STM3: Mixture of Multiscale Mamba for Long-Term Spatio-Temporal Time-Series Prediction
Haolong Chen, Liang Zhang, Zhengyuan Xin, Guangxu Zhu
Key claim
STM3 achieves state-of-the-art results in time-series prediction.
STM3 effectively captures complex long-term spatio-temporal dependencies using a unique architecture. It significantly outperforms the second-best model on the PEMSD8 dataset by 7.1% in MAE, showcasing its robustness in time-series prediction.
The proposed STM3 introduces a novel framework for efficiently capturing multiscale spatio-temporal dependencies.
The methodology is solid, with extensive experiments demonstrating superior performance on real-world benchmarks.
Deep reliability assessment
The methodology supports the claim that STM3 effectively captures long-term spatio-temporal dependencies through its innovative architecture, but the extent of its superiority over existing models may be overstated without comprehensive comparisons across all relevant datasets. Additionally, the theoretical proofs provided may not fully account for practical implementation challenges in diverse real-world scenarios.
Reproducibility
Yes, the code is available at https://github.com/IfReasonable/STM3_KDD26.
Discussion questions
- What assumptions about the nature of spatio-temporal dependencies might limit the applicability of STM3 in different domains?
- How can builders leverage the findings of this paper to improve their own models in practical applications?
- What specific conditions or datasets would lead to STM3 underperforming compared to its claimed results?
Key figure
Figure 1 illustrates the main structure of STM3, highlighting the integration of Multiscale Preprocessing, Adaptive Graph Causal Convolution Network, and Disentangled Mixture of Multiscale Mamba.