Pith. sign in

REVIEW 5 major objections 5 minor 30 references

Designing Adaptive Algorithms Based on Reinforcement Learning for Dynamic Optimization of Sliding Window Size in Multi-Dimensional Data Streams

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes RL-Window, a reinforcement learning agent that dynamically selects sliding window sizes for multi-dimensional data streams, claiming it outperforms ADWIN and CNN-based adaptive methods in accuracy, drift robustness, and…

desk verdict Plausible RL-window-sizing idea, but the method is internally inconsistent and the evaluation is irreproducible as written. read the letter →

arxiv 2507.06901 v1 pith:PUO4JTCN submitted 2025-07-09 cs.LG

Abolfazl Zarghani , Sadegh Abedi This is my paper
classification cs.LG
keywords reinforcementlearningslidingwindowdatastreamsconceptdriftduelingDQNadaptivewindowingmulti-dimensionaltimeseriesprioritizedexperiencereplay
verification ladder T0 review T1 audit T2 compute T3 formal

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 proposes RL-Window, a reinforcement learning agent that dynamically selects sliding window sizes for multi-dimensional data streams. The agent observes statistical features of the stream, such as variance, inter-dimension correlations, rate of change, distributional entropy, and drift signals, and picks from a discrete set of window sizes to maximize a reward that balances classification accuracy, computational cost, and window stability. The authors claim this learned policy outperforms fixed-size windows, the statistical method ADWIN, and supervised CNN-based window selection on three benchmark datasets, while maintaining competitive computational cost. The contribution matters because fixed windows fail under concept drift, and existing adaptive methods either ignore inter-dimension dependencies or need labeled data.

What carries the argument

The core object is the Dueling DQN agent with a composite reward. The state is a hand-engineered vector of stream statistics computed over the last m data points: per-dimension variance, pairwise correlations, rate of change, distributional entropy, out-of-order indicators, spectral features, and drift signals. The action space is a discrete set of window sizes (20 to 200). The reward combines a classification-correctness term, a computational-cost penalty, and a window-stability penalty, so the agent learns to balance accuracy, speed, and smooth adaptation. Prioritized experience replay and a target network stabilize training under non-stationarity.

What would settle it

Train RL-Window with the state vector replaced by random noise while keeping everything else identical. If accuracy stays near the reported 90–92% levels rather than collapsing toward baseline, then the hand-designed features are not doing the claimed work. A second check: run the same experiment on a stream where the optimal window size is known analytically (e.g., a piecewise-stationary process with a known change point); the learned policy should track the known optimal window after training.

Watch

Extended reading notes

Core claim

The central claim is that window-size selection in multi-dimensional data streams can be framed and solved as a reinforcement learning problem. Using a Dueling Deep Q-Network with prioritized experience replay, the agent learns a policy that maps stream statistics to one of ten discrete window sizes. On UCI HAR, PAMAP2, and Yahoo! Finance streams, this policy achieves the highest classification accuracy (89.7–92.1%), the smallest post-drift accuracy drop (−3.2 to −4.1%), and the lowest window-size variability among adaptive baselines, while keeping per-instance latency and energy use comparable to simpler methods.

Load-bearing premise

The entire approach rests on the assumption that the hand-designed statistical features computed over the last m points carry enough information for the agent to pick better window sizes than fixed or statistical baselines; without that, the policy has nothing to learn from.

Editorial extensions

If this is right

  • RL-Window's learned policy transfers across three heterogeneous multi-dimensional streams, suggesting one training procedure can serve different sensor and financial applications.
  • Because RL-Window keeps smaller average windows than ADWIN and CNN-Adaptive while losing less accuracy at drift, it offers a better accuracy-responsiveness trade-off.
  • The composite reward allows tuning toward efficiency: raising the cost penalty yields leaner windows, lowering it yields longer context, so deployment can be tailored to resource budgets.
  • The stability penalty reduces window-size churn, which matters for downstream models that are retrained on the current window.
  • The method's competitive latency and energy per instance indicate it can run in edge or real-time settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reward design makes the method task-agnostic: the same RL-Window machinery could be attached to anomaly detection or forecasting instead of classification, and the gains may transfer.
  • A natural next experiment is to test whether the learned policy transfers across datasets without retraining; the paper does not report cross-dataset transfer, but its feature set is deliberately dataset-independent.
  • The reported gains over ADWIN are the strongest on the Yahoo! Finance stream, where synthetic drift was injected; a real-drift benchmark (e.g., from sensors in changing environments) would test whether the drift signals in the state are the source of the robustness.
  • The ablation results imply that prioritized replay, dueling architecture, and spectral features each contribute; ranking their contributions would help practitioners decide where to simplify.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes RL-Window, a reinforcement-learning agent that selects the sliding-window size for multi-dimensional data streams. The agent observes statistical features of recent stream points, chooses a window size from a discrete set, and receives a reward based on downstream classification accuracy, computational cost, and window-size stability. The evaluation compares RL-Window against fixed-size windows, ADWIN, CNN-Adaptive, Stream-X RL, and SMAUG-Inspired on UCI HAR, PAMAP2, and Yahoo! Finance Stream, reporting classification accuracy, drift robustness, average window size, computational cost, stability, energy efficiency, and latency. The central claim is that RL-Window outperforms the baselines on accuracy, drift robustness, and computational efficiency.

