Learning by Shifting: Temporal View Construction for Time Series Contrastive Learning

Abdul-Kazeem Shamba; Gavin Taylor; Kerstin Bach

arxiv: 2606.21957 · v1 · pith:LO6XJ6WJnew · submitted 2026-06-20 · 💻 cs.LG

Learning by Shifting: Temporal View Construction for Time Series Contrastive Learning

Abdul-Kazeem Shamba , Kerstin Bach , Gavin Taylor This is my paper

Pith reviewed 2026-06-26 12:07 UTC · model grok-4.3

classification 💻 cs.LG

keywords time seriescontrastive learningself-supervised learningtemporal shiftrepresentation learningShiFTview construction

0 comments

The pith

A simple deterministic temporal shift suffices to learn strong time series representations in contrastive learning.

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

The paper sets out to establish that contrastive learning for time series does not require elaborate hand-crafted augmentations or masking schemes. Instead, constructing positive pairs through a fixed, deterministic temporal shift directly encodes the desired shift invariance and yields effective representations. A sympathetic reader would care because this removes the need to inject domain-specific assumptions that can limit generalization or create spurious correlations. The resulting method, called ShiFT, reaches state-of-the-art accuracy on six real-world datasets plus the UCR and UEA archives while shortening training time. The work also supplies an empirical study of how batch size and the number of negatives affect downstream performance.

Core claim

By replacing complex view-construction heuristics with a deterministic temporal shift, the ShiFT procedure produces positive pairs whose only guaranteed relation is a known time offset. Training a contrastive objective on these pairs learns representations that are invariant to such shifts and that transfer to classification tasks across diverse temporal datasets, outperforming prior self-supervised baselines while using less compute.

What carries the argument

The deterministic temporal view construction inside Shift Invariant Feature Training (ShiFT), which generates each positive pair by applying a fixed time offset to the original series.

If this is right

Representations learned this way generalize across varied temporal patterns without dataset-specific tuning.
Training time decreases because view generation is a cheap deterministic operation rather than a stochastic augmentation pipeline.
Downstream classification performance remains stable across a range of batch sizes and negative counts, offering practical guidance for implementation.
The same shift-based pairing can be applied to the full UCR and UEA archives and still yields state-of-the-art results.

Where Pith is reading between the lines

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

Minimal, structure-preserving views may be preferable to rich but biased augmentations whenever the data possess an obvious invariance such as time translation.
The approach could be tested on irregularly sampled or multivariate series to check whether the same deterministic shift remains sufficient.
If the shift distance is treated as a hyper-parameter, one could measure how downstream accuracy varies with that distance to identify an optimal scale of invariance.

Load-bearing premise

Hand-crafted augmentations necessarily embed strong domain assumptions that hurt generalization, whereas a plain temporal shift introduces no new spurious correlations.

What would settle it

A controlled experiment in which ShiFT is trained and evaluated on the same six benchmarks but achieves materially lower accuracy than at least one established augmentation-based contrastive baseline.

Figures

Figures reproduced from arXiv: 2606.21957 by Abdul-Kazeem Shamba, Gavin Taylor, Kerstin Bach.

**Figure 2.** Figure 2: Comparison of view construction strategies for contrastive learning in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of ShiFT. A time series instance X ∈ R T ×C is partitioned into two overlapping shifted views Vi , Vj ∈ R L×C via strided subwindowing. Each view is independently encoded by a shared encoder fθ into a representation h ∈ R 1×F , then projected by an MLP head into z ∈ R 1×D. The NT-Xent loss LNT-Xent is computed over the projected representations. The MLP head is discarded at evaluation time. [8], E… view at source ↗

**Figure 4.** Figure 4: Effect of projection head type and output dimension on linear probing accuracy (%), averaged over the six large-scale datasets. accuracy to 62.62%, the lowest among all configurations. Beyond the inconsistent semantic sharing between views, circular rolling introduces artificial discontinuities at the wraparound boundary and fails the non-triviality criterion: at small shift magnitudes the two views are n… view at source ↗

