Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Label-Free Concept Drift Assessment for Reliable AI in Emerging Wireless Applications

T0 review · 4 major / 6 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 Two unsupervised detectors, CFPT and TabAutoDrift, catch concept drift and signal retraining without ground-truth labels, beating classical drift detectors on two wireless benchmarks.

desk verdict The drift-detection idea is real but the evaluation can't be trusted as printed: the alarm threshold is unspecified and a baseline row is arithmetically wrong. read the letter →

arxiv 2508.00042 v2 pith:J26HMC2F submitted 2025-07-31 cs.NI cs.LG

Athanasios Tziouvaras , Carolina Fortuna , George Floros , Kostas Kolomvatsos , Panagiotis Sarigiannidis , Marko Grobelnik , Blaž Bertalanič This is my paper
classification cs.NIcs.LG
keywords conceptdriftlabel-freedetectionwirelessnetworksexpectedutilitypseudo-labelingmodelretrainingfingerprintinglocalizationlinkanomaly
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

This paper tries to establish that concept drift in deployed wireless machine-learning models can be detected without any post-deployment ground-truth labels, using an expected-utility score computed from a surrogate model's macro-F1 curve. If correct, this gives network operators a way to decide when to retrain a model purely from unlabeled operational data, which is exactly the situation in 6G-era wireless systems where true labels are scarce. The authors propose two model-agnostic detectors, CFPT and TabAutoDrift, and report drift-detection F1 scores of 0.88–0.94 on outdoor fingerprinting localization and 0.80–1.00 on link-anomaly detection, up to 0.24 higher than classical detectors.

What carries the argument

The central machinery is the expected-utility score $U$, the mean absolute epoch-wise difference between two macro-F1 curves, defined in Eq. 4 for CFPT and Eq. 5 for TabAutoDrift. CFPT generates pseudo-labels by keeping only the deployed model's high-confidence predictions on the unlabeled new batch, then re-trains an XGBoost classifier on those pseudo-labels and compares its macro-F1 curve to the curve from training on the original data. TabAutoDrift uses an unsupervised TabNet pass on the new batch to learn its feature structure, transfers those weights into a supervised classifier, and compares that classifier's macro-F1 curve to a classifier that never saw the new data. A large $U$ is the signal that the original model should be retrained.

What would settle it

A direct test would be a constructed drift where $M_0$ is a very confident classifier and all post-drift examples are of a new class that lies inside the convex hull of an old class; if CFPT's pseudo-labeling assigns them high confidence to that old class and $U$ stays near zero while true accuracy drops, the central claim fails. A second test would run TabAutoDrift on a no-drift stream with shuffled feature order; a large $U$ would indicate the score is not specific to drift.

Watch

Extended reading notes

Core claim

The central claim is that concept drift can be assessed without labels by measuring how much a surrogate model's learning curve changes when it is re-trained on new data. CFPT takes the deployed model $M_0$, uses its high-confidence predictions on the unlabeled new batch $D_1$ as pseudo-labels, trains an XGBoost classifier on $D_0$ for a few epochs, re-trains it on $(D_1, L_1)$, and computes the expected utility $U = \frac{1}{n}\sum_i |\bar{F}_1^{(i)}_{(D_1,L_1)} - \bar{F}_1^{(i)}_{(D_0,L_0)}|$. TabAutoDrift instead pre-trains TabNet unsupervised on $D_1$, transfers those weights into two supervised TabNet classifiers, and uses the analogous difference $\frac{1}{n}\sum_i |\bar{F}_1^{(i)}_{(D_0,L_0)} - \bar{F}_1^{(i)}_{(D_0,L_0,D_1)}|$ as a utility signal. The paper argues that a utility value near 1 means the new data and the original data cannot be fit by one model, so retraining $M_0$ is warranted.

Load-bearing premise

The method assumes that when the true data distribution changes, a surrogate model's macro-F1 curve will also change; if the deployed model confidently maps new-class data onto an old class, the utility score stays low and drift goes unnoticed.

Editorial extensions

If this is right

  • Deployed models in AI-native wireless networks can be monitored by an external, model-agnostic observer using only unlabeled operational data.
  • Retraining alarms can be suppressed while the model is still reliable, avoiding the cost of unnecessary model updates.
  • One detector configuration transfers across tabular fingerprinting data and time-series link data without per-domain recalibration.
  • The reported gains over classical detectors come with a runtime cost: CFPT and TabAutoDrift take about 9–19 seconds per assessment versus under 6 seconds for baselines, which the paper argues is affordable for gradual drifts.