Significance. If the claims were reproducible, the paper would address a relevant gap: most adaptive-windowing methods are statistical or supervised, and a general RL formulation for multi-dimensional streams could be useful. The paper also attempts broader evaluation metrics (energy, latency, stability) and provides ablations, which are appropriate for deployment-oriented research. However, as written, the method is not a single, well-defined algorithm: the state and action specifications in Section 3 and Section 4 contradict each other, the reward function is defined inconsistently, and the reported computational-efficiency claims are contradicted by the paper's own tables. No implementation, hyperparameter details, or dataset-generation code is provided. Because the evaluated algorithm cannot be identified from the text, the reported performance gains cannot be attributed to a reproducible method, and the significance of the contribution cannot currently be assessed.

major comments (5)
  1. [§3.1, §3.2, §4.1] The action space is defined inconsistently. Section 3.1 defines W = {20, 40, 60, 80, 100, 120, 140, 160}, which has 8 actions, and Section 3.2 states the DQN output layer has |W| = 8 neurons. Section 4.1 defines W = {20, 40, 60, 80, 100, 120, 140, 160, 180, 200}, which has 10 actions. Table 2, which reports the experimental results, cannot be mapped to a uniquely specified algorithm. The paper must state exactly which action set was used for each reported result.
  2. [§3.1, §4.1] The state-space definition is internally inconsistent and the reported state sizes do not match the stated feature list. Section 3.1 defines features over the last m = 100 points: variance per dimension, pairwise correlations, rate of change, distributional entropy, and out-of-order indicators, with a 10-dimensional state for d = 3. Section 4.1 uses m = 200 and adds spectral features and concept drift signals; the stated formula for HAR (d = 6), 6 + C(6,2) + 6 + 1 + 1 + 6 + 1 = 36, is consistent with that feature list, but for PAMAP2 (d = 12) the same formula gives 12 + C(12,2) + 12 + 1 + 1 + 12 + 1 = 93, not the reported 91, and for Yahoo! Finance (d = 10) it gives 78, not 67. The ablation that removes 'spectral features' confirms that the evaluated model is not the model defined in Section 3. The paper must specify the exact feature set, the length m, and the resulting state dimension for each dataset, and report results only for that specification.
  3. [§3.2, §3.3, §4.1] The reward function and the symbol gamma are used inconsistently. Section 3.1 defines a binary or log-loss reward with a computational-cost penalty lambda * c_t and a discount factor gamma = 0.99. Section 3.3 states that the reward incorporates classification performance, computational cost, and window size stability, but no stability term was defined in Section 3.1. Section 4.1 defines a composite reward with coefficients alpha = 1.0, beta = 0.01, and gamma = 0.005, where gamma is a stability penalty, not the discount factor. Thus the same symbol gamma denotes two different quantities with two different values, and the reward used in the experiments is not the reward defined in the methodology. The paper must define one reward function with distinct symbols and give the exact values used.
  4. [Abstract, §4.4, Tables 2–5] The abstract's claim that RL-Window outperforms baselines in computational efficiency is contradicted by the reported results. In Table 2, RL-Window's computational cost is 2.3 ms on HAR, 2.9 ms on PAMAP2, and 2.7 ms on Yahoo! Finance, which is higher than Fixed-Size Window (1.8, 2.5, 2.2) and ADWIN (2.1, 2.8, 2.5). Tables 3–5 likewise show RL-Window's energy and latency are not the lowest (e.g., energy 1.1 mJ vs. ADWIN 1.0 mJ; latency 2.5 ms vs. ADWIN 2.3 ms). The efficiency claim should be corrected to state that RL-Window is competitive with, but not superior to, ADWIN and fixed windows on these metrics.
  5. [§4.1, §4.3, Table 3] The stability metric is not an independent evaluation metric because a stability penalty is part of the RL reward. Section 4.1 defines the reward as including 'window size stability' with coefficient gamma = 0.005, and Section 4.3 defines Stability as the average absolute change in window size. Reporting that RL-Window has the lowest stability value is therefore partly a consequence of the reward design, not an independent finding. This should be acknowledged, and stability should be presented as a reward-alignment result or evaluated under a reward without the stability penalty.
minor comments (5)
  1. [§3.2, §4.1] The architecture is described inconsistently: Section 3.2 describes a DQN with Double DQN and prioritized experience replay, while the abstract and the ablation study refer to a Dueling DQN, and Section 4.1 also says 'Dueling DQN'. The paper should state the exact network architecture and which components were used in each experiment.
  2. [Table 1] The UCI HAR dataset is attributed to reference [23], but reference [23] is the PAMAP2 paper by Reiss and Stricker. The correct source for UCI HAR should be cited.
  3. [§4.1] The paper says the Yahoo! Finance Stream is 'synthetic' and 'Custom' but gives no description of how the data was generated, what the 10 dimensions represent, or how concept drift was injected. This prevents any independent interpretation of the results on that dataset.
  4. [§2.1] There is a typographical error in the ADWIN discussion: 'However, its reliance on stat(formatting) fails to handle complex inter-dimensional dependencies' contains a broken parenthetical fragment and should be rewritten.
  5. [§3.1] The symbol alpha is used for the learning rate in Section 3.2 (alpha = 0.001) and for the reward accuracy coefficient in Section 4.1 (alpha = 1.0). Distinct symbols should be used to avoid ambiguity.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild constructed-metric circularity in the stability result; central accuracy, drift, and efficiency claims are held-out and remain independent.

