Understanding Mobility Dynamics and Predicting Urban Traffic State via Improved Unsupervised Learning

Traffic dynamic evolution concerning multiple coupling factors. The key to traffic forecasting is to deal with the multi-modal coupling spatiotemporal factors in the observation data, such as weather information (temperature, wind), different time scales (hours, days, weeks), and some uncontrollable random factors (traffic accidents, etc.). To this end, this paper proposes the semantic factorization-based traffic prediction generative adversarial network (SFTPGAN), which is an improved semantic factorization method based on unsupervised learning. It can automatically find meaningful semantic information in traffic dynamics evolution through its network structure and visualize the impact of each factor on the traffic dynamics evolution by changing the direction of each semantic individually. We evaluate the model on a large-scale GPS trajectory data set in the main urban area of Beijing and find it works well in searching semantic information.