EUVIMEDEUVIMED entdecken
Crossref

Time-MaC: Multimodal Time Series Forecasting With Reconstruction-Oriented MAE and Coupled-Mamba Fusion

Chenxu Deng, Hui Wang, Feng Xu, Guoliang Luo, Zizhu Fan

IEEE Signal Processing Letters · 2026 · 33 · 4146-4150

DOI: 10.1109/lsp.2026.3739688 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Chenxu Deng und 4 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 28 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Ministry of Education Humanities and Social Sciences Project, Open Project of Shandong Provincial Key Laboratory of Advanced Computing.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Aligning temporal dynamics with structurally informative visual cues remains challenging in multimodal time-series forecasting. Time-MaC couples temporal memory with reconstruction-oriented visual representations through an adapted Coupled-Mamba fusion module for multivariate forecasting. The framework addresses practical integration issues in Time-VLM-style pipelines by (i) introducing a temporal memory branch for long-range signal preservation, (ii) converting multivariate time series into image-oriented inputs and extracting reconstruction-conditioned embeddings with a pretrained MAE backbone, and (iii) integrating temporal and visual cues through coupled state-space fusion. By using reconstruction-oriented visual features, Time-MaC emphasizes structural patterns such as periodic textures, spectral variations, and trend transitions that are often weakened in purely numerical forecasting pipelines. The adapted fusion design further preserves modality-specific dynamics before the final prediction. Furthermore, an optional post-hoc online calibration plugin is implemented to support adaptive, coverage-oriented prediction intervals under chronological evaluation. Experiments on seven standard long-term forecasting benchmarks show that Time-MaC achieves competitive overall forecasting performance and improves the robustness of the complete model under severe missing-observation settings.

Förderung

  • National Natural Science Foundation of China
  • Ministry of Education Humanities and Social Sciences Project
  • Open Project of Shandong Provincial Key Laboratory of Advanced Computing

Zitieren

APA

Deng, C., Wang, H., Xu, F., Luo, G., & Fan, Z. (2026). Time-MaC: Multimodal Time Series Forecasting With Reconstruction-Oriented MAE and Coupled-Mamba Fusion. IEEE Signal Processing Letters, 33, 4146-4150. https://doi.org/10.1109/lsp.2026.3739688

Vancouver

Deng C, Wang H, Xu F, Luo G, Fan Z. Time-MaC: Multimodal Time Series Forecasting With Reconstruction-Oriented MAE and Coupled-Mamba Fusion. IEEE Signal Processing Letters. 2026;33:4146-4150. doi:10.1109/lsp.2026.3739688

BibTeX
@article{Deng2026,
  author = {Deng, Chenxu and Wang, Hui and Xu, Feng and Luo, Guoliang and Fan, Zizhu},
  title = {Time-MaC: Multimodal Time Series Forecasting With Reconstruction-Oriented MAE and Coupled-Mamba Fusion},
  journal = {IEEE Signal Processing Letters},
  year = {2026},
  volume = {33},
  pages = {4146--4150},
  doi = {10.1109/lsp.2026.3739688}
}
Kostenlos starten →