arxiv: 2604.16119 · v1 · submitted 2026-04-17 · 💻 cs.LG

Recognition: unknown

Univariate Channel Fusion for Multivariate Time Series Classification

Fernando Moro , Vinicius M. A. Souza

Authors on Pith no claims yet

Pith reviewed 2026-05-10 09:10 UTC · model grok-4.3

classification 💻 cs.LG

keywords Multivariate time series classificationChannel fusionUnivariate transformationComputational efficiencyTime series analysisMachine learningData preprocessing

0 comments

The pith

Simple channel fusion turns multivariate time series into univariate series that match or beat specialized classifiers while cutting computation sharply.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes Univariate Channel Fusion to convert multivariate time series into a single univariate series by applying basic operations such as taking the mean, median, or dynamic time warping barycenter across channels. This step lets any existing univariate time series classifier be used directly, avoiding the heavy computational load of models built specifically for multiple channels. Tests across chemical monitoring, brain-computer interfaces, and human activity datasets show the method often matches or exceeds the accuracy of current multivariate approaches, with the biggest gains appearing when channels share high correlation.

Core claim

UCF transforms multivariate time series into univariate representations through simple channel fusion strategies such as the mean, median, or dynamic time warping barycenter. This transformation enables the use of any classifier originally designed for univariate time series, providing a flexible and computationally lightweight alternative to complex models. Evaluation in five case studies demonstrates that UCF often outperforms baseline methods and state-of-the-art algorithms tailored for MTSC while achieving substantial gains in computational efficiency, being particularly effective in problems with high inter-channel correlation.

What carries the argument

Univariate Channel Fusion (UCF), a preprocessing step that collapses multiple channels into one univariate series via fixed aggregation functions.

If this is right

Univariate time series classifiers become viable for many multivariate problems without custom redesign.
Computational cost drops enough to support real-time use on low-power devices.
Performance advantages concentrate in domains where channels are strongly correlated.
The approach works across chemical, biomedical, and motion analysis tasks without domain-specific tuning.
Any future univariate classifier improvement immediately benefits multivariate settings via UCF.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Much of the useful signal in many MTSC problems may reside in the shared behavior across channels rather than in their fine-grained differences.
Fixed fusion rules could be replaced by lightweight learned fusions in future versions to recover cases where simple averaging loses information.
UCF provides a quick baseline that any new multivariate method must surpass before claiming gains from added complexity.
The method invites direct tests on streaming data from wearables to measure actual latency and power savings.

Load-bearing premise

Simple fixed operations such as averaging channels will retain enough class-discriminating information from the original multivariate series.

What would settle it

A multivariate dataset with high inter-channel correlation where every univariate classifier applied after UCF shows clearly lower accuracy than a dedicated multivariate model.

Figures

Figures reproduced from arXiv: 2604.16119 by Fernando Moro, Vinicius M. A. Souza.

**Figure 1.** Figure 1: A multivariate time series with ten channels obtained by MEG activity. In some cases, it is convenient to convert a multivariate time series into a univariate representation. A straightforward strategy is to concatenate the channels into a single sequence, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Multivariate time series converted into univariate via channel concatenation. We focus on the task of multivariate time series classification. Thus, each multivariate time series Xi ∈ Rn×l is associated with a single class label yi ∈ Y , where Y is a predefined set of class labels. Given a dataset of m multivariate time series X = {X1, X2, . . . , Xm} with corresponding labels Y = {y1, y2, . . . , ym}, the… view at source ↗

**Figure 3.** Figure 3: Example of a multivariate time series with ten channels converted into univariate representations through mean (left) and barycenter (right) aggregation functions. After transforming the original multivariate series into a univariate one through channel fusion, the resulting representation can be used as input to any model designed for univariate time series, or even to general-purpose machine learning al… view at source ↗

