pith. sign in

arxiv: 2606.31532 · v1 · pith:QYLNHWLOnew · submitted 2026-06-30 · 💻 cs.AI

A time-series classification framework for individual-level absenteeism prediction under severe class imbalance

Kwong Ho Li , Matthew Roughan , Wathsala Karunarathne This is my paper

Pith reviewed 2026-07-01 05:34 UTC · model grok-4.3

classification 💻 cs.AI
keywords absenteeism predictiontime series classificationclass imbalanceLSTM-FCNfocal lossgeometric mean lossproactive predictionworkforce planning
0
0 comments X

The pith

A TSC framework predicts future individual absences by separating historical sequences from future labels and applying LSTM-FCN with imbalance-aware losses.

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

The paper establishes a framework for predicting future staff absences at the individual level by treating attendance histories as time series and forecasting ahead, rather than classifying current states. This matters because existing methods only reproduce known outcomes and ignore sequential patterns in attendance data. The authors use a simulated dataset based on UCI data to handle the lack of public longitudinal records and test losses designed for class imbalance. They find that LSTM-FCN combined with Binary Focal Loss or Geometric Mean loss achieves strong specificity and balanced accuracy around 80 percent on held-out data. Performance depends mainly on the balancing parameter in focal loss, while geometric mean loss requires no tuning.

Core claim

The central claim is that separating historical attendance sequences from future absence labels via a TSC framework, combined with LSTM-FCN architecture and imbalance losses like BFL with alpha set to 1/(1+rho) or G-Mean, enables proactive prediction with specificity 0.813 and balanced accuracy 0.888, yielding about 80% on test data.

What carries the argument

The LSTM-Fully Convolutional Network (LSTM-FCN) architecture paired with Binary Focal Loss or Geometric Mean loss, applied to separated historical time series and future labels.

Load-bearing premise

The simulated dataset calibrated to the UCI collection accurately captures the sequential behavioural structure and severe class imbalance of real individual attendance histories without introducing artifacts or data leakage.

What would settle it

Evaluating the model on actual longitudinal attendance data from a real organization and observing whether the balanced accuracy remains near 80% or drops significantly due to differences in sequence structure.

Figures

Figures reproduced from arXiv: 2606.31532 by Kwong Ho Li, Matthew Roughan, Wathsala Karunarathne.

