Skip to content

DC-Mamber: A Dual Channel Prediction Model Based on Mamba and Linear Transformer for Multivariate Time Series Forecasting

Home / Papers / DC-Mamber: A Dual Channel Prediction Model Based on Mamba and Linear Transformer for Multivariate Time Series Forecasting

DC-Mamber: A Dual Channel Prediction Model Based on Mamba and Linear Transformer for Multivariate Time Series Forecasting

Authors: Bing Fan, Shusen Ma, Yun‐Bo Zhao, Gang Xu Date: 2026-04-21 Paper ID: openalex:2507.04381

Summary

DC-Mamber addresses the trade-off between local temporal modeling and global context in multivariate time series forecasting by combining state space models (Mamba) with linear Transformers. The architecture uses a dual-channel approach to simultaneously extract intra-variable patterns and global dependencies, mitigating the quadratic complexity of standard Transformers while overcoming Mamba’s limitations in global context sensitivity. A specialized feature-fusion module then synthesizes these distinct representations to produce more accurate predictions. Experimental results demonstrate that this hybrid approach outperforms existing channel-independent and channel-mixing baselines.

Key Contributions

  • Proposes DC-Mamber, a dual-channel architecture that integrates Mamba for local temporal feature extraction and a linear Transformer for global dependency modeling.
  • Implements a Dual-Channel Embedding Layer that processes input streams through distinct temporal strategies to capture both intra-variable and inter-variable patterns.
  • Achieves state-of-the-art forecasting accuracy across four multivariate time series datasets by balancing linear complexity with global contextual integration.

Open Questions & Future Work

Archivist Review

I have approved one open question concerning the optimal fusion of heterogeneous features in dual-channel forecasting architectures, as this remains a broader research challenge beyond this specific paper. The model architecture itself, DC-Mamber, and its specific embedding layer were rejected as they are implementation-specific contributions rather than foundational concepts for the vault.

Approved Open Questions

  • Optimal Feature Fusion Methods: Establishing robust fusion mechanisms is critical for hybrid architectures that attempt to merge disparate feature representations, which is a common challenge in multi-modal or multi-branch neural network designs.

Rejected Candidates

  • [concept] DC-Mamber (dc-mamber) - not_novel: The proposed model is a specific architecture implementation that does not represent a sufficiently novel or generalizable architectural paradigm beyond what is already captured by existing hybrid state space-transformer models.
  • [concept] Dual-Channel Embedding Layer (dual-channel-embedding-layer) - paper_local: This is an implementation detail specific to the DC-Mamber architecture rather than a widely reusable mechanism.

Metadata & Links

url
https://arxiv.org/abs/2507.04381
paper_id
2507.04381
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-24T05:50:16Z