**Figure 5.** Figure 5: Effect of batch size on linear probing accuracy (%) across four representative largescale datasets [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: t-SNE projections of learned representations on the [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

read the original abstract

Supervised learning demands large quantities of labeled data, a bottleneck that is expensive and reliant on domain-specific expertise. Self-supervised learning, particularly contrastive learning, has emerged as a compelling alternative, enabling rich representation learning directly from unlabeled data. Yet its success hinges critically on the design of positive and negative sample pairs. Existing approaches for time series rely on hand-crafted augmentations and masking heuristics that embed strong domain assumptions, often limiting generalization across diverse temporal patterns and potentially introducing spurious correlations. In this work, we challenge this paradigm by demonstrating that explicitly encoding temporal shift invariance through a simple, deterministic view construction is sufficient to learn strong representations for time series classification. By exploiting temporal structure, our method, Shift Invariant Feature Training (ShiFT), achieves state-of-the-art performance on six diverse real-world time series benchmark datasets, as well as the UCR and UEA archives, while reducing training time. Beyond empirical performance, we present a systematic analysis of contrastive learning dynamics in time series settings, examining the effects of batch size and the number of negatives on downstream performance. Our findings provide practical insights for designing efficient contrastive learning frameworks for time series representation learning. The source code is publicly available at https://github.com/sfi-norwai/ShiFT.

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

ShiFT uses a fixed temporal shift for contrastive pairs in time series and claims SOTA with less training time, but the abstract gives no numbers or ablations to check it.

read the letter

The main takeaway is a straightforward alternative to hand-crafted augmentations in time series contrastive learning: just apply a deterministic temporal shift to create positive views. The paper argues this encodes shift invariance without the domain assumptions that come with masking or noise, and it reports better results on UCR/UEA plus six other datasets along with shorter training.

What the work does is keep the view construction simple and tied to the data's temporal order. It also includes a section looking at batch size and number of negatives, which gives some practical guidance on training dynamics. Public code is a plus for anyone wanting to test the method.

The weak part is the lack of visible evidence for the performance claims. The abstract states SOTA and faster training but shows none of the tables, baselines, statistical tests, or ablations that would let a reader judge whether the shift actually holds up or only works on stationary series. The concern about non-stationary cases or timing-dependent labels is reasonable and not addressed in the provided text.

This is aimed at people building self-supervised models for sensor or temporal data where labels are scarce. Readers who want a minimal baseline or are already comparing augmentation strategies could find the shift idea and the batch/negative analysis worth checking.

Send it to review. The idea is clean enough and the code is out, so referees can evaluate the experiments directly.

Referee Report

2 major / 1 minor

Summary. The paper proposes ShiFT (Shift Invariant Feature Training), a contrastive learning framework for time series that constructs positive pairs via a simple deterministic temporal shift to explicitly encode shift invariance. It claims this suffices to learn strong representations, achieving SOTA on six real-world benchmarks plus UCR/UEA archives, reducing training time relative to augmentation-heavy baselines, and includes analysis of batch size and negative sample count effects on downstream classification performance. Code is released publicly.

Significance. If the empirical results hold under the stated construction, the work would show that complex hand-crafted augmentations are unnecessary for effective time-series contrastive learning and that a minimal deterministic shift can avoid the domain assumptions they introduce. Public code supports reproducibility; the dynamics analysis on batch size and negatives could offer practical guidance for the field.

major comments (2)

[Abstract] Abstract: the SOTA claim and reduced training time assertion are presented without reference to any baselines, statistical tests, ablation results, or dataset names, rendering the central empirical contribution unverifiable from the provided text and undermining assessment of the sufficiency claim.
[Abstract] Method/Experiments (implied by abstract's contrast with existing approaches): the assertion that the deterministic shift 'avoids spurious correlations' is load-bearing for the novelty but relies on the untested premise that temporal shifts preserve downstream labels across all evaluated benchmarks; no analysis or stratification addresses non-stationary series or cases where class depends on event timing (e.g., early vs. late anomalies), as required by the central claim.

minor comments (1)

[Abstract] Abstract: the phrase 'six diverse real-world time series benchmark datasets' is stated without enumeration, reducing clarity on the scope of the claimed generalization.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below, indicating planned revisions to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the SOTA claim and reduced training time assertion are presented without reference to any baselines, statistical tests, ablation results, or dataset names, rendering the central empirical contribution unverifiable from the provided text and undermining assessment of the sufficiency claim.

Authors: We agree that the abstract's brevity limits immediate verifiability of the empirical claims. In the revised version we will expand the abstract to name the six real-world benchmarks, reference the UCR/UEA archives, and note that detailed baseline comparisons (including training-time measurements) appear in Section 4, while preserving conciseness. revision: yes
Referee: [Abstract] Method/Experiments (implied by abstract's contrast with existing approaches): the assertion that the deterministic shift 'avoids spurious correlations' is load-bearing for the novelty but relies on the untested premise that temporal shifts preserve downstream labels across all evaluated benchmarks; no analysis or stratification addresses non-stationary series or cases where class depends on event timing (e.g., early vs. late anomalies), as required by the central claim.