Figure 1
Figure 1. Figure 1: Illustration of the sliding window formulation for time series prediction without staff background information. The past n consecutive days of attendance form the input sequence X, and the subsequent m days form the target label y. To construct the TSC dataset, we implemented a custom function to transform daily attendance records into fixed-length time series segments and to partition the data into traini… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the sliding window formulation for time series prediction with staff background information. As in [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of batch size on validation performance of the LSTM-FCN model trained without background information. Batch sizes range from 16 to 1,024 in doublings. All runs use window size 80, output horizon 5 days, G-Mean loss (eps = 10−12, use log = True), learning rate 10−3 , and early stopping with patience 10. G-Mean loss has no α or γ parameters; no loss function assumptions are made in this sweep. size sw… view at source ↗
Figure 4
Figure 4. Figure 4: Effect of input window size on validation performance across all evaluated model architectures, with and without background information, for a 5-day prediction horizon. Precision, F1-score, specificity, and balanced accuracy are reported for window sizes ranging from 5 to 320 days. All models were trained using a batch size of 1,024 (established in Section 4.1), output horizon 5 days, G-Mean loss (eps = 10… view at source ↗
Figure 5
Figure 5. Figure 5: Present-to-absent gradient norm ratio (log scale) per epoch under BFL (α = 0.75, γ = 2, baseline parameterisation) and G-Mean loss, computed on the LSTM-FCN model with window size 80, batch size 1,024, without background information. Both models are initialised from the same random seed. The orange colour dashed reference line shows the theoretical prediction of ρα/(1 − α) ≈ 119 for the baseline BFL config… view at source ↗
Figure 6
Figure 6. Figure 6: Empirical mean predicted probability p per epoch for the present class, plotted against the theoretical gradient ratio curve ρα/(1 − α) · g1/g0 for the baseline BFL config￾uration (α = 0.75, γ = 2, ρ = 39.5 from the training partition). Dots represent the actual mean p at each training epoch; colour indicates epoch number (light to dark). Both BFL and G-Mean operate in the range p ≈ 0.8–0.87 throughout tra… view at source ↗
Figure 7
Figure 7. Figure 7: Effect of α on the BFL gradient ratio (ρ = 42.31 computed from data; result at p = 0.5 holds for any γ). Left: Pure focusing ratio g1/g0 across p for all γ values, independent of α and ρ. All curves converge to g1/g0 = 1 exactly at p = 0.5, proving by symmetry that γ has zero effect at initialisation. When p < 0.5 (model wrong about present employees), γ amplifies majority dominance; when p > 0.5 (model co… view at source ↗
Figure 8
Figure 8. Figure 8: Effect of γ on the BFL gradient ratio (ρ = 42.31 computed from data; no α or γ assumed). Left: Pure focusing ratio g1/g0 across p for all γ values, identical to [PITH_FULL_IMAGE:figures/full_fig_p027_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Mean predicted probability p = σ(z) per class during training across 20 epochs under BFL loss (left, baseline parameterisation α = 0.75, γ = 2) and G-Mean loss (right), measured on the LSTM-FCN model. Shaded bands indicate ±1 standard deviation. Under BFL, the absent-class mean p starts at 0.775 at epoch 1 and declines to 0.62 by epoch 20, remaining above the decision boundary of 0.5 throughout training, i… view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of p per class at epochs 1, 2, 5, 10, and 20 under BFL loss with baseline parameterisation (α = 0.75, γ = 2). These figures demonstrate the pre￾dicted probability distributions that produce the specificity collapse of 0.124 observed in [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of p per class at epochs 1, 2, 5, 10, and 20 under G-Mean loss (empirical; no assumptions about α or γ). Top row (present, y = 1, majority): Bimodal distribution with a large spike near p ≈ 1 and a smaller spike near p ≈ 0, reflecting confident predictions in both directions. Bottom row (absent, y = 0, minority): The distribution collapses to a sharp spike near p ≈ 0 from epoch 2 onwards, wit… view at source ↗
read the original abstract

Staff absenteeism imposes substantial operational costs in high-demand work environments such as healthcare, emergency services, meat processing, construction, and courier and delivery services, where proactive workforce planning depends on reliable individual-level absence prediction. Existing regression and classification approaches share a structural limitation; they map features observed at time t to labels at the same time t, reproducing already-realised outcomes rather than predicting future events, and discard the sequential behavioural structure inherent in individual attendance histories. We propose a Time Series Classification (TSC) framework that separates historical attendance sequences from future absence labels, enabling genuinely proactive prediction. Due to the lack of public longitudinal attendance data, we construct a reproducible simulated dataset calibrated to the UCI dataset. We analyse Binary Focal Loss (BFL) and Geometric Mean (G-Mean) loss under severe class imbalance using only the imbalance ratio $\rho$. For BFL, the initial gradient ratio is $\rho\alpha/(1-\alpha)$, implying the balanced weight $\alpha = 1/(1+\rho) \approx 0.023$. Experiments show that performance is governed mainly by $\alpha$, with BFL achieving specificity 0.813 and balanced accuracy 0.888, comparable to G-Mean. Unlike BFL, G-Mean adapts automatically without parameter calibration. Among three deep learning architectures evaluated, Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and the hybrid LSTM-Fully Convolutional Network (LSTM-FCN), the LSTM-FCN delivers strong precision and specificity. Stable performance is obtained with batch sizes >= 64 and window sizes between 40-80 days, yielding balanced accuracy of approximately 80% on held-out test data.

