A Simple State Space Model Excels at Multivariate Time Series Classification
A new paper says simpler structured state space models beat Mamba-style variants on multivariate time series classification across 59 datasets.
Intelligence analysis by GPT-5.4 Mini
The paper argues that more complex input-dependent state space models are not necessary for top performance in multivariate time series classification. In its tests, a simpler diagonal SSM and two lightweight variants, MS4 and MS4N, were faster and often more accurate than Mamba-based models.
A group of researchers looked at computer models that try to understand patterns in many signals over time, like heartbeats or sensor readings. They expected the fancier models to win, but the simpler ones did better.
They also tried two small upgrades to the simple model. Those upgrades helped it stay steady and work well without making it much bigger.
It is a bit like finding out that a lighter bicycle can go faster than a heavier one on the same road. The paper says bigger and more complicated is not always better.
Analysis
What the paper studies
The authors examine structured state space models for multivariate time series classification, a setting where models must recognize patterns across many signals over time. They focus on two broad families: diagonal SSMs such as S4D, and input-dependent SSMs from the Mamba family.
Main finding
The paper says the expected winner was not the most complex model. Instead, S4D "consistently outperforms" Mamba-based variants on both accuracy and efficiency, according to the abstract. That result is presented as a challenge to the idea that adding more architectural complexity automatically improves time-series classification.
What they add
Building on that result, the authors introduce MS4, which adds a linear input projection and channel-mixing mechanism to S4D. They also propose MS4N, a normalized version that stabilizes state dynamics with little added cost.
Reported results
The abstract says the models were evaluated on 59 datasets across MONSTER and the UEA benchmark, compared against 15 baselines. MS4 and MS4N are described as consistently outperforming Mamba-based models while staying more efficient. The paper also says MS4N matches or exceeds competing deep learning models that are about 2x and 10x larger in parameter count.
Bottom line
The paper positions lightweight structured SSMs as a practical alternative to scaling up complexity for multivariate time series classification. Its central claim is not that bigger models are useless, but that simpler structured designs can be stronger choices for this task.
Key points
- The paper studies multivariate time series classification with structured state space models.
- It says simpler S4D models beat Mamba-style variants in both accuracy and efficiency.
- The authors propose MS4 and MS4N as lightweight S4D-based improvements.
- The evaluation spans 59 datasets across MONSTER and the UEA benchmark.
- The abstract frames lightweight structured SSMs as a strong alternative to adding complexity.