Authors: The empirical results across the chosen benchmarks provide supporting evidence that the deterministic shift yields strong representations without the spurious correlations introduced by more complex augmentations. Nevertheless, we acknowledge the absence of explicit stratification by non-stationarity or timing-dependent label structure. We will add a dedicated limitations paragraph in the revised manuscript that discusses this modeling assumption and identifies datasets where the approach may be less suitable. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method with independent benchmark validation

full rationale

The paper describes an empirical contrastive learning approach (ShiFT) that constructs positive pairs via deterministic temporal shifts and evaluates it on UCR/UEA and other benchmarks. No equations, derivations, or fitted parameters are presented that reduce to the inputs by construction. Central performance claims rest on external dataset comparisons rather than self-referential fitting or self-citation chains. The method is self-contained against public benchmarks and released code, satisfying the criteria for a non-circular empirical result.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The method rests on standard contrastive learning assumptions plus the novel claim that temporal shift suffices; no new entities postulated and few explicit free parameters beyond typical hyperparameters like batch size.

free parameters (2)

batch size
Examined as a variable affecting downstream performance; chosen during experiments.
number of negatives
Systematically analyzed for impact on performance; a design choice in the contrastive setup.

axioms (1)

domain assumption Positive pairs created by temporal shifts are semantically similar enough for effective contrastive learning
Invoked when stating that shift invariance is sufficient without additional augmentations.

pith-pipeline@v0.9.1-grok · 5757 in / 1169 out tokens · 24492 ms · 2026-06-26T12:07:27.579853+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 3 canonical work pages

[1]

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 (2018), https://arxiv.org/abs/1811.00075

Pith/arXiv arXiv 2018
[2]

In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 16 A.K

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 16 A.K. Shamba et al

2021
[3]

In: International conference on machine learning

Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597–1607. PMLR (2020)

2020
[4]

IEEE/CAA Journal of Automatica Sinica6(6), 1293–1305 (2019)

Dau, H.A., Bagnall, A., Kamgar, K., Yeh, C.C.M., Zhu, Y., Gharghabi, S., Ratanamahatana, C.A., Keogh, E.: The ucr time series archive. IEEE/CAA Journal of Automatica Sinica6(6), 1293–1305 (2019)

2019
[5]

Dong, J., Wu, H., Zhang, H., Zhang, L., Wang, J., Long, M.: Simmtm: A simple pre- trainingframeworkformaskedtime-seriesmodeling.AdvancesinNeuralInformation Processing Systems36, 29996–30025 (2023)

2023
[6]

In: The Fourteenth International Conference on Learning Representations (2026),https://openreview.net/forum?id=vTLmHAkoIW

Dou, Z., Yao, Z., Xie, Z., Wen, X., Xiao, T., Pei, D.: AutoDA-timeseries: Automated data augmentation for time series. In: The Fourteenth International Conference on Learning Representations (2026),https://openreview.net/forum?id=vTLmHAkoIW

2026
[7]

In: International Conference on Learning Representations (2018),https://openreview.net/forum?id=S1v4N2l0-

Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by pre- dicting image rotations. In: International Conference on Learning Representations (2018),https://openreview.net/forum?id=S1v4N2l0-

2018
[8]

Christoph Hofer, Florian Graf, Bastian Rieck, Marc Niethammer, and Roland Kwitt

Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P., Mark, R., Mietus, J., Moody, G., Peng, C.K., Stanley, H.: Physiobank, physiotoolkit, and physionet : Components of a new research resource for complex physiologic signals. Circulation 101, E215–20 (07 2000).https://doi.org/10.1161/01.CIR.101.23.e215

work page doi:10.1161/01.cir.101.23.e215 2000
[9]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729–9738 (2020)

2020
[10]

Data Mining and Knowledge Discovery34(6), 1936–1962 (2020)

Ismail Fawaz, H., Lucas, B., Forestier, G., Pelletier, C., Schmidt, D.F., Weber, J., Webb, G.I., Idoumghar, L., Muller, P.A., Petitjean, F.: Inceptiontime: Finding alexnet for time series classification. Data Mining and Knowledge Discovery34(6), 1936–1962 (2020)

1936
[11]