  1. fitted input called prediction [§4.1 (Reward), §4.3 (Stability metric), §4.4 (Stability results), Tables 3–5]
    "Reward: Composite reward with α = 1.0, β = 0.01, γ = 0.005, incorporating classification accuracy, computational cost, and window size stability. // Stability: Average absolute change in window size (|wt − wt−1|) during testing, reflecting smoothness of adaptation. // Stability: RL-Window achieves the lowest window size variability (7.8–9.0), compared to 9.5–14.2 for adaptive baselines, due to the reward penalty for large window size changes, ensuring smooth adaptation."

    The stability advantage is the same quantity that the reward function explicitly optimizes: the composite reward contains a 'window size stability' term, and the reported metric is exactly the average absolute window-size change |w_t − w_{t−1}|. The paper itself states that the low variability is 'due to the reward penalty for large window size changes.' Thus the stability result is partly enforced by construction through a fitted reward weight rather than observed as an independent emergent property. This does not collapse the paper's main accuracy or drift comparisons, which are evaluated on held-out test data, but it does mean the stability claim should be read as a restatement of the training objective rather than as a standalone empirical discovery.

full rationale

The paper's central derivation chain is an empirical RL pipeline: a hand-designed state, a discrete window-size action space, a reward, and a Dueling DQN policy, followed by test-set comparisons against ADWIN, CNN-Adaptive, and RL baselines. The headline accuracy, drift-robustness, and computational-cost numbers are held-out measurements and do not reduce to the reward or to any fitted parameter; reporting the reward-aligned accuracy as an evaluation metric is standard RL practice, not circularity. I found no load-bearing self-citations: the references are external works (ADWIN, DQN, prioritized replay, etc.), and no argument in the paper relies on a uniqueness theorem or prior result by the present authors. The one identifiable constructed result is the stability metric, which is explicitly a component of the composite reward and is presented as an outcome; that is a minor, partial circularity. I also note, as a reproducibility and correctness concern rather than circularity, that the state/action specifications in §3.1 and §4.1 conflict (m=100 vs m=200, 8 vs 10 actions, and spectral/drift features appearing only in §4.1 and in the ablation), leaving the exact evaluated method underspecified. This does not make the reported numbers equivalent to their inputs, so it does not raise the circularity score further. Overall, the paper is not derivational circularity: one non-central metric is partly enforced by construction, while the core empirical claims remain independently testable.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters and domain assumptions; none are derived, justified by sensitivity analysis, or made consistent across sections.