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.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a time-series classification framework for proactive individual-level absenteeism prediction that separates historical attendance sequences from future labels. It constructs a reproducible simulated dataset calibrated to the UCI Absenteeism collection, derives a balancing parameter α = 1/(1+ρ) ≈ 0.023 for Binary Focal Loss from the initial gradient ratio, compares it to Geometric Mean loss, and evaluates LSTM, CNN, and LSTM-FCN architectures under severe class imbalance, reporting specificity 0.813, balanced accuracy 0.888, and approximately 80% balanced accuracy on held-out test data with the LSTM-FCN.

Significance. If the simulation faithfully reproduces real longitudinal temporal structure without leakage, the framework would address a practical gap in proactive workforce planning for high-demand sectors. The automatic adaptation of G-Mean loss without parameter tuning and the explicit separation of history from future labels are strengths; the derivation of α directly from ρ is parameter-free in the stated sense.

major comments (2)
  1. [Abstract] Abstract and data-construction description: the simulation is characterized only as 'reproducible simulated dataset calibrated to the UCI dataset' without specifying whether it employs per-individual transition probabilities, higher-order Markov structure, or merely marginal ρ; this is load-bearing for the central claim because the proactive TSC performance (specificity 0.813, balanced accuracy 0.888) cannot be interpreted without evidence that temporal autocorrelation and absence clustering are preserved and that input windows contain no information about future labels.
  2. [Abstract] Abstract: the reported metrics (specificity 0.813, balanced accuracy 0.888, ~80% on held-out test) are stated without error bars, confidence intervals, statistical tests against baselines, or ablation details on window size and batch-size effects, making it impossible to assess whether the LSTM-FCN advantage is robust or an artifact of the particular simulated realization.
minor comments (1)
  1. [Abstract] Notation: ρ is introduced without an explicit definition in the abstract even though it is the sole parameter used for both losses.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful comments, which help improve the clarity and robustness of our manuscript. We address each major comment below and will incorporate revisions accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract and data-construction description: the simulation is characterized only as 'reproducible simulated dataset calibrated to the UCI dataset' without specifying whether it employs per-individual transition probabilities, higher-order Markov structure, or merely marginal ρ; this is load-bearing for the central claim because the proactive TSC performance (specificity 0.813, balanced accuracy 0.888) cannot be interpreted without evidence that temporal autocorrelation and absence clustering are preserved and that input windows contain no information about future labels.

    Authors: We acknowledge that the abstract and data-construction description lack sufficient detail on the simulation methodology. We will revise the manuscript to explicitly describe the simulation process, specifying the use of per-individual transition probabilities derived from the UCI Absenteeism dataset. This approach preserves temporal autocorrelation and absence clustering while ensuring that input windows contain no information about future labels, as the framework separates historical attendance sequences from future absence labels. revision: yes

  2. Referee: [Abstract] Abstract: the reported metrics (specificity 0.813, balanced accuracy 0.888, ~80% on held-out test) are stated without error bars, confidence intervals, statistical tests against baselines, or ablation details on window size and batch-size effects, making it impossible to assess whether the LSTM-FCN advantage is robust or an artifact of the particular simulated realization.

    Authors: We agree that including measures of variability and ablation studies would enhance the assessment of robustness. In the revised manuscript, we will report standard deviations or confidence intervals from multiple experimental runs, include statistical tests comparing to baselines, and provide additional ablation results on the effects of window sizes (40-80 days) and batch sizes (>=64). This will demonstrate that the LSTM-FCN performance is stable and not an artifact of a single realization. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The alpha derivation for Binary Focal Loss follows directly from the standard initial gradient ratio formula using only the given imbalance ratio rho and is independent of any fitted model or test performance. G-Mean loss requires no parameter tuning. The simulated dataset is calibrated to UCI statistics but the reported metrics are measured on held-out test data under explicit temporal separation of history from future labels; no equation reduces the performance numbers to quantities defined solely by tuning on the test set. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes. The central claims rest on empirical evaluation rather than self-referential definitions or fitted-input predictions.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the simulated dataset being representative of real attendance dynamics and on the TSC formulation avoiding temporal leakage; these are domain assumptions rather than derived quantities.

