MSDR: Multi-Step Dependency Relation Networks for Spatial Temporal Forecasting
Dachuan Liu, Jin Wang, Shuo Shang, Peng Han
Abstract
Spatial temporal forecasting plays an important role in improving the quality and performance of Intelligent Transportation Systems. This task is rather challenging due to the complicated and long-range spatial temporal dependencies in traffic network. Existing studies typically employ different deep neural networks to learn the spatial and temporal representations so as to capture the complex and dynamic dependencies. In this paper, we argue that it is insufficient to capture the long-range spatial dependencies from the implicit representations learned by temporal extracting modules. To address this problem, we propose Multi-Step Dependency Relation (MSDR), a brand new variant of recurrent neural network. Instead of only looking at the hidden state from only one latest time step, MSDR explicitly takes those of multiple historical time steps as the input of each time unit. We also develop two strategies to incur the spatial information into the dependency relation embedding between multiple historical time steps and the current one in MSDR. On the basis of it, we propose the Graph-based MSDR (GMSDR) framework to support general spatial temporal forecasting applications by seamlessly integrating graph-based neural networks with MSDR. We evaluate our proposed approach on several popular datasets. The results show that the proposed GMSDR framework outperforms state-of-the-art methods by an obvious margin.