AdaptST: Adaptive Fusion-based Spatio-Temporal Networks for Urban Flow Prediction

Urban flow prediction is of great importance to intelligent traffic system. Recently, it has attracted the attention of many researchers due to the vital impact of deep learning and big data. There are complex nonlinear dependencies such as spatio-temporal dependencies and external factors existing in the urban flow, which makes the accurate prediction challenging. Previous work has applied various deep neural networks in the task, in order to capture complicated dependencies. However, many of them rarely take into consideration all kinds of spatio-temporal dependencies, ignoring or weakening the influence of certain dependencies. In this paper, we propose a method entitled AdaptST (Adaptive fusion-based spatio-temporal networks) based on two kind of adaptive fusions in spatio-temporal networks. One of adaptive fusions is designed in the variant of ConvLSTM (Convolutional LSTM) cell, which fuses multiscale local and global features selectively instead of adding them directly. Another is adopted at the end of the long-term component to further boost the performance by attention-based weighted summation of feature maps at different time steps. Experimental results on real-world traffic dataset verify the effectiveness of our methods.