STM3: Multiscale Mamba Model for Long-Term Spatio-Temporal Prediction
A novel deep learning framework named Spatio-Temporal Mixture of Multiscale Mamba (STM3) has been introduced for predicting long-term spatio-temporal time series. This model tackles two significant issues: the extraction of multiscale data from extended sequences and the modeling of correlations among multiscale temporal information across various nodes. STM3 employs a Multiscale Mamba structure within a Disentangled Mixture-of-Experts (DMoE) system to effectively gather diverse multiscale insights. Additionally, it features an adaptive graph causal network to represent intricate spatial relationships. To enhance representation learning, the model utilizes a stable routing strategy alongside causal contrastive learning, aiming to address the shortcomings of current deep learning techniques in managing long-term spatio-temporal dependencies. The findings are available on arXiv under identifier 2508.12247.
Key facts
- STM3 stands for Spatio-Temporal Mixture of Multiscale Mamba
- Model uses Multiscale Mamba architecture within a Disentangled Mixture-of-Experts (DMoE) framework
- Adaptive graph causal network models spatial dependencies
- Stable routing strategy and causal contrastive learning ensure robust representation
- Addresses challenges of long-term spatio-temporal dependency learning
- Published on arXiv with identifier 2508.12247
- Focuses on long-term spatio-temporal time-series prediction
- Existing deep learning methods struggle with complex long-term dependencies
Entities
Institutions
- arXiv