Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

LLM-generated semantic priors, fused with margin, PageRank, hubness, and facility-location scoring, select a more informative compact set of training windows for few-shot wearable activity recognition, reaching 88.8% macro F1 versus 56.7–62

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

LLM-chosen feature weights and confusability priors improve few-shot exemplar selection for wearable HAR, but the claimed 88.78% is the best model, not the mean (67.02%).

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Plausible idea, but the experiments can't isolate LLM-guided selection from the ML Gate and semantic features, and the abstract's 88.78% is one classifier, not the framework. the 4 major comments →

arxiv 2512.22385 v2 pith:Y4NHYBWT submitted 2025-12-26 cs.CL cs.AIcs.CV

LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition

Elsen Ronando , Sozo Inoue This is my paper
classification cs.CL cs.AIcs.CV
keywords few-shot learningexemplar selectionhuman activity recognitionlarge language modelssemantic priorsfacility locationwearable sensorssmartphone inertial data
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper aims to establish that semantic knowledge from a large language model can improve exemplar selection for few-shot human activity recognition. It argues that purely geometric selection methods, such as random sampling, herding, and k-center, fail for activities with overlapping signal patterns, like walking versus walking upstairs. The proposed framework converts LLM outputs into a structured prior (feature weights, inter-class confusability, and per-class budget multipliers) and a set of semantic feature axes, then scores each candidate with a hybrid function that also includes validation margin, PageRank centrality, and a hubness penalty. A facility-location optimization selects the final exemplars per class. On a standard six-activity smartphone inertial benchmark, the best classifier reaches 88.8% macro F1, and the mean across six classifiers is 67.0%, outperforming the baselines. If correct, the work would show that LLM priors are a practical, low-cost way to choose which sensor windows to label under extreme label scarcity.

Core claim

The central claim is that exemplar selection for few-shot HAR should be a hybrid semantic-structural process rather than a purely geometric one. The framework generates LLM-based semantic features—linear combinations of sensor statistics such as verticality and rotation—and a structured knowledge JSON containing per-class feature weights, an inter-class confusability matrix, and label budget multipliers, all derived only from training-set statistics. Each candidate sample is scored by S(i) = αM(i)+µP(i)−τH(i)+βT(i), where M is the validation margin, P is class-conditional PageRank centrality, H is a hubness penalty, and T is the LLM semantic prior. Exemplars are then chosen via greedy facili

What carries the argument

The load-bearing object is the hybrid scoring function S(i) = αM(i) + µP(i) − τH(i) + βT(i). M(i) is the difference between a candidate's mean cosine similarity to validation windows of its own class and to windows of other classes; P(i) is the PageRank of the candidate in a class-conditional mutual-kNN graph; H(i) counts how often the candidate appears as a neighbor of other samples, penalizing hubs; and T(i) is the LLM semantic prior, which rewards alignment with the LLM's class-specific feature weights and penalizes similarity to confusable classes weighted by the LLM's confusability matrix. The four terms are combined with empirically set weights (α=1.0, µ=0.10, τ=0.10, β=0.15). The resu

Load-bearing premise

The reported superiority over random, herding, and k-center is attributed to the exemplar-selection strategy, but the proposed pipeline also adds LLM semantic features and an ML gate that the baselines do not have, so the isolated contribution of the LLM-guided selection itself is not measured.

What would settle it

Run random, herding, and k-center selection using the exact same feature set (including the LLM semantic features) and the same ML Gate, and compare against the proposed LLM-guided selection without the gate and semantic features; if the macro-F1 gap drops to near zero, the central attribution is wrong. Alternatively, repeat the full comparison on a second inertial dataset with different activity classes; if the LLM-guided advantage disappears or reverses, the result does not generalize.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the method holds, practitioners can choose which sensor windows to label using only training-set statistics and an LLM prompt, reducing annotation cost in real-world wearable deployments.
  • The hybrid scoring suggests geometric and semantic signals are complementary; pure geometry leaves performance on the table, especially for confusable activity classes.
  • The large ablation drop when facility-location is removed (−19.7 F1) indicates that coverage and diversity, not only semantic relevance, are essential; the LLM prior tunes but does not replace structural selection.
  • The ML Gate's +7.8 F1 contribution implies a simple static/dynamic pre-classifier substantially simplifies the few-shot problem, a factor that should be controlled for in future few-shot HAR comparisons.
  • Because the pipeline uses only basic statistical and spectral features, it could in principle be applied to other inertial datasets without architectural changes.

