A Novel Data Fusion Method to Leverage Passively-collected Mobility Data in Generating Spatially-heterogeneous Synthetic Population

Conventional population synthesis methods rely on household travel survey (HTS) data. However, the synthesized population suffers from a low spatial heterogeneity issue due to high data aggregation and low sampling rates of HTS data. Passively collected (PC) data from smartphone devices or transit smart cards have the potential to overcome the limitations of HTS data, thanks to the continuous collection of mobility patterns at a high spatial resolution for a large proportion of the population. However, the mismatched spatial resolution, sampling rate, and attribute information make the fusion of HTS and PC data challenging. This study presents a novel cluster-based data fusion method that exploits the benefits of both HTS and PC data to generate a synthetic population with high spatial heterogeneity. As the number of the value combinations for spatial attributes (e.g., home and work locations) in PC data is much larger than that in HTS data, clustering is adopted to deal with the high-dimensionality issue and link spatial attributes in the two data sources. The data fusion problem is then formulated as tractable multiple low-dimensional optimization subproblems. The properties of the proposed method are analytically derived. Such analytical validation is necessary for an interpretable and trustworthy data fusion, which is infeasible to establish in state-of-the-art deep learning methods. Three experiments are conducted to validate the accuracy, illustrate the data fusion properties, and demonstrate the case study of the proposed method using the HTS and LTE/5G cellular signaling data from Seoul, South Korea.