A Simple State Space Model Excels at Multivariate Time Series Classification

This is a research paper presenting a systematic empirical study and introducing new model variants, challenging prevailing architectural assumptions in sequence modeling.

Challenging the Complexity-Performance Paradigm

The paper's core contribution is its direct challenge to the assumption that increased model complexity necessarily yields better performance for time-series classification (TSC). While Mamba architectures, with their input-dependent state transitions, have shown strong results in many areas, this work finds that the added complexity is not justified for TSC. The simpler, linear time-invariant S4D models perform better. This suggests that the specific mechanisms making Mamba effective for tasks like language modeling may not be as critical, or may even be detrimental, for the characteristics of TSC datasets.

Methodology: A Rigorous Comparative Lens

The authors ground their claims through a large-scale, systematic evaluation across diverse benchmarks. They compare diagonal SSMs (S4D) and input-dependent SSMs (Mamba family) on the MONSTER benchmark (notably massive in scale) and the UEA archive, against 15 baselines. This broad comparative approach across 59 datasets provides robust evidence that the finding is not an artifact of a specific dataset but a general trend, strengthening the paper's central argument.

The MS4/MS4N Practical Innovations

Building on the empirical finding, the authors propose two practical model modifications:

• MS4 introduces a linear input projection and a channel-mixing mechanism to the base S4D model.
• MS4N adds a normalization variant to stabilize state dynamics.
  These are described as lightweight modifications with negligible computational overhead. The key result is that these enhanced models not only maintain the efficiency of S4D but also achieve or surpass the performance of models with 2x to 10x more parameters. This demonstrates that targeted, minor architectural tweaks to a simpler foundation can be more effective than scaling up more complex models.

Implications for the Field

The study's outcomes have two main implications. First, they urge a re-evaluation of model selection for TSC, suggesting that researchers and practitioners should prioritize simpler, efficient SSMs over defaulting to more complex contemporary architectures. Second, they highlight that the optimal sequence modeling architecture is task-dependent; what advances performance in one domain (e.g., language) does not automatically translate to another. The work positions MS4/MS4N as new strong baselines and a compelling direction for efficient time-series analysis.