free parameters (1)
  • alpha = 1/(1+rho) approx 0.023
    Derived analytically from gradient ratio but stated to govern performance; window size (40-80 days) and batch size (>=64) are also selected experimentally.
axioms (1)
  • domain assumption The simulated dataset calibrated to the UCI collection accurately captures sequential behavioural structure and class imbalance of real attendance histories.
    Invoked to justify use of the dataset for evaluating proactive prediction performance.

pith-pipeline@v0.9.1-grok · 5840 in / 1479 out tokens · 39739 ms · 2026-07-01T05:34:11.316804+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

39 extracted references · 25 canonical work pages · 4 internal anchors

  1. [1]

    Classification model for reducing absenteeism of nurses at hos- pitals using machine learning and artificial neural network techniques

    Dalia Alzu’bi et al. “Classification model for reducing absenteeism of nurses at hos- pitals using machine learning and artificial neural network techniques”. In:Interna- tional Journal of System Assurance Engineering and Management15.7 (July 2024), pp. 3266–3278.issn: 0976-4348.doi:10.1007/s13198-024-02334-7.url:https: //doi.org/10.1007/s13198-024-02334-7

    work page doi:10.1007/s13198-024-02334-7.url:https: 2024

  2. [2]

    Trends, variations and prediction of staff sickness absence rates among NHS ambulance services in England: a time series study

    Zahid B Asghar et al. “Trends, variations and prediction of staff sickness absence rates among NHS ambulance services in England: a time series study”. In:BMJ Open11.9 (2021).issn: 2044-6055.doi:10.1136/bmjopen- 2021- 053885. eprint: https : / / bmjopen . bmj . com / content / 11 / 9 / e053885 . full . pdf.url:https : //bmjopen.bmj.com/content/11/9/e053885

    work page doi:10.1136/bmjopen- 2021

  3. [3]

    The Great Time Series Classification Bake Off: An Experimental Evaluation of Recently Proposed Algorithms. Extended Version

    Anthony Bagnall et al.The Great Time Series Classification Bake Off: An Experi- mental Evaluation of Recently Proposed Algorithms. Extended Version. 2016. arXiv: 1602.01711 [cs.LG].url:https://arxiv.org/abs/1602.01711

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  4. [4]

    Psychosocial work stress and occupational stres- sors in emergency medical services

    Rupkatha Bardhan and Traci Byrd. “Psychosocial work stress and occupational stres- sors in emergency medical services”. In:Healthcare. Vol. 11. 7. MDPI. 2023, p. 976

    2023

  5. [5]

    Bishop.Pattern Recognition and Machine Learning

    Christopher M. Bishop.Pattern Recognition and Machine Learning. Information Sci- ence and Statistics. New York: Springer, 2006

    2006

  6. [6]

    Shikha Bordia.Bonafide at LegalLens 2024 Shared Task: Using Lightweight DeBERTa Based Encoder For Legal Violation Detection and Resolution. 2024. arXiv:2410 . 22977 [cs.CL].url:https://arxiv.org/abs/2410.22977

    work page arXiv 2024

  7. [7]

    A deep learning approach for intrusion detection in Internet of Things using focal loss function

    Ayesha S. Dina, A.B. Siddique, and D. Manivannan. “A deep learning approach for intrusion detection in Internet of Things using focal loss function”. In:Internet of Things22 (2023), p. 100699.issn: 2542-6605.doi:https://doi.org/10.1016/ j.iot.2023.100699.url:https://www.sciencedirect.com/science/article/ pii/S2542660523000227

    work page arXiv 2023

  8. [8]

    Stress, fear, and anxiety among construction workers: a systematic review

    Carlos Gomez-Salgado et al. “Stress, fear, and anxiety among construction workers: a systematic review”. In:Frontiers in Public Health11 (2023), p. 1226914. 32

    2023

  9. [9]

