Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition
Open accessThe DOI matches its Crossref record; 54 of the 54 references it cites were found in a registry or an index; 2 of the 12 sources its claims cite were checked for support.
Abstract
Contrastive learning is an emerging and important self-supervised learning paradigm that has been successfully applied to sensor-based human activity recognition (HAR) because it can achieve competitive performance relative to supervised learning. Contrastive learning methods generally involve instance discrimination, which means that the instances are regarded as negatives of each other, and thus their representations are pulled away from each other during the training process. However, instance discrimination could cause overclustering, meaning that the representations of instances from the same class could be overly separated. To alleviate this overclustering phenomenon, we propose a new contrastive learning framework to select negatives by clustering in HAR, which is named clustering for contrastive learning in human activity recognition (ClusterCLHAR). First, ClusterCLHAR clusters the instance representations, and for each instance, only those from different clusters are regarded as negatives. Second, a new contrastive loss function is proposed to mask the same-cluster instances from the negative pairs. We evaluate ClusterCLHAR on three popular benchmark data sets: 1) USC-HAD; 2) MotionSense; and 3) UCI-HAR, using the mean F1-score as an evaluation metric for downstream tasks. The experimental results show that ClusterCLHAR outperforms all the state-of-the-art methods applied to HAR in self-supervised learning and semi-supervised learning.
Overview
CoChat Summary
TL;DR
ClusterCLHAR is a contrastive learning framework for sensor-based human activity recognition that clusters instance representations and prevents same-cluster samples from being treated as negatives. Across three benchmark datasets, it reports state-of-the-art results in both self-supervised and semi-supervised settings.
Key findings
- In self-supervised evaluation, ClusterCLHAR achieved mean F1-scores of 58.85 on USC-HAD, 90.95 on MotionSense, and 92.63 on UCI-HAR, exceeding the compared self-supervised methods.
- With only 1% labeled data and linear evaluation, it reached 77.18 on USC-HAD, 86.69 on MotionSense, and 88.78 on UCI-HAR, improving over the previous best contrastive method by 4.82, 3.3, and 4.89 points, respectively.
- With 1% labeled data and fine-tuning, it reached 78.09 on USC-HAD, 87.35 on MotionSense, and 90.53 on UCI-HAR, improving over the previous best by 3.39, 1.16, and 3.0 points.
- Ablation showed clustering choice mattered: best-to-worst gaps were 9.68%, 2.02%, and 4.5% across the three datasets; cluster-number choice produced gaps of 4.59%, 2.03%, and 3.53%. t-SNE visualization indicated tighter grouping of same-class representations than SimCLR.
Methods
- The framework extends SimCLR with a negative-selection step: representations from one branch are clustered, and instances sharing a cluster label are excluded from negative pairs; a new Cluster-NT-Xent loss masks same-cluster negatives. A hyperparameter α controls what fraction of samples near cluster centers use this strategy.
- Experiments used TPN as encoder, resampling augmentation, a three-layer MLP projection head with 96 neurons, Adam, temperature 0.1, 200 pretraining epochs, and batch sizes of 1024 for USC-HAD and 256 for MotionSense/UCI-HAR. K-means was used for USC-HAD and MotionSense; BIRCH for UCI-HAR, with cluster count defaulting to the true class count.
- Data: USC-HAD (14 subjects, 12 activities, 100 Hz), MotionSense (24 participants, 6 activities, 50 Hz), and UCI-HAR (30 subjects, 6 activities, 50 Hz). Evaluation used mean F1-score, five-fold cross-validation, ten repeated runs, and downstream linear or fine-tuned classifiers with 1% or 10% labels.
Limitations
- The paper states that gains over fully supervised learning with all labels are limited, training time is longer, and clustering quality strongly affects final performance.
- The text also makes evident that clustering method and cluster number substantially change results, batch size and epoch choices are unstable across datasets, and α = 100 is not always optimal.
