MAMGNN: Bidirectional Mamba Based Spatio-Temporal Series Prediction Graph Neural Network

Due to the high value of spatio-temporal series forecasting, it has been regarded as a high-priority research topic in the fields of economics, physics, and transportation. Spatio-temporal graph neural networks can extract temporal and spatial information, thus suitable for spatio-temporal series prediction tasks. After the spatial information has been determined, the effectiveness of temporal correlation information extraction determines the accuracy of the prediction method. However, traditional methods cannot simultaneously balance computational consumption and long-range prediction capability in terms of temporal information. Therefore, we propose MAMGNN which contains a temporal information extraction (TIE) module that can ensure the accuracy of long series prediction at low computational consumption due to its structure. It also employs an inter-frame value variation trend constraint mechanism to ensure the accuracy of small-scale forecasts. Sufficient experiments demonstrate that our method achieves high prediction accuracy on economic and transportation datasets.