    Priya Goyal et al.Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

  10. [10]

    arXiv:1706.02677 [cs.CV].url:https://arxiv.org/abs/1706.02677

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Semi-Supervised Bayesian GANs with Log-Signatures for Uncertainty- Aware Credit Card Fraud Detection

    David Hirnschall. “Semi-Supervised Bayesian GANs with Log-Signatures for Uncertainty- Aware Credit Card Fraud Detection”. In:Mathematics13.19 (2025).issn: 2227-7390. doi:10.3390/math13193229.url:https://www.mdpi.com/2227- 7390/13/19/ 3229

    work page doi:10.3390/math13193229.url:https://www.mdpi.com/2227- 2025

  12. [12]

    Deep learning for time series classification: a review

    Hassan Ismail Fawaz et al. “Deep learning for time series classification: a review”. In: Data Mining and Knowledge Discovery33.4 (July 2019), pp. 917–963.issn: 1573- 756X.doi:10.1007/s10618- 019- 00619- 1.url:https://doi.org/10.1007/ s10618-019-00619-1

    work page doi:10.1007/s10618- 2019

  13. [13]

    InceptionTime: Finding AlexNet for time series classi- fication

    Hassan Ismail Fawaz et al. “InceptionTime: Finding AlexNet for time series classi- fication”. In:Data Mining and Knowledge Discovery34.6 (Sept. 2020), 1936–1962. issn: 1573-756X.doi:10.1007/s10618-020-00710-y.url:http://dx.doi.org/ 10.1007/s10618-020-00710-y

    work page doi:10.1007/s10618-020-00710-y.url:http://dx.doi.org/ 2020

  14. [14]

    Time Series-Based Fault Detection and Classification in IEEE 9-Bus Transmission Lines Using Deep Learning

    Somchat Jiriwibhakorn and Shazia Kanwal. “Time Series-Based Fault Detection and Classification in IEEE 9-Bus Transmission Lines Using Deep Learning”. In:IEEE Access13 (2025), pp. 118135–118146.doi:10.1109/ACCESS.2025.3586045

    work page doi:10.1109/access.2025.3586045 2025

  15. [15]

    LSTM Fully Convolutional Networks for Time Series Classifi- cation

    Fazle Karim et al. “LSTM Fully Convolutional Networks for Time Series Classifi- cation”. In:IEEE Access6 (2018), pp. 1662–1669.doi:10 . 1109 / ACCESS . 2017 . 2779939

    2018

  16. [16]

    Multivariate LSTM-FCNs for time series classification

    Fazle Karim et al. “Multivariate LSTM-FCNs for time series classification”. In:Neu- ral Networks116 (2019), pp. 237–245.issn: 0893-6080.doi:https://doi.org/10. 1016/j.neunet.2019.04.014.url:https://www.sciencedirect.com/science/ article/pii/S0893608019301200

    2019

  17. [17]

    The effects of emergency medical service work on the psycho- logical, physical, and social well-being of ambulance personnel: a systematic review of qualitative research

    Sharon Lawn et al. “The effects of emergency medical service work on the psycho- logical, physical, and social well-being of ambulance personnel: a systematic review of qualitative research”. In:BMC psychiatry20.1 (2020), p. 348

    2020

  18. [18]

    Focal loss for dense object detection

    Tsung-Yi Lin et al. “Focal loss for dense object detection”. In:Proceedings of the IEEE international conference on computer vision. 2017, pp. 2980–2988

    2017

  19. [19]

    Lindell, Marcus and Schwerin, Casper.Processing, Modeling, and Forecasting: A Time Series Analysis of Sick Leave Absences in Sweden and Evaluating the Impact of Macro Factors. eng. Student Paper. 2023

    2023

  20. [20]

    Predicting workplace absenteeism using machine learning: a pilot study in occupational health

    Pablo Llamas Bl´ azquez. “Predicting workplace absenteeism using machine learning: a pilot study in occupational health”. In:Journal of Occupational Medicine and Toxicology20.1 (2025), p. 38.doi:10.1186/s12995-025-00482-5

    work page doi:10.1186/s12995-025-00482-5 2025

  21. [21]

