Modeling Spatial-Temporal Dynamics for Traffic Prediction

Spatial-temporal prediction has many applications such as climate forecasting and urban planning. In particular, traffic prediction has drawn increasing attention in data mining research field for the growing traffic related datasets and for its impacts in real-world applications. For example, an accurate taxi demand prediction can assist taxi companies to pre-allocate taxis to meet with commuting demands. The key challenge of traffic prediction lies in how to model the complex spatial and temporal dependencies. In this paper, we make two important observations which have not been considered by previous studies: (1) the spatial dependency between locations are dynamic; and (2) the temporal dependency follows strong periodicity but is not strictly periodic for its dynamic temporal shifting. Based on these two observations, we propose a novel Spatial-Temporal Dynamic Network (STDN) framework. In this framework, we propose a flow gating mechanism to learn the dynamic similarity between locations via traffic flow. A periodically shifted attention mechanism is designed to handle long-term periodic dependency and periodic temporal shifting. Furthermore, we extend our framework from region-based traffic prediction to traffic prediction for road intersections by using graph convolutional structure. We conduct extensive experiments on several large-scale real traffic datasets and demonstrate the effectiveness of our approach over state-of-the-art methods.