Skip to main navigation Skip to search Skip to main content

ASM-Former: Adaptive Spectral Mixing for Multivariate Time Series Classification

  • Han Fu
  • , Chengcheng Feng
  • , Tianjiao Li*
  • , Xin Qin
  • , Wenjie Wang
  • , Hui Liu
  • , Xu Cheng
  • *Corresponding author for this work
  • Tianjin University of Technology

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

Abstract

The drive towards longer sequences in Multivariate Time Series Classification (MTSC) intensifies the fundamental trade-off between computational efficiency and representational flexibility, creating a pressing need for novel methods that can transcend static filters and adapt to non-stationary spectral dynamics. Attention-based models provide instance-specific adaptation, yet they incur high computational costs that pose scalability challenges. Frequency-domain methods provide a scalable solution by leveraging the Fast Fourier Transform (FFT). However, they typically rely on static, data-independent global filters, limiting their ability to model signal-specific spectral shifts. To overcome this limitation, we propose ASM-Former, a hierarchical architecture featuring a novel Adaptive Spectral Mixing (ASM) module. The ASM module introduces a dynamic, instance-aware filtering mechanism based on a Basis Decomposition strategy. It learns a compact set of shared spectral basis filters and synthesizes a sample-specific filter through a lightweight linear combination, conditioned on the input. This approach enables adaptive feature extraction in the frequency domain without incurring prohibitive parameter overhead. Extensive experiments on 25 benchmark datasets demonstrate that ASM-Former achieves competitive accuracy against strong baselines. Further analyses confirm its computational efficiency, highlighting its scalability for long-sequence modeling. Our work validates that basis-driven dynamic spectral mixing is an effective strategy for enhancing feature representation in MTSC.

Original languageEnglish
Title of host publicationProceedings of the 2026 29th International Conference on Computer Supported Cooperative Work in Design (CSCWD)
Publication date2026
Edition2026
Pages3085-3090
DOIs
Publication statusPublished - 2026
Event29th International Conference on Computer Supported Cooperative Work in Design, CSCWD 2026 - Fuzhou, China
Duration: 13 May 2026 → 15 May 2026

Conference

Conference29th International Conference on Computer Supported Cooperative Work in Design, CSCWD 2026
Country/TerritoryChina
CityFuzhou
Period13/05/2026 → 15/05/2026

Keywords

  • Adaptive Filtering
  • Basis Decomposition
  • Frequency Domain
  • Instance-Aware Modeling
  • Multivariate Time Series Classification

Fingerprint

Dive into the research topics of 'ASM-Former: Adaptive Spectral Mixing for Multivariate Time Series Classification'. Together they form a unique fingerprint.

Cite this