    Supporting Sustainable Workforce Manage- ment for Worker Illness Absence Through Predictive Analytics

    Ida Lumintu and Achmad Maududie. “Supporting Sustainable Workforce Manage- ment for Worker Illness Absence Through Predictive Analytics”. In:Engineering Proceedings84.1 (2025).issn: 2673-4591.doi:10.3390/engproc2025084017.url: https://www.mdpi.com/2673-4591/84/1/17. 33

    work page doi:10.3390/engproc2025084017.url: 2025

  22. [22]

    How should clinicians respond to patients experiencing ongoing present traumatic stress of industrial meat production?

    Rachel MacNair. “How should clinicians respond to patients experiencing ongoing present traumatic stress of industrial meat production?” In:AMA Journal of Ethics 25.4 (2023), pp. 251–255

    2023

  23. [23]

    UCI Machine Learn- ing Repository

    Andrea Martiniano and Ricardo Ferreira.Absenteeism at work. UCI Machine Learn- ing Repository. DOI: https://doi.org/10.24432/C5X882. 2012

    work page doi:10.24432/c5x882 2012

  24. [24]

    Incorporating a Machine Learning Model into a Web-Based Ad- ministrative Decision Support Tool for Predicting Workplace Absenteeism

    Gopal Nath et al. “Incorporating a Machine Learning Model into a Web-Based Ad- ministrative Decision Support Tool for Predicting Workplace Absenteeism”. In:In- formation13.7 (2022).issn: 2078-2489.doi:10.3390/info13070320.url:https: //www.mdpi.com/2078-2489/13/7/320

    work page doi:10.3390/info13070320.url:https: 2022

  25. [25]

    Predicting health-related absenteeism with machine learning: A case study

    Aleksander Piciga and Matjaˇ z Kukar. “Predicting health-related absenteeism with machine learning: A case study”. In: (2024)

    2024

  26. [26]

    An Integrated AI Framework for Occupational Health: Predicting Burnout, Long COVID, and Extended Sick Leave in Healthcare Workers

    Maria Valentina Popa et al. “An Integrated AI Framework for Occupational Health: Predicting Burnout, Long COVID, and Extended Sick Leave in Healthcare Workers”. In:Healthcare13.18 (2025).issn: 2227-9032.doi:10.3390/healthcare13182266. url:https://www.mdpi.com/2227-9032/13/18/2266

    work page doi:10.3390/healthcare13182266 2025

  27. [27]

    Towards effective classification of imbalanced data with convolutional neural networks

    Vidwath Raj, Sven Magg, and Stefan Wermter. “Towards effective classification of imbalanced data with convolutional neural networks”. In:Artificial Neural Networks in Pattern Recognition. Springer. 2016, pp. 150–162

    2016

  28. [28]

    An explicit formula for the A-polynomial of twist knots

    Mayakkannan Raman, Narentheren Kaliappen, and Choo Ling Suan. “A Study on Machine Learning Classifier Models in Analyzing Discipline of Individuals Based on Various Reasons Absenteeism from Work”. In:2020 International Conference on Decision Aid Sciences and Application (DASA). 2020, pp. 360–364.doi:10.1109/ DASA51403.2020.9317017

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  29. [29]

    The psychological impact of slaughterhouse em- ployment: A systematic literature review

    Jessica Slade and Emma Alleyne. “The psychological impact of slaughterhouse em- ployment: A systematic literature review”. In:Trauma, Violence, & Abuse24.2 (2023), pp. 429–440

    2023

  30. [30]

    Don't Decay the Learning Rate, Increase the Batch Size

    Samuel L. Smith et al.Don’t Decay the Learning Rate, Increase the Batch Size. 2018. arXiv:1711.00489 [cs.LG].url:https://arxiv.org/abs/1711.00489

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  31. [31]

    A Deep Convolutional Neural Network for Time Series Classification with Intermediate Targets