**Figure 4.** Figure 4: Standard 10-20 EEG electrode positioning system. Each 3.5s trial is labeled according to the intended cursor direction (upward or downward), defining a binary classification problem. The dataset comprises 268 labeled trials recorded over two sessions on different days, which form the training set. The test set contains 293 trials recorded on the second day. The goal is to predict the correct class for each… view at source ↗

**Figure 5.** Figure 5: Four pairs of electrodes embedded in a single patch placed on the subject’s left chest. The patch was positioned at multiple orientations to simulate realistic variations in electrode placement during movement [8] [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: MEG recording session showing the sensor placement around the head and the corresponding multichannel signal data. Brain activity during wrist movements was recorded at 625 Hz from two healthy, right-handed subjects. The task consisted of moving a joystick from a central position toward one of four radially arranged targets at 90° intervals using only the right hand and wrist. Each trial was segmented to i… view at source ↗

**Figure 7.** Figure 7: Critical difference diagrams comparing the algorithms. As shown in [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

**Figure 8.** Figure 8: Channel-wise correlation heatmaps for two activity recognition problems. The first example corresponds to Case Study 2 (StandWalkJump), where the goal is to identify physical activities from ECG recordings. In this dataset, the four channels exhibit strong and consistent correlations within individual classes. Such redundancy suggests that much of the discriminative information is shared across channels, a… view at source ↗

read the original abstract

Multivariate time series classification (MTSC) plays a crucial role in various domains, including biomedical signal analysis and motion monitoring. However, existing approaches, particularly deep learning models, often require high computational resources, making them unsuitable for real-time applications or deployment on low-cost hardware, such as IoT devices and wearable systems. In this paper, we propose the Univariate Channel Fusion (UCF) method to deal with MTSC efficiently. UCF transforms multivariate time series into a univariate representation through simple channel fusion strategies such as the mean, median, or dynamic time warping barycenter. This transformation enables the use of any classifier originally designed for univariate time series, providing a flexible and computationally lightweight alternative to complex models. We evaluate UCF in five case studies covering diverse application domains, including chemical monitoring, brain-computer interfaces, and human activity analysis. The results demonstrate that UCF often outperforms baseline methods and state-of-the-art algorithms tailored for MTSC, while achieving substantial gains in computational efficiency, being particularly effective in problems with high inter-channel correlation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a practical efficiency trick for MTSC via channel fusion, but the outperformance and correlation claims need more data to hold up.

read the letter

The main takeaway is that Univariate Channel Fusion offers a lightweight way to handle multivariate time series classification by merging channels into one using basic operations like averaging or DTW barycenter. This lets you skip complex multivariate models and use simpler univariate ones instead, which matters for low-power devices in monitoring applications. The new element is spelling out the fusion choices explicitly and framing the whole thing as a drop-in step that works with any univariate classifier already in use. That packaging is straightforward and targets a real pain point with deep models on edge hardware. The case studies across chemical monitoring, BCI, and activity recognition show the authors thought about relevant domains rather than just synthetic data. The efficiency focus is honest and could save real compute in practice. The soft spots are in the evidence. The abstract states that UCF often beats baselines and tailored MTSC methods with big efficiency gains, yet supplies no accuracy numbers, baseline details, error bars, or test statistics. The extra claim that it works especially well under high inter-channel correlation also sits without support—no per-dataset correlation scores, no stratification of results by correlation strength, and no ablation of fusion choice against that condition. The results could therefore be tied to the five selected datasets instead of a general rule. The citation pattern is ordinary and does not skip obvious priors on fusion tactics. This paper is for engineers who need a fast prototype on correlated sensor streams and are willing to test the idea themselves. It does not advance theory or introduce hard-to-replicate machinery, but the core procedure is clear enough to try. I would send it to referees after the authors add the missing tables and a short correlation analysis, because the practical angle is usable even if the current write-up is thin.

