Exploring the effect of location and time on human mobility flow networks

Abstract

Contemporary flow data are used to investigate a number of social science phenomena: examples include flows inferred from mobile phone data, migration flows, commuting flows, or taxi flows between pick‐up and set‐down points. This PhD thesis develops new bespoke advanced quantitative methods for flow data to better understand patterns in flows and their relationship with space and time. Flow data are mathematically represented as geographic networks and are due to their size and complicated structure a typical example of big data. In geography, there is a lack of appropriate methods for their analysis. While some disciplines, such as physics, have recently developed flow methods, these methods are not suitable for flow networks that are geographically constrained, as the methods are not scalable to large geographic networks, nor do they consider the effects of geographic location. Further, in physics, there is no consideration for temporal conditions which affect human mobility. This thesis explores different methods of flow network analysis by gradually experimenting with different approaches and network complexities. We begin the exploration by utilising community detection methods and a commuting network. Community detection allows us to find clusters within the network that have more connections between a location within them than they have to other clusters. We can consider this as a subcomponent of the network that has most of the movement happening within, and only a limited amount of movement towards other subcomponents. Then to complicate things further, we test different types of geographical weighting to the network. In that way, we confirm how to geographically weigh the distances depending on the dataset and decided purpose. Finally, we add a temporal component into the analysis and develop a bespoke method for spatio-temporal network clustering. The method allows us to look at time and space as a continuous variable and to detect patterns that span over different time periods. This would allow us to capture a pattern that only represents a few hours in a day and a pattern that captures the whole weekend behaviour without explicitly saying we are investigating a specific timespan. The results show that space and time make a significant difference in spatio-temporal analysis and the method enables us to simplify and explore huge networks in a seemingly complex way.