DEMamba: Decoupled Enhanced State Space Models with Selective Mechanisms for Multivariate Time Series Forecasting

Junluo Zheng, Yang Liu, Jianyong Chen

Keywords: Multivariate Time Series Forecasting, State Space Models, Selective Mechanis

Abstract: Multivariate time series forecasting (MTSF) is an important and forefront task in many real-world applications. Recently, Mamba has emerged as a powerful alternative to Transformer-based models in sequential modeling, leveraging its selective mechanism to achieve high accuracy and linear complexity. However, existing efforts to apply Mamba to MTSF face limitations, particularly in effectively and efficiently capturing temporal and cross-variate dependencies. In this work, we analyze these limitations and propose DEMamba, an enhanced model tailored for MTSF. Specifically, we redefine the roles of 1D convolution and selective SSMs (S6) in the original Mamba block and design the Time-Variable Decoupled Scanning (TVDS) Mamba block. It decouples the learning of different dimensions within a single Mamba block, with 1D convolution capturing temporal dependencies and S6 capturing cross-variate dependencies. Furthermore, we capture intra-patch feature dependencies with a feed-forward network. These designs significantly enhance the ability of DEMamba to handle complex dependencies. Extensive experimental results on eight real-world datasets demonstrate the effectiveness of DEMamba against previous state-of-the-arts. Code is available at this repository: XX.XXX.

Area: Learning and Adaptation (LEARN)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 963