Where Pith is reading between the lines

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

  • A reader might infer that the headline gain over random sampling is not purely a product of LLM-guided selection: the same pipeline also adds LLM semantic features (+2.26 mean F1) and an ML Gate (+7.83 mean F1), so a fair comparison would give baselines the same feature set and gate; the paper does not report such a controlled comparison.
  • The LLM's 'semantic features' are essentially hand-coded linear combinations; the LLM may be serving as a proxy for domain knowledge that a human expert could specify, meaning the real novelty might be the hybrid integration rather than the knowledge content.
  • A natural extension is active learning: the same hybrid scoring could rank unlabeled windows by expected value, making the method applicable to adaptive labeling beyond a fixed budget.
  • If the LLM confusability matrix were replaced with a data-estimated confusion or margin matrix, the method might retain much of its benefit; whether the LLM adds unique signal beyond that is an open empirical question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper proposes LLM-Guided Exemplar Selection for few-shot wearable-sensor HAR. The pipeline uses GPT-4o-mini to generate semantic feature axes and a structured knowledge JSON (label-feature weights, inter-class confusability, label budget multipliers) from training-set class means. Each candidate is scored as S(i) = αM(i) + μP(i) − τH(i) + βT(i), combining validation margin, PageRank centrality, hubness penalty, and an LLM semantic score; per-class exemplars are selected by greedy facility location. At inference, an ML Gate restricts test instances to static or dynamic label sets. Experiments on UCI-HAR with a subject-disjoint split compare Random, Herding, and K-Center across six classifiers, reporting a best macro F1 of 88.79% and a mean of 67.02%.

Significance. If the central claim were established, the paper would make a useful contribution: it demonstrates a concrete mechanism for injecting LLM-derived semantic knowledge into exemplar selection for HAR, with detailed prompts, JSON knowledge structures, and an ablation breakdown. The pipeline is transparent and the appendices give enough detail to reproduce the LLM interaction and selection procedure. However, the current evaluation does not isolate the exemplar-selection component: the baseline comparison also includes LLM semantic features and an ML Gate, and the reported superiority is not consistent across classifiers. The significance therefore hinges on the controlled comparison the paper currently lacks.

major comments (4)
  1. [Abstract and §4.3.1, Table 2] The claim that the method 'consistently outperformed all baselines' is contradicted by Table 2. Linear SVC falls from 74.45 (Random) and 78.50 (Herding) to 72.85, a -1.60 drop, and HistGradientBoosting is a flat tie at 16.99 across all four methods. Please revise the claim to acknowledge the negative and tied cases, or support 'consistent' with a statistical test.
  2. [Table 2 vs. Tables 5–6] The main comparison does not hold other components fixed. Appendix E (Table 6) shows the ML Gate contributes +7.83 mean F1; without it, the proposed method is below Random Sampling for HistGradientBoosting (5.14 vs. 16.99) and Gaussian NB (28.69 vs. 54.64). Appendix C (Table 5) shows semantic features add +2.26. Since Random, Herding, and K-Center are not evaluated with the same gate and semantic features, the Table 2 improvements cannot be attributed to LLM-guided selection. Please add a controlled ablation with gate/features held fixed across selection methods, or reframe the claim as an end-to-end framework comparison.
  3. [Abstract and §5] The headline 'macro F1-score of 88.78%' is the single best classifier (Logistic Regression; Table 2 gives 88.79), not the framework. The framework mean is 67.02. Reporting a classifier-specific peak as 'the framework achieves' is misleading. Please report the mean and spread across classifiers, or explicitly label the number as the best of the six classifiers.
  4. [§3.2 vs. §4.2] The scoring coefficients are given as α=1.0, μ=0.10, τ=0.10, β=0.15 in §3.2, but §4.2 lists α_prior=0.20, μ_cent=0.10, τ_hub=0.10 and introduces λred=0.25 with no mapping between the two sets. In addition, all results come from a single seed (seed=42); without repeated runs or error bars, the magnitude of the reported improvements (e.g., +27.24 for kNN) cannot be assessed. Please reconcile the parameter settings and report repeated-run statistics.
minor comments (6)
  1. [Abstract / Table 2] The abstract reports 88.78%, while Table 2 and the conclusion report 88.79% for Logistic Regression. Please reconcile the number.
  2. [Table 2, HistGradientBoosting] The HGB row shows exactly 16.99 for all four selection methods. This suggests a degenerate or fallback behavior, possibly the uniform fallback in Algorithm 5 (lines 21–24). Please explain why all strategies give an identical value.
  3. [Algorithm 3] The λ1–λ4 weights in Algorithm 3 are not specified. Please map them to the α, μ, τ, β coefficients in Eq. (2) or remove the placeholder notation.
  4. [Algorithm 5 / Eq. (2)] The symbol τ is used both for the hubness penalty in Eq. (2) and for the ML Gate threshold in Algorithm 5. These are different quantities; please rename one to avoid ambiguity.
  5. [§5] The statement that the framework can be applied to WISDM, PAMAP2, and RealWorld HAR is speculative, since no experiments on those datasets are presented. Please mark this as future work or add evidence.
  6. [Throughout] There are several typos and formatting glitches: 'F eature Construction' (§3.2), 'ore details of the algorithm' (§3.4), and 'F eature' in section headings. Please copyedit.