- The study is limited to three benchmark datasets and does not report real-world deployment; the authors state that real-life applications remain future work.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
1 Valid · 1 Not valid · 7 Not checked · 77 Own results and interpretations
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“Moreover, the labels are affected by various noise sources, such as sensor noise, segmentation problems, and differences in the activities of different people, rendering the annotation process error-prone [11].”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > H UMAN activity recognition (HAR) based on the Internet
What the sources say
- Could not verifycites [11] sourceExperimental — accuracy not yet measured
Our copy of BIRCH is a clustering algorithm paper: its only discussion of 'noise' is the synthetic noise added to generated benchmark datasets ('noise in the form of data points uniformly distributed throughout the overview of the dataset... The parameter rn controls the percentage of data points in the dataset that are considered noise', Abstract, dataset-generator section), plus outlier handling in Phase 4 and the 'outsiders' misplacement issue. Nothing in the paper addresses sensor data, ground-truth annotation/labeling, segmentation problems, or inter-individual differences in activities; the sole mention of labeling is that in Phase 4 'each data point can be labeled with the cluster that it belongs to' (Abstract, Phase 4 discussion), which is algorithm output, not human annotation. Searched: 'sensor data annotation labels noise segmentation activity recognition' and 'labeling error ground truth annotation error-prone' — no passage speaks to the claim, so the attribution can be neither confirmed nor contradicted on our copy. (LLM, conf=0.70)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, the contrastive loss function of the above work is InfoNCE [39] or NT-Xent [40]; both are essentially the instance discrimination task that will regard the same-class samples as negative examples.”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > B. Contrastive learning for HAR
What the sources say
- Partly supportedcites [39] sourceExperimental — accuracy not yet measured
Section 3.1 (p. 6) says the HAR method uses InfoNCE to distinguish a future-timestep positive from negatives drawn from other windows in the batch. Because it does not exclude windows of the same activity class, those windows could be treated as negatives. The paper does not establish the broader claim that both InfoNCE and NT-Xent are instance-discrimination tasks; NT-Xent is attached to a different citation in the citing sentence. (LLM, conf=0.83)
- Not checked yetcites [40]
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“ClusterCLHAR outperforms supervised learning of the corresponding backbone network on the USC-HAD dataset and is inferior to it on the MotionSense and UCI-HAR datasets.”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > A. Self-supervised learning
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“With 1% data labels, ClusterCLHAR outperforms the previous best contrastive learning work by 4.82% for USC-HAD, 3.3% for MotionSense and 4.89% for UCI-HAR with linear evaluation.”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > B. Semi-supervised learning
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“ClusterCLHAR outperforms CSSHAR in self-supervised learning”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > A. Self-supervised learning
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
50 more claims
Create a free account to see all 86 claims and to verify the 10 references still open.
- The quote only partly states this claim. 4
- It may state two separate claims as one. 20
- Probably not a substantial claim. 12
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Both methods were designed to leverage unlabeled data by pre-training on unannotated signals and subsequently fine-tuning on labeled data.”
From Towards User-Generalizable Wearable-Sensor-Based Human Activity Recognition: A Multi-Task Contrastive Learning Approach (2025) · cites this paper as [35]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Other methods, such as CPC [30] and ClusterCLHAR [35], also involved two-stage training.”
From Towards User-Generalizable Wearable-Sensor-Based Human Activity Recognition: A Multi-Task Contrastive Learning Approach (2025) · cites this paper as [35]
In the citing paper
“A well-known limitation of self-supervised contrastive learning is the presence of false negative pairs.”
From Towards User-Generalizable Wearable-Sensor-Based Human Activity Recognition: A Multi-Task Contrastive Learning Approach (2025) · cites this paper as [35]
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.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
“With 1% data labels, ClusterCLHAR outperforms the previous best contrastive learning work by 4.82% for USC-HAD, 3.3% for MotionSense and 4.89% for UCI-HAR with linear evaluation.”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > B. Semi-supervised learning
Reported 3.3 · recomputed 2.93
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > A. Proposed ClusterCLHAR Framework
“l Cluster i,j = -log exp(z i • z j /τ ) 2N k=1 I [k / ∈cluster(i)] exp(z i • z k /τ )(2)”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > VII. CONCLUSIONS
“masks_ab = masks_aa -eye(N) #(N,N) logits_ab = logits_ab -masks_ab * LARGE_NUM #(N,N)”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > B. Semi-supervised learning
“The experimental results show that our proposed framework outperforms all state-of-the-art contrastive learning methods in both linear evaluation and fine-tuning for all three datasets.”
§ Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition > A. Clustering Methods
“MoCoHAR [28] 77.66 85.49 91.72 96.63 [77.56, 77.75] [84.60, 86.36] [91.57, 91.87] [96.36, 96.90] NNCLR [25] 83.39 86.19 92.09 96.40 [82.20, 84.58] [85.13, 87.26] [91.74, 92.44] [96.24, 96.56] ClusterCLHAR (ours) 86.69 87.35 94.02 96.44”
Citations
Cites
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946
(2008) · Journal of Machine Learning Research · Cited by 33,051
(1967) · Cited by 23,719
(1996) · Cited by 20,179
Momentum Contrast for Unsupervised Visual Representation Learning
He, K., Fan, H., Wu, Y., Xie, S., & Girshick, R. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00975 · Cited by 13,347
A simple framework for contrastive learning of visual representations
Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020) · arXiv · doi:10.48550/arxiv.2002.05709 · Cited by 10,857
TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin, M., et al. (2016) · arXiv · doi:10.48550/arxiv.1603.04467 · Cited by 9,408
Representation Learning with Contrastive Predictive Coding
Oord, A. V. D., Li, Y., & Vinyals, O. (2018) · arXiv · doi:10.48550/arxiv.1807.03748 · Cited by 6,242
Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation
Powers, D. M. W. (2020) · arXiv · doi:10.48550/arxiv.2010.16061 · Cited by 5,440
Unsupervised Feature Learning via Non-parametric Instance Discrimination
Wu, Z., Xiong, Y., Yu, S. X., & Lin, D. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00393 · Cited by 3,664
Cited by
Zhou, X., Yang, Q., Zheng, X., Liang, W., Wang, K. I., Ma, J., Pan, Y. F., & Jin, Q. (2024) · IEEE Journal on Selected Areas in Communications · doi:10.1109/jsac.2023.3345431 · Cited by 187
Li, S., Zhu, T., Duan, F., Chen, L., Ning, H., Nugent, C., & Wan, Y. (2024) · IEEE Internet of Things Journal · doi:10.1109/jiot.2024.3463405 · Cited by 83
Self-citationA Multitask Deep Learning Approach for Sensor-Based Human Activity Recognition and Segmentation
Duan, F., Zhu, T., Wang, J., Chen, L., Ning, H., & Wan, Y. (2023) · IEEE Transactions on Instrumentation and Measurement · doi:10.1109/tim.2023.3273673 · Cited by 51
Self-citationEnhancing Human Activity Recognition in Smart Homes with Self-Supervised Learning and Self-Attention
Chen, H., Gouin-Vallerand, C., Bouchard, K., Gaboury, S., Couture, M., Bier, N., & Giroux, S. (2024) · Sensors · doi:10.3390/s24030884 · Cited by 49
Self-supervised Learning for Accelerometer-based Human Activity Recognition: A Survey
Logacjov, A. (2024) · Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies · doi:10.1145/3699767 · Cited by 40
Metadata
- Authors
- Wang, J., Zhu, T., Chen, L. L., Ning, H., & Wan, Y.
- Affiliations
- University of South China (CN)
- University of Ulster (GB)
- University of Science and Technology Beijing (CN)
- Venue
- IEEE Internet of Things Journal
- Published
- 2023-06-15
- Volume
- 10
- Issue
- 12
- Pages
- 10833-10844
- ISSN
- 2327-4662 (ISSN-L)
- 2372-2541
- OpenAlex
- W4318148707
- Cited by
- 52
- References
- 52
- Citations by year
- 2026: 8 · 2025: 13 · 2024: 23 · 2023: 7
- FWCI
- 4.64
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE Internet of Things Journal (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- Ulster University Research Portal (Ulster University) (submittedVersion, public-domain)
- arXiv (Cornell University) (cc-by)
- Funders
- National Natural Science Foundation of China (62006110, 62071213)
- Natural Science Foundation of Hunan Province (2021JJ30574)
- Research Foundation of Education Bureau of Hunan Province (21C0311, 21B0424)
- Hengyang Science and Technology Major Project (202250015428)
- 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
- contrastive learning, human activity recognition, negative selection, clustering, instance discrimination, self-supervised learning, semi-supervised learning, contrastive loss, sensor-based activity recognition
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 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-07 (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 and 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: Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition. https://app.cochat.ai/papers/10.1109/jiot.2023.3239945
- The paper
- Wang, J., Zhu, T., Chen, L. L., Ning, H., & Wan, Y. (2023). Negative Selection by Clustering for Contrastive Learning in Human Activity Recognition. IEEE Internet of Things Journal. https://doi.org/10.1109/jiot.2023.3239945
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.