Mining hourly population dynamics by activity type based on decomposition of sequential snapshot data

The dynamic population distributions by activity type (e.g. working, shopping or in-home) are vital for resource allocation, urban planning and epidemic containment. Although studies have incorporated individual-level human mobility data to map population distribution by activity type, access to such data is hindered due to privacy issues and they rely on auxiliary data to provide priori activity knowledge. This paper presents a method for generating the population dynamics by activity type. We first introduce more readily available sequential snapshot data to construct the population mixture model, then decompose the population mixture, and finally estimate the dynamic population size for each activity. We test the method in the central districts of Guangzhou city, China, based on real-time Tencent user density data. Correlation analysis and accuracy assessment prove that our method can accurately estimate hourly distributions for populations engaging in working, stay-at-home, and socializing activities. The temporal distribution of the working population reproduces the regular work scenarios and socializing population displays complex spatial patterns. We also find that there is an underlying relationship between a region's function and its dynamic population structure. The presented method has great potential for application and could provide new insight for studying urban dynamic functions.