Circularity Check

0 steps flagged

No circular derivation: LLM priors are derived only from training statistics and evaluated on a held-out test split; the ML Gate/semantic-feature confound is an experimental-control issue, not circularity.

full rationale

The claimed result is not derived from its inputs by construction. The LLM semantic axes and knowledge JSON are generated from training-set statistics only (Sec. 4.2: 'using only training-set statistics in the prompts to ensure that no information from the validation or test sets leaks into the LLM'; Algorithms B.1-B.2). Hybrid scoring (Eq. 2) combines validation margin, PageRank, hubness, and the LLM semantic score, and the final Macro-F1 is computed on the subject-disjoint test split (Algorithm B.5, lines 16-26). No equation defines the test F1 in terms of the fitted priors; the selection scores are features of the training/validation pipeline. The paper's self-citations [23-25] are related-work context for LLM-based sensor analysis and HED-LM's limitations; none is invoked to justify the reported result or to force the design choice. The ML Gate and semantic-feature ablations (Tables 5-6) explicitly measure the contributions of components accompanying exemplar selection; whether Random/Herding/K-Center baselines should have received those same components is a legitimate experimental-control concern, not circularity. An internal-consistency problem exists in Table 2 (Linear SVC -1.60, HGB +0.00 vs. the text's 'consistently outperformed all baselines'), but that is a claim-accuracy issue, not a definitional reduction. Therefore no circular step can be quoted, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The ledger reflects that the paper's 'LLM priors' are arbitrary LLM-authored constants plus a small set of validation-tuned coefficients; no new physical entities are introduced. The UCI-HAR split, cosine-similarity choice, and validation-based tuning are domain assumptions that the central comparison depends on.

