MP3: Multi-Period Pattern Pre-training forSpatio-Temporal Forecasting

The article introduces Multi-Period Pattern Pre-training (MP3), a novel plugin for spatio-temporal forecasting that enhances existing spatio-temporal graph neural networks (STGNNs). MP3 employs multi-period temporal and spatial modeling using edge convolution and a causality-enhanced Transformer to effectively identify and leverage temporal mirages in urban data. Experimental results demonstrate that MP3 improves forecasting performance by reducing the Mean Absolute Error (MAE) by 4.7% and Root Mean Square Error (RMSE) by 5.0% across five STGNN baselines on real-world datasets, indicating its strong adaptability and scalability for practitioners in the field.