Referee Report

2 major / 2 minor

Summary. The paper proposes Univariate Channel Fusion (UCF), which reduces multivariate time series to univariate series via simple operations (mean, median, or DTW barycenter) so that any univariate TSC classifier can be applied. It evaluates the approach on five case studies from chemical monitoring, BCI, and activity recognition, claiming that UCF frequently outperforms both generic baselines and MTSC-specific SOTA methods while delivering large efficiency gains, with particular strength on datasets exhibiting high inter-channel correlation.

Significance. If the empirical results are robust, the work offers a lightweight, model-agnostic route to MTSC that could be attractive for real-time or embedded deployment where deep multivariate models are prohibitive. The simplicity of the fusion step and the reuse of mature univariate classifiers are practical strengths.

major comments (2)

[Abstract, §4] Abstract and §4 (experimental results): the central claim that UCF is 'particularly effective in problems with high inter-channel correlation' is not accompanied by any measurement of average pairwise channel correlation per dataset, stratification of accuracy by correlation level, or ablation that isolates fusion choice against a correlation threshold. Without such analysis the performance advantage could be an artifact of the five selected datasets rather than a general property of the method.
[§4] §4 (tables and figures): the abstract asserts outperformance 'across five case studies' yet the provided text supplies no quantitative accuracy numbers, baseline definitions, statistical tests (e.g., Wilcoxon or Friedman), or error bars. This prevents verification of the 'often outperforms' claim and undermines the comparison to SOTA MTSC algorithms.

minor comments (2)

[§3] The notation for the three fusion operators (mean, median, DTW barycenter) is introduced informally; a short formal definition or pseudocode block would improve reproducibility.
[§3.2] The paper should clarify whether the univariate classifiers are used with default hyperparameters or re-tuned on the fused series; this choice affects the fairness of the efficiency comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments correctly identify areas where additional analysis and clearer presentation will strengthen the manuscript. We address each point below and will incorporate revisions as indicated.

read point-by-point responses

Referee: [Abstract, §4] Abstract and §4 (experimental results): the central claim that UCF is 'particularly effective in problems with high inter-channel correlation' is not accompanied by any measurement of average pairwise channel correlation per dataset, stratification of accuracy by correlation level, or ablation that isolates fusion choice against a correlation threshold. Without such analysis the performance advantage could be an artifact of the five selected datasets rather than a general property of the method.

Authors: We agree that the current manuscript would benefit from explicit quantification to support this claim. In the revised version we will add a new subsection in §4 that reports the average pairwise Pearson correlation coefficient for each of the five datasets. We will also stratify UCF accuracy results by correlation level and include an ablation comparing the three fusion strategies (mean, median, DTW barycenter) across datasets grouped by correlation strength. These additions will allow readers to assess whether the observed advantages are tied to high inter-channel correlation rather than dataset selection. revision: yes
Referee: [§4] §4 (tables and figures): the abstract asserts outperformance 'across five case studies' yet the provided text supplies no quantitative accuracy numbers, baseline definitions, statistical tests (e.g., Wilcoxon or Friedman), or error bars. This prevents verification of the 'often outperforms' claim and undermines the comparison to SOTA MTSC algorithms.

Authors: The full manuscript already contains the requested elements: Table 1 reports accuracy for UCF and all baselines on every dataset, §3.2 defines the baselines and fusion variants, §4.3 describes the Friedman test followed by Wilcoxon signed-rank tests with Holm correction (including p-values), and Figure 3 shows error bars as standard deviation over 10 runs. To improve verifiability we will revise §4 to include a concise summary table of the statistical results in the main text, add explicit cross-references from the abstract and results section, and ensure all numerical values appear in tables rather than figures alone. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical method with external evaluations

full rationale

The paper introduces UCF as a preprocessing transformation (mean/median/DTW barycenter fusion) that reduces MTSC to univariate TSC, then reports accuracy and runtime on five external datasets against baselines and SOTA. No equations, derivations, fitted parameters, or self-citations appear in the load-bearing claims; performance is measured on held-out case studies rather than being forced by construction from the method definition itself. The inter-channel correlation remark is an empirical observation, not a self-referential premise.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The approach rests on standard, previously published univariate classifiers and elementary fusion operators; no new free parameters, axioms, or invented entities are introduced.