    Aboozar Taherkhani, Georgina Cosma, and T. M. McGinnity. “A Deep Convolutional Neural Network for Time Series Classification with Intermediate Targets”. In:SN Computer Science4.6 (Oct. 2023), p. 832.issn: 2661-8907.doi:10.1007/s42979- 023-02159-4.url:https://doi.org/10.1007/s42979-023-02159-4

    work page doi:10.1007/s42979- 2023

  32. [32]

    Time series activity classifica- tion using gated recurrent units

    Yi-Fei Tan, Xiaoning Guo, and Soon-Chang Poh. “Time series activity classifica- tion using gated recurrent units”. In:International Journal of Electrical and Com- puter Engineering (IJECE)11 (Aug. 2021), p. 3551.doi:10.11591/ijece.v11i4. pp3551-3558

    work page doi:10.11591/ijece.v11i4 2021

  33. [33]

    Predictive analysis of absen- teeism in MNCs using machine learning algorithm

    Krittika Tewari, Shriya Vandita, and Shruti Jain. “Predictive analysis of absen- teeism in MNCs using machine learning algorithm”. In:Proceedings of ICRIC 2019. Springer. 2020, pp. 3–14.doi:10.1007/978-3-030-29407-6_1. 34

    work page doi:10.1007/978-3-030-29407-6_1 2019

  34. [34]

    Predicting Absenteeism at Work Using Tree-Based Learners

    Zaman Wahid et al. “Predicting Absenteeism at Work Using Tree-Based Learners”. In:Proceedings of the 3rd International Conference on Machine Learning and Soft Computing. ICMLSC ’19. Association for Computing Machinery, 2019, 7–11.isbn: 9781450366120.doi:10 . 1145 / 3310986 . 3310994.url:https : / / doi . org / 10 . 1145/3310986.3310994

    work page arXiv 2019

  35. [35]

    Time series classification from scratch with deep neural networks: A strong baseline

    Zhiguang Wang, Weizhong Yan, and Tim Oates. “Time series classification from scratch with deep neural networks: A strong baseline”. In: May 2017, pp. 1578–1585. doi:10.1109/IJCNN.2017.7966039

    work page doi:10.1109/ijcnn.2017.7966039 2017

  36. [36]

    An equivalence between generalized Maxwell model and fractional Zener model, Mechanics of Materials 100:148-153 (2016)

    Md Doulotuzzaman Xames. “Toward a socially sustainable workforce in service op- erations: Predicting employee burnout with ensemble machine learning”. In:Sustain- able Manufacturing and Service Economics4 (2025), p. 100033.doi:10.1016/j. smse.2025.100033

    work page doi:10.1016/j 2025

  37. [37]

    Prototype Learning for Medical Time Series Classification via Hu- man–Machine Collaboration

    Jia Xie et al. “Prototype Learning for Medical Time Series Classification via Hu- man–Machine Collaboration”. In:Sensors24.8 (2024).issn: 1424-8220.doi:10 . 3390/s24082655.url:https://www.mdpi.com/1424-8220/24/8/2655

    2024

  38. [38]

    A Transformer-based Framework for Multivariate Time Series Representation Learning

    George Zerveas et al. “A Transformer-based Framework for Multivariate Time Series Representation Learning”. In:Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. KDD ’21. Association for Computing Machin- ery, 2021, 2114–2124.isbn: 9781450383325.doi:10.1145/3447548.3467401.url: https://doi.org/10.1145/3447548.3467401

    work page doi:10.1145/3447548.3467401.url: 2021

  39. [39]

    Excess and reduced work absence during COVID-19 in Poland: insights from cause-specific time-series models

    B la˙ zej Lyszczarz and Jakub Wojtasik. “Excess and reduced work absence during COVID-19 in Poland: insights from cause-specific time-series models”. In:Population Health Metrics23.1 (July 2025), p. 35.issn: 1478-7954.doi:10.1186/s12963-025- 00400-1.url:https://doi.org/10.1186/s12963-025-00400-1. 35

    work page doi:10.1186/s12963-025- 2025