LSTDNet: A Long-Range Spatio-Temporal Decoupling Network for Traffic Prediction

Traffic flow forecasting is integral to the functionality of intelligent transportation systems, with the primary challenge lying in the accurate representation of intricate spatio-temporal relationships. However, few existing models address two key problems: how to accurately capture long-range spatial correlations and how to learn from the long-term spatio-temporal context. To address this, we propose a Long-range Spatio-Temporal Decoupling Network (LSTDNet) for traffic prediction. Briefly, LSTDNet uses two masking autoencoders in the pre-training phase to mask and reconstruct traffic flow sequences along the temporal and spatial axes, respectively, in order to capture long-range spatial-temporal correlations from long-term spatio-temporal contexts. Then, the learned spatio-temporal embeddings are fed into a multilayer network in the prediction phase to obtain the final prediction results. Experiments on four publicly available real-world datasets show that LSTDNet can effectively capture long-range spatio-temporal correlations and achieve competitive results.