A Variational Autoencoder Based Generative Model of Urban Human Mobility

Recently, big and heterogeneous human mobility data inspires many revolutionary ideas of implementing machine learning algorithms for solving some traditional social issues, such as zone regulation, air pollution, and disaster evacuation el at.. However, incomplete datasets were provided owing to both the concerns of violation of privacy and some technique issues in many practical applications, which leads to some limitations of the utility of collected data. Variational Autoencoder (VAE), which uses a well-constructed latent space to capture salient features of the training data, shows a significant excellent performance in not only image processing, but also Natural Language Processing domain. By combining VAE and sequence-to-sequence (seq2seq) model, a Sequential Variational Autoencoder (SVAE) is built for the task of human mobility reconstruction. It is the first time that this kind of SVAE model is implemented for solving the issues about human mobility reconstruction. We use navigation GPS data of selected greater Tokyo area to evaluate the performance of the SVAE model. Experimental results demonstrate that the SVAE model can efficiently capture the salient features of human mobility data and generate more reasonable trajectories.