Traffic prediction via spatial-temporal graph learning network based on multi-domain temporal gated self-attention and attentive spectral graph attention network

Abstract

Recently, traffic prediction, which serves as a fundamental component of intelligent transportation systems (ITS), has become essential for traffic planning and management. Nevertheless, existing traffic prediction methods still face challenges in capturing critical information from traffic data, such as spatial-temporal dependencies, dynamics, periodicity, heterogeneity, etc. Therefore, a novel spatial-temporal graph learning network (ST-GLN) based on multi-domain temporal gated self-attention and attentive spectral graph attention network is proposed in this paper. It can efficiently capturing crucial information mentioned above from traffic data. Specifically, an effective data embedding module is first constructed to project raw traffic data into high-dimensional representations, in which essential positional and periodic information is encoded. Second, we design spatial-temporal encoders to extract critical information from traffic data, namely, spatial-temporal dependencies, dynamics, and heterogeneity. Within each spatial-temporal encoder, we employ a temporal encoder consisting of time-domain gated self-attention and residual frequency-domain gated self-attention to capture temporal dependencies, dynamics, and heterogeneity. Meanwhile, a spatial encoder composed of self-scaled dynamic diffusion spectral graph neural network and multi-view graph self-attention is proposed to learn spatial dependencies, dynamics, and heterogeneity. Thirdly, we adopt a multi-layer perceptron as the output embedding module to strengthen its capability of learning deep representations and generating final prediction results. Finally, we conduct extensive experiments on several representative traffic datasets to demonstrate that our method achieves superior performance against existing state-of-the-art approaches. For example, we boost the MAE, RMSE and MAPE by up to 16.83, 21.95 and 14.15% when we use ST-GLN to predict PEMS08. And the MAE, RMSE and MAPE are enhanced by up to 17.17, 23.34 and 13.11% when we use ST-GLN to predict PEMS04.