free parameters (7)
  • Hybrid scoring weights alpha, mu, tau, beta = alpha=1.0, mu=0.10, tau=0.10, beta=0.15 in Eq. 2; Section 4.2 also lists lambda_red=0.25, alpha_prior=0.20, mu_cent=0.10
    Set empirically on the validation split; no sensitivity analysis is reported, and the two sections give inconsistent values for the same names.
  • LLM label_feature_weights = e.g., WALKING: acc_x_mean=0.5, acc_y_mean=-0.5, acc_z_mean=-0.5, gyr_x_std=0.5, gyr_y_std=0.5 (Appendix C.2)
    Output of a GPT-4o-mini prompt using class-wise training means; chosen by the LLM, not derived from test data or independent benchmarks.
  • LLM confusability matrix values = 0.3-0.6 between activity pairs (Appendix C.2)
    LLM-chosen semantic similarities used in Eq. 1 to shape the semantic score; no external validation.
  • LLM label_budget_multipliers and exemplar budgets = WALKING=1.2, WALKING_UPSTAIRS=1.1, WALKING_DOWNSTAIRS=1.1, SITTING=1.0, STANDING=1.0, LAYING=0.9; k_dyn=8, k_stat=2 in S
    LLM-chosen multipliers control per-class exemplar counts, and the concrete k_dyn=8/k_stat=2 values are stated as examples; these directly determine the selected subset.
  • LLM semantic feature weights = verticality=0.8*acc_z_mean+0.6*acc_z_std; rotation=1.5*gyr_z_mean+0.5*gyr_z_std; etc. (Appendix C.1)
    LLM-proposed linear coefficients defining new feature axes; no independent validation or transfer check.
  • k_graph = 10
    Mutual-kNN graph size for PageRank and hubness; fixed without sensitivity analysis.
  • ML Gate threshold tau = 0.50
    Default threshold for static/dynamic routing; no sensitivity analysis.
axioms (5)
  • domain assumption The UCI-HAR subject-wise split (17/4/9) is a valid few-shot benchmark for unseen-user generalization.
    Claims about generalization to unseen users depend on this split being representative; only one split is used.
  • ad hoc to paper The LLM (GPT-4o-mini) output is a valid source of semantic knowledge when prompted with class-wise training means.
    The method's core prior is whatever the LLM returns; there is no check on whether the JSON weights are correct beyond the final F1.
  • domain assumption Validation labels can be used to compute margins and tune coefficients without overfitting that invalidates test results.
    M(i) and coefficient selection use the 1,340-sample validation split; if this tuning overfits the split, the 'unseen-user' claim is weakened.
  • domain assumption Cosine similarity on the combined feature space measures exemplar representativeness.
    Facility-location objective and margin scores both rely on cosine similarity; no alternative similarity is compared.
  • standard math Greedy facility-location maximization selects a representative subset under the chosen budget.
    Standard submodular maximization is assumed for the selection step; the objective is the coverage function of Eq. 4.
invented entities (2)
  • LLM knowledge JSON prior (label_feature_weights, confusability, label_budget_multiplier) no independent evidence
    purpose: Semantic guide for exemplar scoring and per-class budgets.
    Generated by prompting GPT-4o-mini with training-set statistics; no external validation or falsifiable prediction outside this dataset.
  • LLM semantic feature axes (verticality, lateral_movement, rotation, acceleration_magnitude, gyr_std_dev, activity_level) no independent evidence
    purpose: New feature dimensions appended to sensor representations for selection and classification.
    The coefficients are LLM-suggested linear combinations; they have no independent benchmark and no guarantee of transfer to other datasets.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition." pith.science (2026). https://pith.science/paper/Y4NHYBWT

@misc{pith2026251222385,
  author       = {Pith},
  title        = {Pith review of: LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y4NHYBWT}},
  note         = {Machine review of arXiv:2512.22385}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this paper, we propose an LLM-Guided Exemplar Selection framework to address a key limitation in state-of-the-art Human Activity Recognition (HAR) methods: their reliance on large labeled datasets and purely geometric exemplar selection, which often fail to distinguish similar wearable sensor activities such as walking, walking upstairs, and walking downstairs. Our method incorporates semantic reasoning via an LLM-generated knowledge prior that captures feature importance, inter-class confusability, and exemplar budget multipliers, and uses it to guide exemplar scoring and selection. These priors are combined with margin-based validation cues, PageRank centrality, hubness penalization, and facility-location optimization to obtain a compact and informative set of exemplars. Evaluated on the UCI-HAR dataset under strict few-shot conditions, the framework achieves a macro F1-score of 88.78%, outperforming classical approaches such as random sampling, herding, and k-center. The results show that LLM-derived semantic priors, when integrated with structural and geometric cues, provide a stronger foundation for selecting representative sensor exemplars in few-shot wearable-sensor HAR.

Figures

Figures reproduced from arXiv: 2512.22385 by Elsen Ronando, Sozo Inoue.

