Modeling complex correlations between variables in multivariate time series forecasting is an increasingly recognized challenge. Transformer-based models have shown leading performance, but encounter obstacles of weak inductive bias and quadratic computational complexity. Recently emerged State Space Models, such as Mamba, offer comparable performance to Transformer with higher efficiency. Nevertheless, existing methods still face three key challenges. First, most models lack comprehension of graph positional inductive bias. Secondly, the coexistence of valuable information and noise hinders correlation extraction. Furthermore, representative invariant patterns in multivariate time series are overlooked in dynamic modeling. To address these issues, we present a novel Mamba framework called GPS-Mamba to effectively model multivariate correlations. Specifically, we propose the Laplace Scanning algorithm that introduces graph structure information and multivariate correlations into permutation to enhance positional inductive bias. Based on the algorithm, we utilize Mamba-based encoder to model long-term stable correlations and short-term dynamic variation with devised retrieval mechanism, respectively. Finally, we combine the features obtained from two perspectives to achieve more precise prediction. Experimental results on seven real-world datasets demonstrate the superiority of GPS-Mamba over SOTA baselines with high interpretability. The codes are available at https://github.com/yzh8221/GPS-Mamba.
@article{GPS-Mamba_2026,title={GPS-Mamba: Graph permutation scanning state space model for multivariate time series forecasting},journal={Expert Systems with Applications},volume={311},pages={131373},year={2026},issn={0957-4174},doi={https://doi.org/10.1016/j.eswa.2026.131373},url={https://www.sciencedirect.com/science/article/pii/S0957417426002861},author={Yao, Zihao and Zheng, Qi and Zuo, Jiankai and Zhang, Yaying},keywords={Multivariate correlation, Time series forecasting, State space model},}
2025
[TITS’25] TLAST
TLAST: A Time-Lag Aware Spatial-Temporal Transformer for Traffic Flow Forecasting
Qi Zheng, Minhua Shao, and Yaying Zhang*
IEEE Transactions on Intelligent Transportation Systems, 2025
@article{TLAST,author={Zheng, Qi and Shao, Minhua and Zhang, Yaying},journal={IEEE Transactions on Intelligent Transportation Systems},title={TLAST: A Time-Lag Aware Spatial-Temporal Transformer for Traffic Flow Forecasting},year={2025},volume={26},number={9},pages={13144-13158},doi={10.1109/TITS.2025.3583391},}
[AAAI’25] ST-ReP
ST-ReP: Learning Predictive Representations Efficiently for Spatial-Temporal Forecasting
Qi Zheng, Zihao Yao, and Yaying Zhang*
Proceedings of the AAAI Conference on Artificial Intelligence, 2025
@article{STReP,title={ST-ReP: Learning Predictive Representations Efficiently for Spatial-Temporal Forecasting},volume={39},doi={10.1609/aaai.v39i12.33465},number={12},journal={Proceedings of the AAAI Conference on Artificial Intelligence},author={Zheng, Qi and Yao, Zihao and Zhang, Yaying},year={2025},pages={13419-13427},}
[DASFAA’25] CRS-Mamba
Correlation-Aware Reordered Scanning Mamba for Multivariate Time Series Forecasting
Zihao Yao, Qi Zheng, and Yaying Zhang
In Database Systems for Advanced Applications: 30th International Conference, DASFAA 2025, Singapore, Singapore, May 26–29, 2025, Proceedings, Part II, Singapore, Singapore, 2025
The fundamental challenge in Multivariate Time Series forecasting is effectively modeling complex temporal dependencies and variable correlation. Transformer-based models achieve breakthroughs but face challenges with quadratic complexity and permutation invariant bias. A recent model, Mamba, has emerged as a competitive alternative. However, we observe that the issue of scan order sensitivity is not well concerned. In this study, we propose a novel Correlation-aware Reordered Scanning Mamba, namely CRS-Mamba, for multivariate time series forecasting. Specifically, we leverage the downsampling technique to model temporal dependencies. Then, a bidirectional Mamba layer is introduced to extract inter-variate correlations. Moreover, we propose Dimensionality Reduction Scan Algorithm to alleviate scanning sensitivity problem of Mamba. Extensive evaluations show that our approach secures superior performance in prediction accuracy on various datasets. Moreover, ablation studies demonstrate the interpretability of CRS-Mamba.