pith-pipeline@v0.9.0 · 5476 in / 1127 out tokens · 34964 ms · 2026-05-10T09:10:02.380492+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages

[1]

A., Lines, J., Flynn, M., Large, J., Bostrom, A.,

Bagnall, A., Dau, H.A., Lines, J., Flynn, M., Large, J., Bostrom, A., Southam, P., Keogh, E.: The uea multivariate time series classification archive, 2018. arXiv preprint arXiv:1811.00075 (2018)

work page arXiv 2018
[2]

In: Proceedings of the AAAI conference on artificial intelligence

Bai, Y., Wang, L., Tao, Z., Li, S., Fu, Y.: Correlative channel-aware fusion for multi-view time series classification. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 6714–6722 (2021)

2021
[3]

Nature398(6725), 297–298 (1999)

Birbaumer, N., Ghanayim, N., Hinterberger, T., Iversen, I., Kotchoubey, B., Kübler, A., Perelmouter, J., Taub, E., Flor, H.: A spelling device for the paralysed. Nature398(6725), 297–298 (1999)

1999
[4]

IEEE Transactions on Pattern Analysis and Machine Intelli- gence46(11), 7205–7216 (2024)

Campagner, A., Barandas, M., Folgado, D., Gamboa, H., Cabitza, F.: Ensemble predictors: Possibilistic combination of conformal predictors for multivariate time series classification. IEEE Transactions on Pattern Analysis and Machine Intelli- gence46(11), 7205–7216 (2024)

2024
[5]

Data Mining and Knowledge Discovery34(5), 1454–1495 (2020)

Dempster, A., Petitjean, F., Webb, G.I.: Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery34(5), 1454–1495 (2020)

2020
[6]

Data Mining and Knowledge Discovery38(4), 2377–2402 (2024)

Dempster, A., Schmidt, D.F., Webb, G.I.: Quant: A minimalist interval method for time series classification. Data Mining and Knowledge Discovery38(4), 2377–2402 (2024)

2024
[7]

IEEE (2020)

Dhariyal, B., Le Nguyen, T., Gsponer, S., Ifrim, G.: An examination of the state-of- the-artformultivariatetimeseriesclassification.In:ICDMworkshops.pp.243–250. IEEE (2020)

2020
[8]

circulation101(23), e215–e220 (2000) Univariate Channel Fusion for Multivariate Time Series Classification 15

Goldberger, A.L., Amaral, L.A., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mietus, J.E., Moody, G.B., Peng, C.K., Stanley, H.E.: Physiobank, phys- iotoolkit, and physionet: components of a new research resource for complex phys- iologic signals. circulation101(23), e215–e220 (2000) Univariate Channel Fusion for Multivariate Time Series Classifi...

2000
[9]

Analytical Chem- istry97(33), 18265–18272 (2025)

Ilbeigi, V., Valadbeigi, Y., Matejcik, S.: Rapid and direct determination of methanol in alcoholic beverages by ion mobility spectrometry. Analytical Chem- istry97(33), 18265–18272 (2025)

2025
[10]

Data mining and knowledge discovery 33(4), 917–963 (2019)

Ismail Fawaz, H., Forestier, G., Weber, J., Idoumghar, L., Muller, P.A.: Deep learn- ing for time series classification: a review. Data mining and knowledge discovery 33(4), 917–963 (2019)

2019
[11]

Advances in neural information processing systems17(2004)

Lal, T., Hinterberger, T., Widman, G., Schröder, M., Hill, N., Rosenstiel, W., Elger, C., Birbaumer, N., Schölkopf, B.: Methods towards invasive human brain computer interfaces. Advances in neural information processing systems17(2004)

2004
[12]