Sensors (Basel, Switzerland) 21(23) (2021).https://doi.org/10.3390/s21237853

Logacjov, A., Bach, K., Kongsvold, A., Bårdstu, H.B., Mork, P.J.: Harth: A human activity recognition dataset for machine learning. Sensors (Basel, Switzerland) 21(23) (2021).https://doi.org/10.3390/s21237853

work page doi:10.3390/s21237853 2021
[12]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Luo, D., Cheng, W., Wang, Y., Xu, D., Ni, J., Yu, W., Zhang, X., Liu, Y., Chen, Y., Chen, H., et al.: Time series contrastive learning with information-aware augmenta- tions. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 4534–4542 (2023)

2023
[13]

In: Computers in Cardiology

Moody, G.B., Mark, R.G.: A new method for detecting atrial fibrillation using r-r intervals. In: Computers in Cardiology. pp. 227–230 (1983)

1983
[14]

In: European conference on computer vision

Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: European conference on computer vision. pp. 69–84. Springer (2016)

2016
[15]

arXiv preprint arXiv:1807.03748 (2018)

Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

Pith/arXiv arXiv 2018
[16]

In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Pathak, D., Krähenbühl, P., Donahue, J., Darrell, T., Efros, A.A.: Context encoders: Feature learning by inpainting. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2536–2544 (2016).https://doi.org/10.1109/ CVPR.2016.278

2016
[17]

In: 2012 16th international symposium on wearable computers

Reiss, A., Stricker, D.: Introducing a new benchmarked dataset for activity moni- toring. In: 2012 16th international symposium on wearable computers. pp. 108–109. IEEE (2012)

2012
[18]

In: Proceedings of the European Conference on Artificial Intelligence (ECAI)

Shamba, A.K., Bach, K., Taylor, G.: Contrast all the time: Learning time series rep- resentation from temporal consistency. In: Proceedings of the European Conference on Artificial Intelligence (ECAI). Frontiers in Artificial Intelligence and Applications, vol. 413, pp. 2810–2817 (2025),https://doi.org/10.3233/FAIA251137 Learning by Shifting 17

work page doi:10.3233/faia251137 2025
[19]

In: AAAI workshop on activity context representation: tech- niques and languages

Weiss, G.M., Lockhart, J.W.: The impact of personalization on smartphone-based activity recognition. In: AAAI workshop on activity context representation: tech- niques and languages. pp. 98–104. Toronto. (2012)

2012
[20]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Yue, Z., Wang, Y., Duan, J., Yang, T., Huang, C., Tong, Y., Xu, B.: Ts2vec: Towards universal representation of time series. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8980–8987 (2022)

2022
[21]

ACM Transactions on Embedded Computing Systems (TECS)11(3), 1–30 (2012)

Zappi, P., Roggen, D., Farella, E., Tröster, G., Benini, L.: Network-level power- performance trade-off in wearable activity recognition: A dynamic sensor selection approach. ACM Transactions on Embedded Computing Systems (TECS)11(3), 1–30 (2012)

2012

[1] [1]

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 (2018), https://arxiv.org/abs/1811.00075

Pith/arXiv arXiv 2018

[2] [2]

In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 16 A.K

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) 16 A.K. Shamba et al

2021

[3] [3]

In: International conference on machine learning

Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597–1607. PMLR (2020)

2020

[4] [4]

IEEE/CAA Journal of Automatica Sinica6(6), 1293–1305 (2019)

Dau, H.A., Bagnall, A., Kamgar, K., Yeh, C.C.M., Zhu, Y., Gharghabi, S., Ratanamahatana, C.A., Keogh, E.: The ucr time series archive. IEEE/CAA Journal of Automatica Sinica6(6), 1293–1305 (2019)

2019

[5] [5]

Dong, J., Wu, H., Zhang, H., Zhang, L., Wang, J., Long, M.: Simmtm: A simple pre- trainingframeworkformaskedtime-seriesmodeling.AdvancesinNeuralInformation Processing Systems36, 29996–30025 (2023)

2023

[6] [6]

In: The Fourteenth International Conference on Learning Representations (2026),https://openreview.net/forum?id=vTLmHAkoIW

Dou, Z., Yao, Z., Xie, Z., Wen, X., Xiao, T., Pei, D.: AutoDA-timeseries: Automated data augmentation for time series. In: The Fourteenth International Conference on Learning Representations (2026),https://openreview.net/forum?id=vTLmHAkoIW

2026

[7] [7]

In: International Conference on Learning Representations (2018),https://openreview.net/forum?id=S1v4N2l0-

