Simulating Mobility of Large Population using Mobile Application Data

Abstract

Understanding people’s travel patterns and knowing the frequency of their travel activities provide important insights for effective transport management and infrastructure planning. While traditional travel surveys have historically contributed to this valuable information, they come with shortcomings such as being expensive to collect, easily outdated, and having short time coverage. With technology advancing, new data sources on human mobility, such as Mobile Application Data (MAD), passively log people’s geolocations and provide larger and more diverse datasets than traditional travel surveys. However, the resulting information on an individual level is often too sparse to be used in more advanced mobility models. Acknowledging these limitations of MAD, this thesis aims to leverage the strengths of both datasets. By combining the traditional travel survey data with the richer and more extensive dataset from mobile phones, we intend to synthesize comprehensive activity plans for those living in Sweden. This thesis makes two key contributions. Firstly, it enhances the accuracy of identified home locations by analyzing their temporal visitation patterns and comparing them with survey data. The candidate agents whose patterns align closely with the survey data are selected based on the similarity of their temporal distributions. Secondly, it proposes a simple and transferable generative model for synthesising activity plans, which integrates big geodata and survey data. In this model, for each agent, we identify a corresponding “twin traveler” from the travel diary data. We then enrich the activity sequences of these twins with the extensive location data collected from big geodata sources over several months. The proposed model identifies home and work as anchor locations and compares the home location with survey data to exclude unreliable ones. It then transforms user data into activity plans and applies a modified Jaccard similarity to find matching twins between datasets. Finally, it creates synthesized activity plans by combining the activity sequences of survey twins with the extensive location data from mobile app users. The resulting 113 488 synthesized activity plans are then validated against the 18 106 survey responses regarding the essential attributes of individual mobility patterns. We employ the Kullback-Leibler divergence to compare the similarities between the two datasets. The validation shows that our model generally agrees with the survey data. These results indicate that, with some future improvements, generative models combining survey and big geodata sources, as MAD in this thesis, are valuable and promising for future mobility studies.