Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 67 of the 67 references it cites were found in a registry or an index; 8 of the 44 sources its claims cite were checked for support.

Reducing Label Dependency in Human Activity Recognition with Wearables: From Supervised Learning to Novel Weakly Self-Supervised Approaches

Open access
Sheng, Taoran, Huber, Manfred · 2025 · Sensors

Abstract

Human activity recognition (HAR) using wearable sensors has advanced through various machine learning paradigms, each with inherent trade-offs between performance and labeling requirements. While fully supervised techniques achieve high accuracy, they demand extensive labeled datasets that are costly to obtain. Conversely, unsupervised methods eliminate labeling needs but often deliver suboptimal performance. This paper presents a comprehensive investigation across the supervision spectrum for wearable-based HAR, with particular focus on novel approaches that minimize labeling requirements while maintaining competitive accuracy. We develop and empirically compare: (1) traditional fully supervised learning, (2) basic unsupervised learning, (3) a weakly supervised learning approach with constraints, (4) a multi-task learning approach with knowledge sharing, (5) a self-supervised approach based on domain expertise, and (6) a novel weakly self-supervised learning framework that leverages domain knowledge and minimal labeled data. Experiments across benchmark datasets demonstrate that: (i) our weakly supervised methods achieve performance comparable to fully supervised approaches while significantly reducing supervision requirements; (ii) the proposed multi-task framework enhances performance through knowledge sharing between related tasks; (iii) our weakly self-supervised approach demonstrates remarkable efficiency with just 10% of labeled data. These results not only highlight the complementary strengths of different learning paradigms, offering insights into tailoring HAR solutions based on the availability of labeled data, but also establish that our novel weakly self-supervised framework offers a promising solution for practical HAR applications where labeled data are limited.

Overview

CoChat Summary

TL;DR

The paper compares supervised, unsupervised, weakly supervised, self-supervised, and hybrid weakly self-supervised methods for wearable-sensor human activity recognition (HAR) to reduce label dependence. Across three benchmark datasets, the weakly supervised and hybrid methods are reported to approach or exceed fully supervised performance while using far fewer labels.

Key findings

  • The weakly supervised single-task method achieved 94.36% accuracy on UCI-Smartphone, 98.56% on PAMAP2, and 97.96% on REALDISP, substantially outperforming the unsupervised baseline (63.69%, 77.06%, 64.01%).
  • The weakly supervised multi-task method achieved 98.85%, 98.93%, and 98.22% accuracy on the three datasets, and on UCI-Smartphone it exceeded the supervised TCN baseline (98.85% vs. 98.66%).
  • The self-supervised method outperformed the standard autoencoder on all datasets: 74.01% vs. 63.69% on UCI-Smartphone, 85.43% vs. 77.06% on PAMAP2, and 68.12% vs. 64.01% on REALDISP.
  • The weakly self-supervised method improved further with more labels: with 1% labeled data it reached 80.15%, 91.28%, and 75.15%, and with 10% labeled data it reached 93.52%, 99.04%, and 92.23%; on PAMAP2, 99.04% exceeded the fully supervised TCN baseline of 95.27%.

Methods

  • The study systematically compares six paradigms: fully supervised learning, unsupervised autoencoders on handcrafted statistical features, weakly supervised Siamese/contrastive learning with pairwise similarity constraints, weakly supervised multi-task learning with an added person-identification task, self-supervised learning with temporal and feature consistency losses, and a two-stage weakly self-supervised framework combining self-supervised pretraining with weak supervision.
  • Evaluation was conducted on three benchmark HAR datasets: UCI-Smartphone (30 participants, 12 activities), PAMAP2 (9 participants, 12 activities), and REALDISP (17 participants, 33 activities).
  • The paper reports using accuracy and F1-score as metrics, with stratified splits and k-means clustering on learned latent representations for the non-classification approaches.

Limitations

  • The paper notes that fully supervised methods require extensive labeled data, while unsupervised methods often perform worse due to lack of guidance; this is stated as a general trade-off in the paper.
  • The authors state that temporal-consistency self-supervision can struggle on activity transitions such as sit-to-stand or stand-to-walk, where signal patterns are ambiguous; this may limit effectiveness.
  • Evident from the text: the evaluation is restricted to three benchmark datasets and to ideal-placement settings for REALDISP, so the reported results may not fully cover broader real-world deployment variability.

