DPEM:Dual-Perspective Enhanced Mamba for Multivariate Time Series Forecasting

DPEM Model Architecture

Contributions:

-- A Dual-Perspective Enhanced Mamba DPEM framework is introduced for effective multivariate time series forecasting. -- DPEM utilizes a dual-perspective approach to separately model temporal dependencies and spatial dependencies, enabling comprehensive spatiotemporal feature extraction. -- We designed a cross attention mechanism that enables bidirectional interactions between temporal-spatial and spatial-temporal features, resulting in a deeper fusion of the spatiotemporal dependencies. -- Extensive experiments conducted on 8 real-world datasets demonstrate superior performance compared to SOTA baseline models.

Getting Start:

Installation

pip install -r requirements.txt

scikit-learn==1.3.0

numpy==1.26.4

matplotlib==3.7.0

torch==2.0.1

reformer-pytorch==1.4.4

mamba-ssm==1.2.0

Datasets

Detailed dataset descriptions. $Dim$ denotes the variate number of each dataset. $Dataset$ $Size$ denotes the total number of time points in (Train, Validation, Test) split respectively. $Prediction$ $Length$ denotes the future time points to be predicted. $Frequency$ denotes the sampling interval of time points.

Train and evaluate

# Example: Exchange
bash ./scripts/multivariate_forecasting/Exchange/exchange_96.sh

Main results:

Acknowledgement:

We are grateful for the following awesome projects when implementing DPEM:

• iTransformer
• Mamba
• S-Mamba
• [Time-Series-Library] (https://github.com/thuml/Time-Series-Library)

Citation

@article{hou2025dpem, title={DPEM: Dual-Perspective Enhanced Mamba for multivariate time series forecasting}, author={Hou, Rui and Liu, Qiao and He, Peng and Liu, Yao and Huang, Yanwen and Xie, Jun and Xie, Yang and Dai, Tingting}, journal={Information Fusion}, pages={103250}, year={2025}, publisher={Elsevier} }