Spatio-Temporal Pyramid Networks for Traffic Forecasting

Traffic flow forecasting is an important part of smart city construction. Accurate traffic flow forecasting helps traffic management agencies to make timely adjustments, thus improving pedestrian travel efficiency and road utilization. However, this work is challenging due to the dynamic stochastic factors affecting the variation of traffic data and the spatially hidden behavior. Existing approaches generally use attention mechanism or graph neural networks to model correlation in temporal and spatial terms, and despite some progress in performance, they still ignore a number of practical situations: (1) Anomalous data due to traffic accidents or traffic congestion can affect the accuracy of modeling in the current moment and further create potential optimization problems for model training. (2) According to the directedness of the road, the hiding behavior between nodes should also be unidirectional and dynamic. In this paper, we propose a dynamic graph network with a pyramid structure, named PYNet, and use it for traffic flow forecasting tasks. Specifically, first we propose the Pyramid Constructor for transforming multivariate time series into a pyramid network with a multilevel structure, where the higher the level, the larger the range of time scales represented. Second, we perform Trend-Aware Attention top-down in the pyramid network, which gradually enables the lower-level time series to learn their long-term dependence in multiples, and effectively reduces the impact of outliers. Furthermore, to fully capture the hidden behavior in the spatial dimension, we learn an adaptive unidirectional graph and perform forward and backward diffusion convolution on the graph. Experimental results on two types of datasets show that PYNet outperforms the state-of-the-art baseline.