Unveiling social vibrancy in urban spaces with app usage

TL;DR

The paper tackles how online app usage maps onto urban vibrancy by linking high-frequency digital traces to sociospatial features across 18 French cities. It introduces a two-stage methodology: constructing digital signatures from app usage and clustering them with time-series k-means, then relating cluster labels to OpenStreetMap-derived third-place features via regularised multinomial logistic regression. The study finds three robust clusters with monocentric or polycentric patterns; diversity of third-place features generally outperforms simple counts in predicting cluster membership, though patterns vary by city and time of week. The work demonstrates the feasibility and value of combining fine-grained mobile data with geospatial features to understand urban vibrancy and the social fabric of urban spaces, while noting limitations such as data period and potential demographic biases.

Abstract

Urban vibrancy is an important measure of the energetic nature of a city that is related to why and how people use urban spaces, and it is inherently connected with our social behaviour. Increasingly, people use a wide range of mobile phone apps in their daily lives to connect socially, search for information, make decisions, and arrange travel, amongst many other reasons. However, the relationship between online app usage and urban vibrancy remains unclear, particularly regarding how sociospatial behaviours interact with urban features. Here, we use app-usage data as a digital signature to investigate this question. To do this, we use a high-resolution data source of mobile service-level traffic volumes across eighteen cities in France. We investigate the social component of cities using socially relevant urban features constructed from OpenStreetMap 'Points of Interest'. We developed a methodology for identifying and classifying multidimensional app usage time series based on similarity. We used these in predictive models to interpret the results for each city and across France. Across cities, there were spatial behavioural archetypes, characterised by multidimensional properties. We found patterns between the week and the weekend, and across cities, and the country. These archetypes correspond to changes in socially relevant urban features that impact urban vibrancy. Our results add further evidence for the importance of using computational approaches to understand urban environments, the use of sociological concepts in computational science, and urban vibrancy in cities.

Figures (14)

• Figure 1: Clustered app usage in space and time for Paris. Time series and maps for each cluster for weekdays (Monday--Thursday) (\ref{['fig:usv-local_paris_week']}) and weekends (Friday--Sunday) (\ref{['fig:usv-local_paris_weekend']}) for Paris at the local scale. The time series displays the average app usage for all cells in Paris. To facilitate comparison across app categories, the data were scaled using the min-max scaling technique, but this scaling was applied only for visualization purposes. The averaged app categories are shown as black lines, with each cluster centre visualized at the forefront of these. The size of each cluster is indicated in the top corner of each figure. Clusters are presented in size order, with the largest cluster assigned the smallest number. The maps depict the spatial arrangement of the clusters, and the legend shows the colour coding for the clusters in both the time series and the maps.
• Figure 2: Local-level multinomial logistic regression predictive model estimates for the weekdays and weekends in Paris. The figure contains estimates for the count and diversity of all third place features as well as each third place category. The y-axis displays each coefficient. Coefficients for the week (darker shade) and weekend (lighter shade). Data preprocessing was applied to the NetMob23 data set martinez-duriveNetMob23DatasetHighresolution2023. The legend shows the colour coding for the clusters.
• Figure 3: Global-level multinomial logistic regression predictive model estimates for the weekdays and weekends. The figure contains estimates for the count and diversity of all third place features as well as each third place category. The y-axis displays each coefficient. Coefficients for the week (darker shade) and weekend (lighter shade). Data preprocessing was applied to the NetMob23 data set martinez-duriveNetMob23DatasetHighresolution2023. The legend shows the colour coding for the clusters.
• Figure SI 1: Flowchart showing digital signatures, POIs, clustering, and modelling process in 18 French cities.
• Figure SI 2: Mobile apps and their assigned Apple Store categories. The apps are given in the left-hand column and the app categories are given in the right-hand column.