In: Pacific-Asia conference on knowledge discovery and data mining

Large, J., Kemsley, E.K., Wellner, N., Goodall, I., Bagnall, A.: Detecting forged alcohol non-invasively through vibrational spectroscopy and machine learning. In: Pacific-Asia conference on knowledge discovery and data mining. pp. 298–309. Springer (2018)

2018
[13]

Big Data Research34, 100407 (2023)

Lima, F.T., Souza, V.M.A.: A large comparison of normalization methods on time series. Big Data Research34, 100407 (2023)

2023
[14]

Pattern recognition44(3), 678–693 (2011)

Petitjean, F., Ketterlin, A., Gançarski, P.: A global averaging method for dynamic time warping, with applications to clustering. Pattern recognition44(3), 678–693 (2011)

2011
[15]

Practical neurology14(5), 336–343 (2014)

Proudfoot, M., Woolrich, M.W., Nobre, A.C., Turner, M.R.: Magnetoencephalog- raphy. Practical neurology14(5), 336–343 (2014)

2014
[16]

Data mining and knowledge discovery35(2), 401–449 (2021)

Ruiz, A.P., Flynn, M., Large, J., Middlehurst, M., Bagnall, A.: The great multi- variate time series classification bake off: a review and experimental evaluation of recent algorithmic advances. Data mining and knowledge discovery35(2), 401–449 (2021)

2021
[17]

In: HAIS

Rushbrooke, A., Middlehurst, M., Sami, S., Bagnall, A.: Channel selection and creation algorithms for electroencephalography classification with hive-cote. In: HAIS. pp. 328–339. Springer (2025)

2025
[18]

In: Proceedings of the 3rd ECML/PKDD Workshop on Advanced Analytics and Learning on Temporal Data (AALTD)

Schäfer, P., Leser, U.: Multivariate time series classification with WEASEL+MUSE. In: Proceedings of the 3rd ECML/PKDD Workshop on Advanced Analytics and Learning on Temporal Data (AALTD). pp. 1–11 (2018)

2018
[19]

Data mining and knowledge discovery31(1), 1–31 (2017)

Shokoohi-Yekta, M., Hu, B., Jin, H., Wang, J., Keogh, E.: Generalizing dtw to the multi-dimensional case requires an adaptive approach. Data mining and knowledge discovery31(1), 1–31 (2017)

2017
[20]

IEEE Sensors Journal22(1), 544–554 (2021)

Silva, L.T., Souza, V.M.A., Batista, G.E.: An open-source tool for classification models in resource-constrained hardware. IEEE Sensors Journal22(1), 544–554 (2021)

2021
[21]

Engineering Applications of Artificial Intelli- gence74, 198–211 (2018)

Souza, V.M.A.: Asphalt pavement classification using smartphone accelerometer and complexity invariant distance. Engineering Applications of Artificial Intelli- gence74, 198–211 (2018)

2018
[22]

Knowledge and Information Systems63(6), 1497–1527 (2021)

Souza, V.M.A., Parmezan, A.R., Chowdhury, F.A., Mueen, A.: Efficient unsuper- vised drift detector for fast and high-dimensional data streams. Knowledge and Information Systems63(6), 1497–1527 (2021)

2021
[23]

Knowledge-Based Systems309, 112864 (2025)

Souza, V.M.A., Veiga, P.S., Ribeiro, A.G.R.: Visemble: A fast ensemble approach for time series classification with multiple visual representations. Knowledge-Based Systems309, 112864 (2025)

2025
[24]

Frontiers in neuroscience6, 55 (2012)

Tangermann, M., Müller, K.R., Aertsen, A., Birbaumer, N., Braun, C., Brunner, C., Leeb, R., Mehring, C., Miller, K.J., Müller-Putz, G.R., et al.: Review of the bci competition iv. Frontiers in neuroscience6, 55 (2012)

2012
[25]

In: Proceedings of the AAAI conference on artificial intelligence

Zhang, X., Gao, Y., Lin, J., Lu, C.T.: Tapnet: Multivariate time series classification with attentional prototypical network. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 6845–6852 (2020)

2020