DNN-Based Prediction Model for Spatio-Temporal Data

Advances in location-acquisition and wireless communication technologies have led to wider availability of spatiotemporal (ST) data, which has unique spatial properties (i.e. geographical hierarchy and distance) and temporal properties (i.e. closeness, period and trend). In this paper, we propose a Deep-learning-based prediction model for patio Temporal data (DeepST). We leverage ST domain knowledge to design the architecture of DeepST, which is comprised of two components: spatio-temporal and global. The spatio-temporal component employs the framework of convolutional neural networks to simultaneously model spatial near and distant dependencies, and temporal closeness, period and trend. The global component is used to capture global factors, such as day of the week, weekday or weekend. Using DeepST, we build a real-time crowd flow forecasting system called UrbanFlow. Experiment results on diverse ST datasets verify DeepST's ability to capture ST data's spatio-temporal properties, showing the advantages of DeepST beyond four baseline methods.