Figure 1
Figure 1. Figure 1: Motivation for the proposed LLM-Guided Exemplar Selection. (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall workflow of the proposed LLM-Guided Exemplar Selection [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Prompt design for LLM-based semantic feature generation. [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Prompt design for LLM knowledge representation. [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LLM response for semantic feature synthesis prompt. [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: LLM Knowledge (priors): feature weights per label, confusability, [PITH_FULL_IMAGE:figures/full_fig_p030_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Confusion matrices of our proposed approach across six baseline [PITH_FULL_IMAGE:figures/full_fig_p031_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Global UMAP projection of the training set with exemplars selected [PITH_FULL_IMAGE:figures/full_fig_p033_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Per-class UMAP visualizations showing exemplar distributions for [PITH_FULL_IMAGE:figures/full_fig_p035_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Computation time comparison between baseline exemplar selection [PITH_FULL_IMAGE:figures/full_fig_p036_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TRACE: Temporal Reasoning over Context and Evidence for Activity Recognition in Smart Homes

    cs.HC 2026-05 unverdicted novelty 4.0

    TRACE improves activity recognition accuracy and temporal coherence in smart homes by integrating multi-source sensor evidence with contextual priors.

Reference graph

Works this paper leans on

36 extracted references · cited by 1 Pith paper

  1. [1]

    Anguita, Alessandro Ghio, L

    D. Anguita, Alessandro Ghio, L. Oneto, Xavier Parra, and Jorge Luis Reyes-Ortiz. A public domain dataset for human activity recognition using smartphones. InThe European Symposium on Artificial Neural Networks, 2013

    2013

  2. [2]

    Comprehensive machine and deep learning analysis of sensor-based human activ- ity recognition.Neural Computing and Applications, 35(17):12793– 12831, June 2023

    Hossam Magdy Balaha and Asmaa El-Sayed Hassan. Comprehensive machine and deep learning analysis of sensor-based human activ- ity recognition.Neural Computing and Applications, 35(17):12793– 12831, June 2023

    2023

  3. [3]

    Sylvio Barbon Junior, Paolo Ceravolo, Sven Groppe, Mustafa Jar- rar, Samira Maghool, Florence Sèdes, Soror Sahri, and Maurice Van Keulen. Are large language models the new interface for data pipelines? InProceedings of the International Workshop on Big Data in Emergent Distributed Environments, BiDEDE ’24, New York, NY, USA, 2024. Association for Computin...

    2024

  4. [4]

    A com- prehensive study of class incremental learning algorithms for visual tasks.Neural Networks, 135:38–54, 2021

    Eden Belouadah, Adrian Popescu, and Ioannis Kanellos. A com- prehensive study of class incremental learning algorithms for visual tasks.Neural Networks, 135:38–54, 2021

    2021

  5. [5]

    Ensem-har: An ensem- ble deep learning model for smartphone sensor-based human activity recognition for measurement of elderly health monitoring.Biosen- sors, 12(6), 2022

    Debarshi Bhattacharya, Deepak Sharma, Wonjoon Kim, Muham- mad Fazal Ijaz, and Pawan Kumar Singh. Ensem-har: An ensem- ble deep learning model for smartphone sensor-based human activity recognition for measurement of elderly health monitoring.Biosen- sors, 12(6), 2022. LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition IJ...

    2022

  6. [6]

    Carvalho and Rute C

    Liliana I. Carvalho and Rute C. Sofia. A review on scaling mobile sensing platforms for human activity recognition: Challenges and recommendations for future research.IoT, 1(2):451–473, 2020

    2020

  7. [7]

    Improving augmentation efficiency for few-shot learning.IEEE Access, 10:17697–17706, 2022

    Wonhee Cho and Eunwoo Kim. Improving augmentation efficiency for few-shot learning.IEEE Access, 10:17697–17706, 2022

    2022

  8. [8]

    Rosen, Gerbrand Ceder, Kristin A

    John Dagdelen, Alexander Dunn, Sanghoon Lee, Nicholas Walker, Andrew S. Rosen, Gerbrand Ceder, Kristin A. Persson, and Anub- hav Jain. Structured information extraction from scientific text with large language models.Nature Communications, 15(1):1418, Febru- ary 2024

    2024

  9. [9]

    Human activity recognition using inertial, physiological and environmental sensors: A comprehensive survey.IEEE Access, 8:210816–210836, 2020

    Florenc Demrozi, Graziano Pravadelli, Azra Bihorac, and Parisa Rashidi. Human activity recognition using inertial, physiological and environmental sensors: A comprehensive survey.IEEE Access, 8:210816–210836, 2020

    2020

  10. [10]

    Few-shot class-incremental learning via rela- tion knowledge distillation.Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):1255–1263, May 2021

    Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class-incremental learning via rela- tion knowledge distillation.Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):1255–1263, May 2021

    2021

  11. [11]

    Eldrandaly

    Basma K. Eldrandaly. Activbench: Leveraging human activity in- ference from smartphone sensors for human computer interactions. Journal of Cognitive Human-Computer Interaction, 2023

    2023

  12. [12]

    Emilio Ferrara. Large language models for wearable sensor-based human activity recognition, health monitoring, and behavioral mod- eling: A survey of early trends, datasets, and challenges.Sensors, 24(15), 2024

    2024

  13. [13]

    Mgfs: A multi-label graph-based feature selection algorithm via pagerank centrality.Expert Systems with Applications, 142:113024, 2020

    Amin Hashemi, Mohammad Bagher Dowlatshahi, and Hossein Nezamabadi-pour. Mgfs: A multi-label graph-based feature selection algorithm via pagerank centrality.Expert Systems with Applications, 142:113024, 2020

    2020

  14. [14]

    Learning from less data: A unified data subset selection and active learning framework for computer vision, 2019

    Vishal Kaushal, Rishabh Iyer, Suraj Kothawade, Rohan Mahadev, Khoshrav Doctor, and Ganesh Ramakrishnan. Learning from less data: A unified data subset selection and active learning framework for computer vision, 2019

    2019

  15. [15]

    Human activity recognition based on multienvironment sensor data

    Yang Li, Guanci Yang, Zhidong Su, Shaobo Li, and Yang Wang. Human activity recognition based on multienvironment sensor data. Information Fusion, 91:47–63, 2023

    2023

  16. [16]

    Zechen Li, Shohreh Deldari, Linyao Chen, Hao Xue, and Flora D. Salim. Sensorllm: Aligning large language models with motion sen- sors for human activity recognition, 2025

    2025

  17. [17]

    Negative margin matters: Understanding margin in few-shot classification, 2020

    Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters: Understanding margin in few-shot classification, 2020

    2020

  18. [18]

    Category-extensible hu- man activity recognition based on doppler radar by few-shot learn- ing.IEEE Sensors Journal, 22(22):21952–21960, 2022

    Ziyu Liu, Chaoyang Wu, and Wenbin Ye. Category-extensible hu- man activity recognition based on doppler radar by few-shot learn- ing.IEEE Sensors Journal, 22(22):21952–21960, 2022. LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition IJABC: International Journal of Activity and Behavior Computing18

    2022

  19. [19]

    A hybrid tuple selection pipeline for smartphone based human activity recognition.Expert Systems with Applications, 217:119536, 2023

    Ayan Kumar Panja, Adityar Rayala, Abhay Agarwala, Sarmistha Neogy, and Chandreyee Chowdhury. A hybrid tuple selection pipeline for smartphone based human activity recognition.Expert Systems with Applications, 217:119536, 2023

    2023

  20. [20]

    Automatic combination of sample selection strategies for few-shot learning, 2024

    Branislav Pecher, Ivan Srba, Maria Bielikova, and Joaquin Van- schoren. Automatic combination of sample selection strategies for few-shot learning, 2024

    2024

  21. [21]

    A weighted k-center algorithm for data subset selection, 2023

    Srikumar Ramalingam, Pranjal Awasthi, and Sanjiv Kumar. A weighted k-center algorithm for data subset selection, 2023

    2023

  22. [22]

    Learning optimal inter-class margin adaptively for few-shot class-incremental learning via neural collapse-based meta-learning

    Hang Ran, Weijun Li, Lusi Li, Songsong Tian, Xin Ning, and Prayag Tiwari. Learning optimal inter-class margin adaptively for few-shot class-incremental learning via neural collapse-based meta-learning. Information Processing & Management, 61(3):103664, 2024

    2024

  23. [23]

    Elsen Ronando and Sozo Inoue. Improving fatigue detection with fea- ture engineering on physical activity accelerometer data using large language models.International Journal of Activity and Behavior Computing, 2024(2):1–22, 2024

    2024

  24. [24]

    Leveraging large language models to enhance understanding of accelerometer data on physical fatigue detection question answering