Gidaris, S., Singh, P., Komodakis, N.: Unsupervised representation learning by pre- dicting image rotations. In: International Conference on Learning Representations (2018),https://openreview.net/forum?id=S1v4N2l0-

2018

[8] [8]

Christoph Hofer, Florian Graf, Bastian Rieck, Marc Niethammer, and Roland Kwitt

Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P., Mark, R., Mietus, J., Moody, G., Peng, C.K., Stanley, H.: Physiobank, physiotoolkit, and physionet : Components of a new research resource for complex physiologic signals. Circulation 101, E215–20 (07 2000).https://doi.org/10.1161/01.CIR.101.23.e215

work page doi:10.1161/01.cir.101.23.e215 2000

[9] [9]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729–9738 (2020)

2020

[10] [10]

Data Mining and Knowledge Discovery34(6), 1936–1962 (2020)

Ismail Fawaz, H., Lucas, B., Forestier, G., Pelletier, C., Schmidt, D.F., Weber, J., Webb, G.I., Idoumghar, L., Muller, P.A., Petitjean, F.: Inceptiontime: Finding alexnet for time series classification. Data Mining and Knowledge Discovery34(6), 1936–1962 (2020)

1936

[11] [11]

Sensors (Basel, Switzerland) 21(23) (2021).https://doi.org/10.3390/s21237853

Logacjov, A., Bach, K., Kongsvold, A., Bårdstu, H.B., Mork, P.J.: Harth: A human activity recognition dataset for machine learning. Sensors (Basel, Switzerland) 21(23) (2021).https://doi.org/10.3390/s21237853

work page doi:10.3390/s21237853 2021

[12] [12]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Luo, D., Cheng, W., Wang, Y., Xu, D., Ni, J., Yu, W., Zhang, X., Liu, Y., Chen, Y., Chen, H., et al.: Time series contrastive learning with information-aware augmenta- tions. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 4534–4542 (2023)

2023

[13] [13]

In: Computers in Cardiology

Moody, G.B., Mark, R.G.: A new method for detecting atrial fibrillation using r-r intervals. In: Computers in Cardiology. pp. 227–230 (1983)

1983

[14] [14]

In: European conference on computer vision

Noroozi, M., Favaro, P.: Unsupervised learning of visual representations by solving jigsaw puzzles. In: European conference on computer vision. pp. 69–84. Springer (2016)

2016

[15] [15]

arXiv preprint arXiv:1807.03748 (2018)

Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

Pith/arXiv arXiv 2018

[16] [16]

In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Pathak, D., Krähenbühl, P., Donahue, J., Darrell, T., Efros, A.A.: Context encoders: Feature learning by inpainting. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2536–2544 (2016).https://doi.org/10.1109/ CVPR.2016.278

2016

[17] [17]

In: 2012 16th international symposium on wearable computers

Reiss, A., Stricker, D.: Introducing a new benchmarked dataset for activity moni- toring. In: 2012 16th international symposium on wearable computers. pp. 108–109. IEEE (2012)

2012

[18] [18]

In: Proceedings of the European Conference on Artificial Intelligence (ECAI)

Shamba, A.K., Bach, K., Taylor, G.: Contrast all the time: Learning time series rep- resentation from temporal consistency. In: Proceedings of the European Conference on Artificial Intelligence (ECAI). Frontiers in Artificial Intelligence and Applications, vol. 413, pp. 2810–2817 (2025),https://doi.org/10.3233/FAIA251137 Learning by Shifting 17

work page doi:10.3233/faia251137 2025

[19] [19]

In: AAAI workshop on activity context representation: tech- niques and languages

Weiss, G.M., Lockhart, J.W.: The impact of personalization on smartphone-based activity recognition. In: AAAI workshop on activity context representation: tech- niques and languages. pp. 98–104. Toronto. (2012)

2012

[20] [20]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Yue, Z., Wang, Y., Duan, J., Yang, T., Huang, C., Tong, Y., Xu, B.: Ts2vec: Towards universal representation of time series. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8980–8987 (2022)

2022

[21] [21]

ACM Transactions on Embedded Computing Systems (TECS)11(3), 1–30 (2012)

Zappi, P., Roggen, D., Farella, E., Tröster, G., Benini, L.: Network-level power- performance trade-off in wearable activity recognition: A dynamic sensor selection approach. ACM Transactions on Embedded Computing Systems (TECS)11(3), 1–30 (2012)

2012