MambaSL: Exploring Single-Layer Mamba for Time Series Classification
MambaSL: Exploring Single-Layer Mamba for Time Series Classification
Authors: Yoo-Min Jung, Leekyung Kim Date: 2026-04-16 Paper ID: openalex:2604.15174
Summary
MambaSL explores the potential of single-layer selective state space models (SSMs) as a lightweight backbone for time series classification (TSC). The authors redesign selective SSM and projection layers guided by domain-specific hypotheses and provide a robust, unified evaluation protocol across all 30 University of East Anglia (UEA) datasets. Empirical results show that MambaSL achieves state-of-the-art performance, highlighting the effectiveness of highly efficient, shallow SSMs for sequential classification tasks.
Key Contributions
- Introduces MambaSL, a minimally redesigned single-layer selective state space model framework for time series classification.
- Establishes a unified, reproducible evaluation protocol for time series classification across all 30 UEA archive datasets.
- Demonstrates state-of-the-art classification performance using a lightweight, single-layer Mamba backbone compared to 20 strong existing baselines.
Open Questions & Future Work
Archivist Review
I have approved the UEA archive as a critical dataset for time series classification benchmarking and the open question regarding SSM time-variance configurations. The MambaSL architecture itself was rejected as a local implementation detail, as it is a specific instantiation rather than a generalized methodological advancement. The review process prioritized foundational research questions over model-specific frameworks.
Approved Open Questions
- Optimal SSM Time-Variance Configuration: Identifying the correct balance of time-invariance versus time-variance is critical for adapting Mamba-based models to time series data, where signal properties vary significantly. Misconfiguring these SSM components can impair classification accuracy.
Rejected Candidates
- [concept] MambaSL (
mambasl) - paper_local: This is a specific framework architecture and implementation rather than a general-purpose, reusable concept in the broader time series field.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.15174
- paper_id
- 2604.15174
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
- uea-archive
- concept_slugs
-
- dataset_slugs
- uea-archive
- skill
- TimeSeriesSkill
- created_at
- 2026-04-19T05:44:27Z