    Elsen Ronando and Sozo Inoue. Leveraging large language models to enhance understanding of accelerometer data on physical fatigue detection question answering. In112th Mobile Computing and New Social Systems, 83rd Ubiquitous Computing Systems, 41st Consumer Devices & Systems, 30th Aging Society Design Joint Research, 2024

    2024

  25. [25]

    Few-shot optimization for sensor data using large language models: A case study on fatigue detection

    Elsen Ronando and Sozo Inoue. Few-shot optimization for sensor data using large language models: A case study on fatigue detection. Sensors, 25(11), 2025

    2025

  26. [26]

    Structuredrag: Json response formatting with large language mod- els, 2024

    Connor Shorten, Charles Pierse, Thomas Benjamin Smith, Erika Cardenas, Akanksha Sharma, John Trengrove, and Bob van Luijt. Structuredrag: Json response formatting with large language mod- els, 2024

    2024

  27. [27]

    A hubness information-based k-nearest neighbor approach for multi- label learning.Mathematics, 13(7), 2025

    Zeyu Teng, Shanshan Tang, Min Huang, and Xingwei Wang. A hubness information-based k-nearest neighbor approach for multi- label learning.Mathematics, 13(7), 2025

    2025

  28. [28]

    Cross- domain har: Few-shot transfer learning for human activity recogni- tion.ACM Trans

    Megha Thukral, Harish Haresamudram, and Thomas Plötz. Cross- domain har: Few-shot transfer learning for human activity recogni- tion.ACM Trans. Intell. Syst. Technol., 16(1), January 2025

    2025

  29. [29]

    A new deep-learning method for human activity recogni- tion.Sensors, 23(5), 2023

    Roberta Vrskova, Patrik Kamencay, Robert Hudec, and Peter Sykora. A new deep-learning method for human activity recogni- tion.Sensors, 23(5), 2023

    2023

  30. [30]

    Stankovic

    Lahiru Wijayasingha and John A. Stankovic. Generalized few-shot learning for wearable sensor-based human activity recognition. In 2022 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops), pages 328–334, 2022

    2022

  31. [31]

    Adnan Yazici, Dana Zhumabekova, Aidana Nurakhmetova, Zhanggir Yergaliyev, Hakan Yekta Yatbaz, Zaida Makisheva, Michael Lewis, and Enver Ever. A smart e-health framework for monitoring the LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition IJABC: International Journal of Activity and Behavior Computing19 health of the eld...

    2023

  32. [32]

    Csi-based human activity recognition with graph few-shot learning.IEEE Internet of Things Journal, 9(6):4139–4151, 2022

    Yong Zhang, Yang Chen, Yujie Wang, Qingqing Liu, and Andong Cheng. Csi-based human activity recognition with graph few-shot learning.IEEE Internet of Things Journal, 9(6):4139–4151, 2022

    2022

  33. [33]

    features

    Yifan Zhou and Lei Yu. Few-shot learning via weighted prototypes from graph structure.Pattern Recognition Letters, 176:230–235, 2023. LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition IJABC: International Journal of Activity and Behavior Computing20 A Appendix: LLM Prompting Design LLM-based semantic feature prompting <...

    2023

  34. [34]

    Label-F eature W eights.This part assigns a numerical impor- tance score to each feature for every activity class. For example, for the classWALKING, the LLM assigns higher positive weights to the stan- dard deviation of gyroscope signals (gyr_x_std, gyr_y_std) and mod- erate positive weights to acc_x_mean, indicating that walking involves rhythmic latera...

  35. [35]

    label_feature_weights

    Class Confusability .This component quantifies the degree of similarity between classes, with values ranging from 0 (distinct) to 1 (highly similar). As expected, activities such asWALKING,WALK- ING_UPSTAIRS, andWALKING_DOWNSTAIRSexhibit high mutual confusability values (0.5–0.6), reflecting their overlapping motion pat- terns. In contrast, static–dynamic...

  36. [36]

    Label Budget Multiplier.Finally, this component controls the number of exemplars allocated per class based on activity characteris- tics. Dynamic activities (WALKING,WALKING_UPSTAIRS,WALK- ING_DOWNSTAIRS) receive slightly higher multipliers (1.1–1.2), allow- ing more exemplars to represent their complex and variable movements. In contrast, static classes ...

    1967

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.