MP3 pre-training cuts forecasting errors by 5% across 5 datasets

Spatio-temporal forecasting faces a critical challenge called 'temporal mirage': similar short-window inputs can lead to wildly divergent future trends, and vice versa. Existing spatio-temporal graph neural networks (STGNNs) struggle to distinguish these mirages due to incomplete period observations, heterogeneous global spatial correlations, and cross-period superposition causality. To address this, the authors introduce MP3 (Multi-Period Pattern Pre-training), a new plug-and-play pre-training plugin designed to strengthen any STGNN backbone. MP3 introduces three core innovations: multi-period temporal modeling (using edge convolution to identify patterns), multi-period spatial modeling (with a bottleneck projection and global memory bank to capture heterogeneous spatial relations), and cross-period pattern interaction (via a causality-enhanced Transformer to capture dependencies across different period patterns).

Experiments across five STGNN baselines and five real-world datasets (including a large-scale California dataset) demonstrate MP3's effectiveness, scalability, and adaptability. On average, MP3 reduces Mean Absolute Error (MAE) by 4.7% and Root Mean Square Error (RMSE) by 5.0%—consistent and robust improvements without modifying the underlying model architecture. The code is publicly available, making it easy for practitioners in traffic, climate, and energy domains to integrate MP3 into their existing forecasting pipelines. The approach is particularly valuable for urban applications where long-term dependencies and periodic patterns are critical.