Revealing Urban Dynamics by Learning Online and Offline Behaviours Together

Abstract

Urban problems and diseases accompanied by the pace of urbanization have drawn attention to the importance of understanding urban dynamics, while a deep and comprehensive understanding is challenging due to our diversified lifestyles in the modern city. In this paper, we propose an urban dynamics modeling system to characterize the regularity of urban activity dynamics as well as urban functions by learning residents' online and offline behaviours together. Built on a state-sharing hidden Markov model, our system utilizes online activities of App usage and offline activities of mobility in different urban regions and different time slots for learning. The learnt state sequence of each region reveals urban dynamics with the corresponding urban functions. We evaluate our system via a large-scale mobile network accessing dataset, which discovers ten hidden states characterizing different life modes and eight representative dynamic patterns corresponding to different urban functions. These discovered dynamic patterns and inferred functions are validated by social media check-ins and the land-use published by the government with 81% accuracy. Based on our model, we propose two applications, crowd flow prediction and popular App prediction, which outperforms the state-of-the-art approaches by 36.1% and 15.7%, respectively. This study paves the way for extensive city-related applications including urban demand analysis, land-use planning, and activity prediction.