@inproceedings{CRS-Mamba,author={Yao, Zihao and Zheng, Qi and Zhang, Yaying},title={Correlation-Aware Reordered Scanning Mamba for Multivariate Time Series Forecasting},year={2025},isbn={978-981-95-3829-4},publisher={Springer-Verlag},address={Berlin, Heidelberg},url={https://doi.org/10.1007/978-981-95-3830-0_29},doi={10.1007/978-981-95-3830-0_29},booktitle={Database Systems for Advanced Applications: 30th International Conference, DASFAA 2025, Singapore, Singapore, May 26–29, 2025, Proceedings, Part II},pages={425–435},numpages={11},keywords={Multivariate Time Series, Forecasting, Mamba},location={Singapore, Singapore},}
2024
[SMC’24] TMBEA
Temporal MLP Bridges the Gap Between Embedding and Attention for Multivariate Time Series Forecasting
Zhinan Xie, Qi Zheng*, and Yaying Zhang
In 2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2024
@inproceedings{TMBEA,author={Xie, Zhinan and Zheng, Qi and Zhang, Yaying},booktitle={2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC)},title={Temporal MLP Bridges the Gap Between Embedding and Attention for Multivariate Time Series Forecasting},year={2024},volume={},number={},pages={2373-2378},keywords={Bridges;Attention mechanisms;Atmospheric modeling;Time series analysis;Resists;Feature extraction;Air pollution;Robustness;Forecasting;Cybernetics},doi={10.1109/SMC54092.2024.10831557},}
2023
[DASFAA’23] TAGnn
TAGnn: Time Adjoint Graph Neural Network for Traffic Forecasting
Qi Zheng and Yaying Zhang*
In Database Systems for Advanced Applications: 28th International Conference, DASFAA 2023, Tianjin, China, April 17–20, 2023, Proceedings, Part I, 2023
@inproceedings{TAGnn,author={Zheng, Qi and Zhang, Yaying},title={TAGnn: Time Adjoint Graph Neural Network for Traffic Forecasting},year={2023},doi={10.1007/978-3-031-30637-2_24},booktitle={Database Systems for Advanced Applications: 28th International Conference, DASFAA 2023, Tianjin, China, April 17–20, 2023, Proceedings, Part I},pages={369–379},}
@inproceedings{STHGCN,author={Zhang, Canyang and Zheng, Qi and Zhang, Yaying},booktitle={2023 IEEE International Conference on Big Data (BigData)},title={Spatial-Temporal Flow Holistic Interaction Graph Convolution Network for Bidirectional Traffic Flow Forecasting},year={2023},volume={},number={},pages={1262-1268},keywords={Deep learning;Convolution;Estimation;Predictive models;Big Data;Feature extraction;Cognition},doi={10.1109/BigData59044.2023.10386746},}
2022
[TBD’22] DSTAGCN
DSTAGCN: Dynamic Spatial-Temporal Adjacent Graph Convolutional Network for Traffic Forecasting
@inproceedings{FCLSTM,author={Yu, Huiyun and Zheng, Qi and Qian, Shuyun and Zhang, Yaying},booktitle={2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC)},title={A Fuzzy-based Convolutional LSTM Network Approach for Citywide Traffic Flow Prediction},year={2022},volume={},number={},pages={3360-3367},keywords={Deep learning;Training;Uncertainty;Smart cities;Neural networks;Transportation;Predictive models},doi={10.1109/ITSC55140.2022.9922491},}
