Diagnosing Temporal Misalignment in Multichannel Time-Series Classification with Minimum Description Length

Sebastian Buschjäger; Michael Frichert; Daniel Kuhse; Jian-Jia Chen

cs.LGSep 13, 2026

Diagnosing Temporal Misalignment in Multichannel Time-Series Classification with Minimum Description Length

Authors: Sebastian Buschjäger, Michael Frichert, Daniel Kuhse, Jian-Jia Chen

Organizations: Lamarr Institute, TU Dortmund University, Germany · Cyber-Physical Systems, RWTH Aachen, Germany

Abstract

Multichannel time-series classification commonly assumes synchronized sensor streams, although latency, clock drift, and preprocessing can introduce relative delays during data collection or after deployment. Existing synchronization solutions are often hardware-specific and difficult to apply retrospectively. Consequently, synchronization problems may remain undetected while classification performance is suboptimal. We introduce a classifier- and label-free diagnostic based on minimum description length (MDL). Our method applies candidate temporal shifts to sensor groups and measures how efficiently one group can be encoded through a representation of the remaining channels. An increased codelength indicates that the shift destroys shared temporal structure, whereas the minimum identifies the alignment most strongly supported by the data. Unlike learned synchronization methods, the diagnostic requires neither retraining nor a trusted aligned reference and can therefore test both training and deployment data for misalignments. Experiments on two controlled synthetic tasks and nine real-world datasets show that the metric exposes alignment structure and can recover accuracy under induced deployment drift. A whole-dataset audit further identifies stable nonzero MDL optima in established benchmarks including FordChallenge, Opportunity, PAMAP2, and UCIActivity, revealing potential systematic offsets that conventional model evaluation does not expose. Our method thus provides a general-purpose tool for detecting, understanding, and correcting temporal misalignment throughout the time-series learning pipeline. Our code is available under https://github.com/sbuschjaeger/mdl-temporal-misalignment.

Figures & tables

Dataset	$n$	$T$	$C$	$\|\mathcal{G}\|$	$\|\mathcal{Y}\|$
CrowdSourced	12,289	256	14	14	2
DreamerA	170,246	256	14	14	2
DreamerV	170,246	256	14	14	2
STEW	28,512	256	14	14	2
Opportunity	17,386	100	113	8	5
PAMAP2	38,856	100	52	4	12

Table 1: Datasets used in the experiments. The table reports the number of windows

n

, window length

T

, channels

C

, channel groups

|\mathcal{G}|

, and classes

|\mathcal{Y}|

. Real-world datasets (top group) are part of MONSTER ( Dempster et al. 2025 ) .

Panel of Figure 1 — Figure 1: Results on the synthetic experiments. The left column shows results for the sensitive datasets, and the right column shows results for the robust one. The first row depicts an example of both datasets, where class labels are depicted as shaded regions. The second row shows the alignment costs according to our metric, and the last row shows the accuracy for all four classifiers with the shaded region depicting the standard deviation over 5 cross-validation folds.

Dataset	Group	Global $\tau_{G}^{\star}$	Training-fold $\tau_{G}^{\star}$	$Q_{G}(0)$ [kbit]
FordChallenge	env	6	4–8	248.5
Opportunity	r shoe	15	8–15	32.2
PAMAP2	ankle	20	8–20	807.4
UCIActivity	gyro	56	56	389.4
Skoda	s05	0	0–0	0.0
CrowdSourced	O2	0	0–0	0.0

Table 2: Sensor groups with the largest misalignment in each dataset. ∗ Reached the maximum lag tested, so a larger lag misalignment might be present in the data.

Figure 2: Accuracy of the four classifiers on 9 real-world datasets with imposed lags with and without correction. The solid lines show the average accuracy of the model across the five folds over imposed lags on the test data. The dashed line shows the accuracy after correcting the imposed lag via the alignment loss. For visual clarity, uncertainty bands are not displayed. See appendix for uncertainty bands.

Dataset	Our metric [%]	SyncNet [%]	Diff [pp]
FordChallenge	93.66	94.41	-0.75
DREAMERV	98.95	100.0	-1.05
Opportunity	94.38	100.0	-5.62
UCIActivity	94.38	100.0	-5.62
PAMAP2	89.44	100.0	-10.56
CrowdSourced	55.11	78.07	-22.96

Table 3: Baseline-relative lag correction using the label-free MDL diagnostic and learned SyncNet baseline. Columns report mean absolute residual error normalized by maximum window length without lag. Higher is better.

Setting	Value
Random seed	$42+{}$ fold index
Validation split	stratified 15% of training fold
Input normalization	per-channel training mean/std.
Training / evaluation batch	256 / 2048
Maximum epochs / early stopping	100 / patience 10
Classifier optimizer	Adam, $10^{-4}$

Table 4: Training and evaluation settings used for every dataset and fold.

Dataset	Evaluated lags $\mathcal{T}_{\Delta}$ [samples]	Selected group
FordChallenge	0, 1, 2, 3, 4, 5, 6, 8	vehicle
Opportunity	0, 1, 2, 3, 5, 8, 10, 15, 20	l shoe
PAMAP2	0, 1, 2, 3, 5, 8, 10, 15, 20	ankle
UCIActivity	0, 1, 2, 3, 5, 7, 10, 15, 20, 25, 30, 36, 42, 48, 56, 64	gyro
Skoda	0, 1, 2, 3, 5, 8, 10, 15, 20	s06
CrowdSourced	0, 2, 5, 8, 12, 18, 25, 35, 50, 65, 80	F7

Table 5: Configured non-negative lag grids and the sensor group used for correction and the SyncNet comparison. The notation

a{:}s{:}b

denotes values from

a

b

in steps of

s

Figure 3: Test accuracy under imposed sensor delays before and after label-free MDL correction. Lines show fold means and bands show one standard deviation over five folds. Solid lines are evaluated at the imposed lag. Dashed lines use the residual lag selected by the MDL alignment cost.