free parameters (8)
  • reward accuracy coefficient alpha = 1.0
    Composite reward weights classification accuracy; chosen by hand in Section 4.1 without sensitivity analysis.
  • computational cost penalty beta (also denoted lambda) = 0.01
    Penalty per millisecond of processing time; lambda=0.01 in Section 3.1 and beta=0.01 in Section 4.1, with no tuning study.
  • window stability penalty gamma = 0.005
    Penalty for |w_t - w_{t-1}| in Section 4.1; conflicts with the discount factor gamma=0.99 used in Section 3.1.
  • discount factor gamma = 0.99
    Discount factor in the RL objective in Section 3.1; same symbol is reused for the reward coefficient in Section 4.1.
  • state history length m = 100 (Section 3.1) vs 200 (Section 4.1)
    Number of recent points used to compute state features; contradictory values appear in different sections.
  • action set W = 8 values (20-160) in Section 3.1; 10 values (20-200) in Section 4.1
    Discrete window size choices; inconsistent between the problem formulation and the experiments, affecting the output layer size.
  • exploration epsilon schedule = 1.0 to 0.05 over 50,000 steps
    Epsilon-greedy decay chosen by hand; directly affects the learned policy.
  • replay buffer and batch sizes = 100,000/64 in Section 3.2; 200,000/128 in Section 4.1
    Replay buffer size and mini-batch size differ between the method description and the implementation details.
assumptions (5)
  • standard math Bellman optimality and DQN convergence assumptions
    The agent is trained with a DQN loss in Section 3.2, relying on standard RL convergence assumptions that are not verified for non-stationary streams.
  • domain assumption State features are sufficient statistics for optimal window size selection
    The manually engineered features (variance, correlations, rate of change, entropy, spectral features, drift signals) are assumed to carry enough information for the RL policy; no feature importance or sensitivity analysis is provided.
  • domain assumption Injected synthetic drift is representative of real concept drift
    Drift robustness is measured by altering the mean and variance of selected test dimensions every 10,000 instances; this synthetic process may not reflect natural drift patterns.
  • domain assumption Transformer classifier retrained every 5000 steps is a fair shared downstream evaluator
    All methods use the same classifier and retraining schedule; if the classifier is insensitive to window size, the comparison may distort the reported differences.
  • domain assumption A fixed 1-second reorder buffer resolves out-of-order events
    Section 3.2 assumes a 1-second reorder horizon is sufficient for temporal consistency, with no evidence for the datasets used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Designing Adaptive Algorithms Based on Reinforcement Learning for Dynamic Optimization of Sliding Window Size in Multi-Dimensional Data Streams." pith.science (2026). https://pith.science/paper/PUO4JTCN

@misc{pith2026250706901,
  author       = {Pith},
  title        = {Pith review of: Designing Adaptive Algorithms Based on Reinforcement Learning for Dynamic Optimization of Sliding Window Size in Multi-Dimensional Data Streams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PUO4JTCN}},
  note         = {Machine review of arXiv:2507.06901}
}
read the original abstract

Multi-dimensional data streams, prevalent in applications like IoT, financial markets, and real-time analytics, pose significant challenges due to their high velocity, unbounded nature, and complex inter-dimensional dependencies. Sliding window techniques are critical for processing such streams, but fixed-size windows struggle to adapt to dynamic changes like concept drift or bursty patterns. This paper proposes a novel reinforcement learning (RL)-based approach to dynamically optimize sliding window sizes for multi-dimensional data streams. By formulating window size selection as an RL problem, we enable an agent to learn an adaptive policy based on stream characteristics, such as variance, correlations, and temporal trends. Our method, RL-Window, leverages a Dueling Deep Q-Network (DQN) with prioritized experience replay to handle non-stationarity and high-dimensionality. Evaluations on benchmark datasets (UCI HAR, PAMAP2, Yahoo! Finance Stream) demonstrate that RL-Window outperforms state-of-the-art methods like ADWIN and CNN-Adaptive in classification accuracy, drift robustness, and computational efficiency. Additional qualitative analyses, extended metrics (e.g., energy efficiency, latency), and a comprehensive dataset characterization further highlight its adaptability and stability, making it suitable for real-time applications.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages

  1. [1]