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

Integrity signals

Registry

Crossref · 10.3390/s25134032

Cites

67 of 67 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 (a) ResNet, (b) Siamese architecture, and (c) TCN block.
Figure 1 (a) ResNet, (b) Siamese architecture, and (c) TCN block.CC BY · PubMed Central
Figure 2 The TCN architecture used in the supervised approach.
Figure 2 The TCN architecture used in the supervised approach.CC BY · PubMed Central
Figure 3 The residual autoencoder architecture used in the unsupervised approach.
Figure 3 The residual autoencoder architecture used in the unsupervised approach.CC BY · PubMed Central
Figure 4 The Siamese architecture used in the weakly supervised single-task approach.
Figure 4 The Siamese architecture used in the weakly supervised single-task approach.CC BY · PubMed Central
Figure 5 The extended Siamese architecture used in the weakly supervised multi-task approach.
Figure 5 The extended Siamese architecture used in the weakly supervised multi-task approach.CC BY · PubMed Central
Figure 6 The residual autoencoder architecture used in the self-supervised model.
Figure 6 The residual autoencoder architecture used in the self-supervised model.CC BY · PubMed Central
Figure 7 The Siamese residual autoencoder architecture used in the weakly self-supervised model.
Figure 7 The Siamese residual autoencoder architecture used in the weakly self-supervised model.CC BY · PubMed Central
Figure 8 Visualization of representation vectors learned by weakly supervised single-task, weakly supervised multi-task, and self-supervised approaches on the UCI-Smartphone dataset. Different colors indicate different activity types in the dataset. (a) Weakly supervised single-task approach. (b) Weakly supervised multi-task approach. (c) Self-supervised approach.
Figure 8 Visualization of representation vectors learned by weakly supervised single-task, weakly supervised multi-task, and self-supervised approaches on the UCI-Smartphone dataset. Different colors indicate different activity types in the dataset. (a) Weakly supervised single-task approach. (b) Weakly supervised multi-task approach. (c) Self-supervised approach.CC BY · PubMed Central
Figure 9 Visualization of representation vectors learned by the weakly self-supervised approach on the UCI-Smartphone dataset, using 1%, 5%, and 10% of labeled data during training. Different colors indicate different activity types in the dataset. (a) Weakly self-supervised approach with 1% labeled data. (b) Weakly self-supervised approach with 5% labeled data. (c) Weakly self-supervised approach with 10% labeled data.
Figure 9 Visualization of representation vectors learned by the weakly self-supervised approach on the UCI-Smartphone dataset, using 1%, 5%, and 10% of labeled data during training. Different colors indicate different activity types in the dataset. (a) Weakly self-supervised approach with 1% labeled data. (b) Weakly self-supervised approach with 5% labeled data. (c) Weakly self-supervised approach with 10% labeled data.CC BY · PubMed Central

Verification

Claims the paper makes

5 Valid · 0 Not valid · 18 Not checked · 94 Own results and interpretations

46 more claims

Create a free account to see all 117 claims and to verify the 36 references still open.

See all claims — free account
  • Probably not a substantial claim. 34
  • It may state two separate claims as one. 41

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.

1 of 1

Citations

Cites

Show all 63 references

Research works printed but not linked: 8

  1. Ruder, S. (2017). An Overview of Multi-Task Learning in Deep Neural Networks. arXiv.
  2. Chen, R., Luo, H., Zhao, F., Meng, X., Xie, Z., and Zhu, Y. (2021). Modeling Accurate Human Activity Recognition for Embedded Devices Using Multi-level Distillation. arXiv.
  3. Simonyan, K., and Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv.
  4. McInnes, L., Healy, J., and Melville, J. (2020). UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv.
  5. Mikolov, T., Grave, E., Bojanowski, P., Puhrsch, C., and Joulin, A. (2017). Advances in pre-training distributed word representations. arXiv.
  6. Kingma, D.P., and Welling, M. (2022). Auto-Encoding Variational Bayes. arXiv.
  7. Chen, T., Kornblith, S., Norouzi, M., and Hinton, G.E. (2020). A Simple Framework for Contrastive Learning of Visual Representations. arXiv.
  8. Aljalbout, E., Golkov, V., Siddiqui, Y., and Cremers, D. (2018). Clustering with Deep Learning: Taxonomy and New Methods. arXiv.

Cited by

See all 6

Metadata

Authors
Sheng, T., & Huber, M.
Affiliations
The University of Texas at Arlington (US)
Corresponding author
Sheng, Taoran
Venue
Sensors
Publisher
MDPI AG
Published
2025-06-28
Volume
25
Issue
13
Pages
4032
ISSN
1424-8220 (ISSN-L)
OpenAlex
W4411802278
Cited by
5
References
63
Citations by year
2026: 4 · 2025: 1
FWCI
1.76
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
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
  • Physical Sciences › Computer Science › Signal Processing › Time Series Analysis and Forecasting
Keywords
human activity recognition, wearable sensors, weakly supervised learning, multi-task learning, knowledge sharing, domain knowledge, limited labeled data, unsupervised learning, supervised learning
MeSH terms
Machine Learning, Supervised Machine Learning, Unsupervised Machine Learning, Wearable Electronic Devices, Algorithms, Humans, Human Activities

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 (gpt-6-sol).

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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Reducing Label Dependency in Human Activity Recognition with Wearables: From Supervised Learning to Novel Weakly Self-Supervised Approaches. https://app.cochat.ai/papers/10.3390/s25134032
The paper
Sheng, T., & Huber, M. (2025). Reducing Label Dependency in Human Activity Recognition with Wearables: From Supervised Learning to Novel Weakly Self-Supervised Approaches. Sensors. https://doi.org/10.3390/s25134032

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.