Figure 4: Label-free alignment cost over all evaluated real-world datasets. Lines show fold means and bands show one standard deviation. Each panel uses the classifier-selected group from Table 5 . Raw bit scales are dataset-specific.

Dataset	MDL	SyncNet	Cases
FordChallenge	2.03	1.79	33
Opportunity	4.50	0.00	22
PAMAP2	8.44	0.00	9
UCIActivity	3.60	0.00	10
Skoda	0.00	0.00	45
CrowdSourced	79.00	38.60	5

Table 6: Baseline-relative lag errors in samples (lower is better). Cases counts the jointly reachable fold–lag pairs.

Dataset	Nominal	Lagged	MDL	SyncNet
FordChallenge	86.5	86.5	86.5	86.5
Opportunity	81.6	81.1	81.2	81.6
PAMAP2	74.0	72.6	72.8	74.0
UCIActivity	92.4	89.9	89.9	92.4
Skoda	91.4	91.1	91.4	91.4
CrowdSourced	67.1	66.0	67.3	66.3

Table 7: Classifier accuracy after lag correction [%]. Values average CNN, ResNet, late fusion, and Magnitude CNN over five folds and the jointly reachable imposed lags.

Dataset	Group	U lag	C lag
FordChallenge	vehicle	0 [0–8]	6 [0–8]
Opportunity	l shoe	5 [5–15]	15 [0–15]
PAMAP2	ankle	20 [8–20]	3 [0–5]
UCIActivity	gyro	56 [56–56]	64 [64–64]
Skoda	s06	0 [0–0]	0 [0–0]
CrowdSourced	F7	80 [80–80]	65 [65–80]

Table 8: Preferred lags for the label-free (U) and class-conditional (C) metrics, reported as median [minimum–maximum] over five folds.

Dataset	U $A_{G}$	C $A_{G}$
FordChallenge	15.9	17.2
Opportunity	94.1	122.2
PAMAP2	278.1	51.7
UCIActivity	167.5	166.1
Skoda	28.0	82.8
CrowdSourced	1.7	25.2

Table 9: Mean raw alignment cost

A_{G}

[kbit] for the label-free (U) and class-conditional (C) metrics.

Explore similar work

The Effect of Missingness-Pattern Mismatch on Method Selection for Time-Series Classification: A Controlled Empirical Study
Oct 4, 2026Ruiqi Zhao, Zishun Yuan, Zhentao Wang +3Time-Series Classification Missing Data
The University of Tokyo, Tokyo, Japan · Stony Brook University, SUNY Korea, Incheon, South Korea · Fuyao University of Science and Technology, Fujian, China +3
LEARN-TS: LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Multivariate Time-Series Anomaly Detection
Sep 30, 2026Jahyeob Koo, Kio Yun, Byoungmo Koo +1Time-Series Anomaly Detection Time Series
Department of Industrial and Management Engineering, Korea University Seoul, Republic of Korea
Inference-Time Decision Calibration for Temporal Classification
Jun 14, 2026Arthur Chagas, Arthur Buzelin, Yan Aquino +4Classifier Inference-Time
Department of Computer Science (DCC), Universidade Federal de Minas Gerais (UFMG)

Oct 4, 2026, cs.LGJ/K move · Enter open · S save

The Effect of Missingness-Pattern Mismatch on Method Selection for Time-Series Classification: A Controlled Empirical Study

Ruiqi Zhao, Zishun Yuan, Zhentao Wang, Jiahao Quan+2

The University of Tokyo, Tokyo, Japan · Stony Brook University, SUNY Korea, Incheon, South Korea · Fuyao University of Science and Technology, Fujian, China+3

Time-Series Classification Missing Data

Open paper AskPDF arXiv

Classifiers for time-series classification are commonly selected on validation data, but the temporal pattern of missing observations at deployment may differ from the pattern seen during validation. We examine whether such a mismatch affects validation-based classifier selection. In a controlled

2 \times 2

design, validation and test sets of 64 univariate UCR datasets were masked with either random point missingness or circular block missingness at six rates from 5% to 30%, imputed by linear interpolation, and used to select among three prespecified candidates: 1NN-DTW, MiniRocket with a Ridge classifier, and a statistical-feature Random Forest. Training data remained complete, and selections made under matched and mismatched validation patterns were compared on the same masked test sets. Mismatched validation reduced the test balanced accuracy of the selected classifier by 1.14 percentage points on average (95% CI 0.79 to 1.51), with losses on 49 of the 64 datasets. The loss was negligible at 5% missingness and increased to 2.46 percentage points at 30%. It was concentrated in point-masked deployment (1.84 percentage points), where block-masked validation shifted selection away from the usually best candidate, while the effect for block-masked deployment was small and not significant. Mismatch changed the selected classifier in 35.5% of paired comparisons, but a changed selection did not always reduce performance. A supplementary analysis with non-wrapping linear blocks reproduced these findings with a larger effect (1.67 percentage points). Matching the missingness pattern of validation data to the expected deployment pattern is therefore a simple safeguard for method selection, particularly at higher missingness rates.

Abstract

Explore similar work

The Effect of Missingness-Pattern Mismatch on Method Selection for Time-Series Classification: A Controlled Empirical Study

LEARN-TS: LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Multivariate Time-Series Anomaly Detection

Inference-Time Decision Calibration for Temporal Classification