combine and extend existing approaches to develop algorithms to create synthetic networks that are realistic over a greater range of social network properties.
Despite the differences, basic concepts are shared. The questions asked in both traditions examine the structure of network(s) that represent entities and the relationships between them. However, strong assumptions are required to make the mathematics of networks tractable, typically that edges occur randomly within some broad constraints. Consequently, networks generated for mathematical studies have little in common with real world social networks. For example, social networks have a high clustering coefficient and degree assortativity. Other properties, such as betweenness and path lengths, are also important for epidemic spread and other dynamic processes occurring over the network.
Simulation studies of dynamic processes such as epidemics require synthetic networks with structures that are similar to real world social networks if their results are to be applied to real world issues. Network generation algorithms can reproduce one or two of these real world structural features, but algorithms are not available for three or more features. This project will combine and extend existing approaches to develop algorithms to create synthetic networks that are realistic over a greater range of social network properties.
