MultiCell

Abstract

Exploring cellphone network data has been proved to be a very effective way to understand urban populations because of the high penetration rate of cellphones. However, the state-of-the-art population models driven by cellphone data are typically built upon single cellphone networks, assuming the users in a particular cellphone network used are representative of all residents in the studied city with multiple cellphone networks. This assumption usually does not hold in the real world due to strategic spatial coverages and business concentrations of cellphone companies, which lead to data biases, and thus overfitting of resultant population models. To address this issue, we design a model called MultiCell to model real-time urban populations from multiple cellphone networks with two novel techniques: (i) a network realignment technique to integrate individual cell-tower spatial distributions from multiple cellphone networks for finer granular population modeling; (ii) a data fusion technique based on cross-network training to design a population model based on multiple network data. We implement MultiCell in the Chinese city Shenzhen based on three cellphone networks with 10 million active users and their daily data records at 11 thousand cell towers. We evaluate MultiCell by comparing it to the state-of-the-art models driven by single cellphone networks, and the evaluation results show that MultiCell outperforms them by 27% in terms of accuracy. Finally, we cross-validate MultiCell with three transportation systems with more than 8 million passengers to investigate its performances.