    Learning from Time-Changing Data with Adap- tive Windowing

    Bifet, A., and Gavalda, R. Learning from Time-Changing Data with Adap- tive Windowing . In Proceedings of the 2007 SIAM International Conference on Data Mining , pages 443–448. Soci- ety for Industrial and Applied Mathe- matics, 2007. https://epubs.siam.org/ doi/10.1137/1.9781611972771.42

    doi:10.1137/1.9781611972771.42 2007

  2. [2]

    L., and Carrera, D

    Baig, S., Iqbal, W., Berral, J. L., and Carrera, D. Adaptive Sliding Windows for Improved Estimation of Data Center Resource Utiliza- tion. Future Generation Computer Systems, 102:437–448, 2019. https: //www.sciencedirect.com/science/ article/pii/S0167739X19309203

    work page 2019

  3. [3]

    SMAUG: A Sliding Multidimensional Task Window-Based MARL Framework for Adaptive Real-Time Subtask Recognition

    Zhang, W., and Li, X. SMAUG: A Slid- ing Multidimensional Task Window-Based MARL Framework for Adaptive Real- Time Subtask Recognition. arXiv preprint arXiv:2403.01816, 2024. https://arxiv. org/abs/2403.01816

    work page Pith review arXiv 2024

  4. [4]

    Streaming Deep Reinforcement Learning Finally Works

    Elsayed, M., and others. Streaming Deep Reinforcement Learning Finally Works . arXiv preprint arXiv:2410.14606 , 2024. https://arxiv.org/abs/2410.14606

    arXiv 2024

  5. [5]

    Effects of Sliding Window Variation in the Performance of Acceleration-Based Human Activity Recognition Using Deep Learning Mod- els

    Garcia-Gonzalez, D., Rivero-Juarez, A., and Trivino, G. Effects of Sliding Window Variation in the Performance of Acceleration-Based Human Activity Recognition Using Deep Learning Mod- els. PeerJ Computer Science , 8:e995,

    work page

  6. [6]

    In Proceedings of the 2015 IEEE International Conference on Data Science and Advanced Analytics , pages 1–10, 2015

    Ikonomovska, E., Gama, J., and Se- basti˜ ao, R.Hoeffding Trees with Change Point Detection . In Proceedings of the 2015 IEEE International Conference on Data Science and Advanced Analytics , pages 1–10, 2015. https://ieeexplore. ieee.org/document/7344780 12

    work page arXiv 2015

  7. [7]

    Adaptive Windowing for Online Learning from Multiple Inter- related Data Streams

    Dˇ zeroski, S., Ikonomovska, E., Driessens, K., and Gama, J. Adaptive Windowing for Online Learning from Multiple Inter- related Data Streams. In Proceedings of the 2011 IEEE 11th International Conference on Data Mining Workshops , pages 907– 912, 2011. https://ieeexplore.ieee. org/document/6137448

    work page arXiv 2011

  8. [8]

    Online Machine Learning in Big Data Streams

    Bencz´ ur, A. A., and others. Online Ma- chine Learning in Big Data Streams . arXiv preprint arXiv:1802.05872 , 2018. https://arxiv.org/abs/1802.05872

    work page Pith review arXiv 2018

Show all 30 references
  1. [9]

    Machine Learning for Streaming Data: State of the Art, Challenges, and Opportunities

    Assem, H., and others. Machine Learning for Streaming Data: State of the Art, Challenges, and Opportunities . SIGKDD Explorations Newsletter , 21(2):6–22,

  2. [10]

    Maintaining Stream Statis- tics over Sliding Windows

    Datar, M., Gionis, A., Indyk, P., and Motwani, R. Maintaining Stream Statis- tics over Sliding Windows . SIAM Journal on Computing , 31(6):1794–1813, 2002. https://epubs.siam.org/doi/abs/10. 1137/S0097539701398363

    2002

  3. [11]

    S., and Barto, A

    Sutton, R. S., and Barto, A. G. Reinforcement Learning: An Intro- duction. MIT Press, 2018. https: //mitpress.mit.edu/9780262039246/ reinforcement-learning/

    2018

  4. [12]

    Density-Based Data Streams Clustering over Sliding Windows

    Ren, J., and Ma, X. Density-Based Data Streams Clustering over Sliding Windows. In Proceedings of the 2009 In- ternational Conference on Fuzzy Systems and Knowledge Discovery , pages 248–252,

    2009

  5. [13]

