CoChat

Comparing self-supervised learning techniques for wearable human activity recognition

Open access
Ek, Sannara, Presotto, Riccardo, Civitarese, Gabriele, Portet, François, Lalanda, Philippe, Bettini, Claudio · 2025 · CCF Transactions on Pervasive Computing and Interaction

The DOI matches its Crossref record; 57 of the 57 references it cites were found in a registry or an index; 1 of the 30 sources its claims cite were checked for support.

Abstract

Human Activity Recognition (HAR) based on the sensors of mobile/wearable devices aims to detect the physical activities performed by humans in their daily lives. Although supervised learning methods are the most effective in this task, their effectiveness is constrained to using a large amount of labeled data during training. While collecting raw unlabeled data can be relatively easy, annotating data is challenging due to costs, intrusiveness, and time constraints. To address these challenges, this paper explores alternative approaches for accurate HAR using a limited amount of labeled data. In particular, we have adapted recent Self-Supervised Learning (SSL) algorithms to the HAR domain and compared their effectiveness. We investigate three state-of-the-art SSL techniques of different families: contrastive, generative, and predictive. Additionally, we evaluate the impact of the underlying neural network on the recognition rate by comparing state-of-the-art CNN and transformer architectures. Our results show that a Masked Auto Encoder approach significantly outperforms other SSL approaches, including SimCLR, commonly considered one of the best-performing SSL methods in the HAR domain. The code and the pre-trained SSL models are publicly available for further research and development.

Overview

CoChat Summary

## TL;DR

The paper compares three SSL approaches for wearable HAR—SimCLR, MAE, and data2vec—across CNN (iSPLInception) and transformer (HART) backbones using a Leave-One-Dataset-Out benchmark. MAE is the most robust method, especially in the frozen-feature-extractor setting and in low-label regimes, while SimCLR performs worst overall.

## Key findings

  • With the feature extractor frozen, MAE outperformed the other SSL methods and the supervised baseline on average: 80.90 ± 10.88 (iSPL) and 82.56 ± 11.38 (HART), versus supervised 65.74 ± 16.56 and 74.95 ± 15.92.
  • In the frozen, dataset-level results, MAE was best on every dataset reported; for example, on HHAR it reached 66.11 ± 1.05 (iSPL) and 66.64 ± 1.61 (HART), and on UCI 92.17 ± 0.83 (iSPL) and 93.40 ± 0.38 (HART).
  • In the unfrozen setting with full training data, there was no clear winner; performances were very close, with all methods reaching about the same recognition rates for a given architecture.
  • In scarce-label experiments, MAE was consistently strongest overall and achieved above 60% F1 with only 10 samples per class in the frozen setting; SimCLR performed very poorly, especially on RealWorld.

## Methods

  • The study adapts and compares three SSL families for HAR: SimCLR (contrastive), MAE (generative/masking), and data2vec (predictive), each paired with two architectures: iSPLInception and HART.
  • Evaluation uses Leave-One-Dataset-Out (LODO) across six public inertial-sensor datasets: HHAR, MobiAct, MotionSense, RealWorld, UCI, and PAMAP2; data were downsampled to 50 Hz, windowed into 128-sample segments with 50% overlap, yielding 147 subjects, 10 activities, 213,289 samples (about 151 hours), and 12 devices.
  • Main analysis: pretraining for 300 epochs with batch size 128, then fine-tuning for 100 epochs with batch size 64; performance is reported using macro F1, averaged over 5 random runs, under both frozen and unfrozen feature-extractor settings and across label-scarcity scenarios from 1 to 100 samples per class.

## Limitations

  • The paper states that the study was limited to three SSL techniques and two models, and that it was conducted only as offline experiments; behavior in real-world settings remains to be evaluated.
  • The authors note that the available HAR datasets are still much smaller and less diverse than those in vision or text, even though LODO helps mitigate limited data.
  • Evident from the text: the label-scarcity experiments are reported only on RealWorld and MotionSense for presentation purposes, even though complete results exist in the repository.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s42486-024-00182-9

Cites

57 of 57 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 Valid · 3 Not valid · 17 Not checked · 75 Own results and interpretations

50 more claims

Create a free account to see all 96 claims and to verify the 29 references still open.

See all claims — free account
  • It may state two separate claims as one. 19
  • Probably not a substantial claim. 25

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Research works printed but not linked: 8

  1. Anguita, D., Ghio, A., Oneto, L., Parra, X., Reyes-Ortiz, J.L.: A public domain dataset for human activity recognition using smartphones. In: 21st European Symposium on Artificial Neural Networks, ESANN 2013, Bruges, Belgium, April 24–26, 2013 (2013)
  2. Baevski, A., Hsu, W.-N., Xu, Q., Babu, A., Gu, J., Auli, M.: Data2vec: A general framework for self-supervised learning in speech, vision and language. In: Int. Conference on Machine Learning, pp. 1298–1312 (2022). PMLR
  3. 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 (2020). PMLR
  4. Devlin, J., Chang, M.-W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019). https://api.semanticscholar.org/CorpusID:52967399
  5. Kalantidis, Y., Sariyildiz, M.B., Pion, N., Weinzaepfel, P., Larlus, D.: Hard negative mixing for contrastive learning. Adv. Neural. Inf. Process. Syst. 33, 21798–21809 (2020)
  6. Miao, S., Chen, L., Hu, R.: Spatial-temporal masked autoencoder for multi-device wearable human activity recognition. Proc. ACM Interact. Mob. Wear. Ubiq. Technol. 7(4), 1–25 (2024)
  7. Qin, X., Chen, Y., Wang, J., Yu, C.: Cross-dataset activity recognition via adaptive spatial-temporal transfer learning. ACM IMWUT 3(4), 1–25 (2019)
  8. Zhang, X., Zhao, Z., Tsiligkaridis, T., Zitnik, M.: Self-supervised contrastive pre-training for time series via time-frequency consistency. Adv. Neural. Inf. Process. Syst. 35, 3988–4003 (2022)

Cited by

See all 15

Metadata

Authors
Ek, S., Presotto, R., Civitarese, G., Portet, F., Lalanda, P., & Bettini, C.
Affiliations
  • Centre National de la Recherche Scientifique (FR)
  • University of Milan (IT)
Corresponding author
Ek, Sannara
Venue
CCF Transactions on Pervasive Computing and Interaction
Published
2025-03-30
Volume
7
Issue
3
Pages
324-341
ISSN
  • 2524-521X (ISSN-L)
  • 2524-5228
OpenAlex
W4409016795
Cited by
15
References
57
Citations by year
2026: 6 · 2025: 8
FWCI
4.26
Type
journal article
Language
en
Open access
Licence
Locations
  • CCF Transactions on Pervasive Computing and Interaction (publishedVersion, public-domain)
  • SSRN Electronic Journal (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion, public-domain)
  • HAL (submittedVersion, cc-by)
  • Archivio Istituzionale della Ricerca (Universita Degli Studi Di Milano) (submittedVersion, other-oa)
  • arXiv (Cornell University)
Funders
  • Naval Group
  • Nextgeneration
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Health Sciences › Medicine › Public Health, Environmental and Occupational Health › Physical Activity and Health
Keywords
human activity recognition, wearable sensors, self-supervised learning, contrastive learning, generative models, prediction model, masked autoencoder, SimCLR, limited labeled data, pre-trained models

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.