Rethinking Spatio-Temporal Transformer for Traffic Prediction:Multi-level Multi-view Augmented Learning Framework

Traffic prediction has become an essential component of Intelligent Transportation Systems (ITS), which encompasses various applications such as traffic management[21], route planning[2] and congestion avoidance[12]. The main challenge lies in efficiently capturing the complex and time-varying spatio-temporal dependencies of traffic data. Recurrent Neural Networks (RNNs)[4, 23] and their variants, such as LSTM[35] and GRU[8], are used to capture temporal dependencies of traffic data. Nonetheless, these methods fail to model spatial correlations. To address this limitation, recent research has combined Convolutional Neural Networks (CNNs)[14, 16, 29] and RNNs to capture spatio-temporal dependencies of grid-based traffic data, with models like ST-ResNet[37] and STDN[34] proposed for this purpose. However, CNNs have inherent limitations in handling common non-Euclidean data representations. Recently, Spatio-Temporal Graph Neural Networks (STGNNs) have been developed for traffic prediction. These models combine GNNs with either RNNs or Temporal Convo-lutional Networks (TCNs) to capture the spatio-temporal correlations of traffic data.