    Human-level control through deep reinforcement learn- ing

    Mnih, V., and others. Human-level control through deep reinforcement learn- ing. Nature, 518(7540):529–533, 2015. https://www.nature.com/articles/ nature14236

    2015

  6. [14]

    Prioritized Experience Replay

    Schaul, T., Quan, J., Antonoglou, I., and Silver, D. Prioritized Experience Replay . arXiv preprint arXiv:1511.05952 , 2015. https://arxiv.org/abs/1511.05952

    2015 arXiv

  7. [15]

    Dueling Network Architectures for Deep Reinforce- ment Learning

    Wang, Z., Schaul, T., Hessel, M., Hasselt, H., Lanctot, M., and Freitas, N. Dueling Network Architectures for Deep Reinforce- ment Learning . In International Confer- ence on Machine Learning , pages 1995– 2003, 2016. https://proceedings.mlr. press/v48/wangf16.html

    1995

  8. [16]

    V., Guez, A., and Silver, D

    Hasselt, H. V., Guez, A., and Silver, D. Deep Reinforcement Learning with Dou- ble Q-Learning . In AAAI Conference on Artificial Intelligence , pages 2094–2100,

  9. [17]

    Noisy Net- works for Exploration

    Fortunato, M., and others. Noisy Net- works for Exploration . arXiv preprint arXiv:1706.10295, 2017. https://arxiv. org/abs/1706.10295

    2017 arXiv

  10. [18]

    B., and Johnson, C

    Yousif, M., Smith, A. B., and Johnson, C. D. Intelligent Processing of Data Streams on the Edge Using Reinforcement Learn- ing. In Proceedings of the IEEE Confer- ence on Edge Computing , 2023

    2023

  11. [19]

    P., and others

    Lillicrap, T. P., and others. Continuous control with deep reinforcement learning . arXiv preprint arXiv:1509.02971 , 2015. https://arxiv.org/abs/1509.02971

    2015 arXiv

  12. [20]

    Batch Nor- malization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

    Ioffe, S., and Szegedy, C. Batch Nor- malization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In International Conference on Machine Learning , pages 448–456, 2015. https://proceedings.mlr.press/v37/ ioffe15.html

    2015

  13. [21]

    Automa- tion of Sequential Feature Scanning in Streaming Environments

    Nagaraju, V., and others. Automa- tion of Sequential Feature Scanning in Streaming Environments . IEEE Transactions on Big Data , 2025 (In Press). https://ieeexplore.ieee.org/ document/pending

    2025

  14. [22]

    A Transformer-based Framework for 13 Multivariate Time Series Representation Learning

    Zerveas, G., Jayaraman, S., Patel, D., Bhamidipaty, A., and Eickhoff, C. A Transformer-based Framework for 13 Multivariate Time Series Representation Learning. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Dis- covery & Data Mining , pages 2114– 2124, 2021. htt...

    2021

  15. [23]

    Intro- ducing a New Benchmarked Dataset for Activity Monitoring

    Reiss, A., and Stricker, D. Intro- ducing a New Benchmarked Dataset for Activity Monitoring . In 2012 16th International Symposium on Wearable Computers , pages 108–109,

    2012

  16. [24]

    Stream-Based Active Learning in the Presence of Verification Latency

    Kottke, D., Lemaire, V., Krempl, G., and Spiliopoulou, M. Stream-Based Active Learning in the Presence of Verification Latency . In Discovery Sci- ence: 23rd International Conference, DS 2020, Thessaloniki, Greece, Octo- ber 19–21, 2020, Proceedings , pages 394–

    2020

  17. [408]

    https://doi

    Springer, Cham, 2021. https://doi. org/10.1007/978-3-030-61527-7_26 14

    2021 doi

  18. [2009]

    net/publication/221086643_ Density-Based_Data_Streams_ Clustering_over_Sliding_Windows

    https://www.researchgate. net/publication/221086643_ Density-Based_Data_Streams_ Clustering_over_Sliding_Windows

  19. [2012]

    https://ieeexplore.ieee.org/ document/6293462

  20. [2016]

    php/AAAI/article/view/10295

    https://ojs.aaai.org/index. php/AAAI/article/view/10295

  21. [2019]

    https://www.researchgate.net/ publication/337581742_Machine_ learning_for_streaming_data_ state_of_the_art_challenges_and_ opportunities

  22. [2022]

    https://www.ncbi.nlm.nih.gov/ pmc/articles/PMC9455026/

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.