Reading between the lines

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

  • Editorial inference: Because the paper does not state the alarm threshold on $U$, reproducing the reported F1 numbers requires choosing one; a natural calibration is to threshold $U$ against its distribution on a no-drift validation set.
  • Editorial inference: The same expected-utility idea could be applied to regression or ranking tasks by replacing macro-F1 with a task-appropriate per-epoch performance score, since the mechanism only needs a learning curve.
  • Editorial inference: A stress test outside the paper's reported regimes would be a drift where new data belongs to a class the deployed model already maps confidently onto an old class; if $U$ stays near zero while true accuracy collapses, the label-free signal would miss the drift.
  • Editorial inference: The utility score likely reacts to covariate shift as well as to the source-3 concept drift the paper targets, which could make the detector useful more broadly but also means operators may need to distinguish which type of drift triggered the alarm.
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

4 major / 6 minor

Summary. The paper proposes two label-free, batch concept-drift detectors for wireless applications: CFPT, which pseudo-labels a new unlabeled batch using the deployed model and measures the absolute macro-F1 difference of an XGB classifier before and after retraining, and TabAutoDrift, which uses TabNet representation learning and a similar expected-utility score. The utility U (Eqs. 4 and 5) is intended to trigger a retraining alarm. The authors evaluate the methods on an outdoor BLE fingerprinting dataset (31 batches) and a wireless link-anomaly dataset (9 batches), reporting drift-detection F1 scores up to 0.94 and 1.00, respectively, and higher reward than ADWIN, DDM, CUSUM, PageHinkley, and STEPD. The paper also includes an ablation of training/retraining epochs and a runtime comparison.

Significance. The problem is relevant: monitoring deployed ML models without labels is important for AI-native 6G networks and other label-scarce operational environments. The paper uses two real-world wireless datasets and proposes model-agnostic, label-free detectors, which are potentially useful contributions. It also provides an ablation of training epochs and runtime measurements. However, the current manuscript lacks the decision-threshold specification and the stream-to-batch conversion protocol needed to substantiate the headline F1 comparison. If these gaps are filled and the numbers remain, the methods would be a solid empirical contribution.

major comments (4)
  1. [Sec. VI.C, Eqs. (4)-(5), Tables I-II] The alarm decision rule is underspecified. The paper defines only the utility U and states that "each method uses a different prediction confidence threshold that is uniquely tailored for itself," but no threshold value or selection procedure is reported. Since Tables I and II are built from binary alarms, the reported TP/FP/TN/FN counts and F1 scores cannot be reproduced, and one cannot rule out that thresholds were chosen on the evaluation batches. Please report the exact decision rule (e.g., st = 1 iff U > θ), the θ values for CFPT and TabAutoDrift, and how each θ was selected (e.g., on a validation split before seeing test batches).
  2. [Sec. VI.D, Tables I-II] The protocol for converting online stream detectors (ADWIN, DDM, CUSUM, PageHinkley, STEPD) into per-batch binary decisions is missing. These detectors emit warnings over time, whereas the ground truth here is batch-level. Without a rule such as "warn if at least one alarm occurs inside the batch" or a delayed-decision rule, neither the alarm counts nor the F1 comparison with the proposed batch methods is well-defined. Please specify the conversion and ensure it is applied symmetrically to all methods.
  3. [Table I] The ADWIN row is arithmetically inconsistent. With TP=6/10 and FP=0/20, precision must be 6/(6+0)=1.00, recall 0.60, and F1=0.75, not the printed 0.60/0.60/0.60. As printed, the table understates this baseline and thus inflates the reported margin of the proposed methods. Please correct the table and re-check the other rows for consistency.
  4. [Sec. V.A] The rationale for CFPT is stated in reverse. The text says that if ¯F1(D1,L1)→1 then the retraining objective "cannot easily converge," but a macro-F1 close to 1 indicates that the XGB model fits the pseudo-labeled data well, i.e., the objective is minimized. The subsequent bullets also switch between low F1 and high F1. This matters because the method's failure mode is exactly when the deployed model M0 confidently maps new-class data to an old class, producing confident but wrong pseudo-labels that leave U low and the drift undetected. Please correct the explanation and add an experiment or discussion addressing this failure mode.
minor comments (6)
  1. [Section III] The set notation "Di = {d0, d2, d3, ..., dj}" appears to be missing d1; please fix the sequence.
  2. [Eqs. (2)-(3)] The summation indices I and J in Eq. (2) are undefined; please replace them with explicit bounds or define them in the text.
  3. [Abstract and contributions] The contribution bullet says the methods achieve an overall F1 of 0.9 in fingerprinting and 0.8 in links, while Table II reports CFPT achieving 1.00 on links; please reconcile the numbers across the abstract, contributions, and results.
  4. [Figures 3 and 4] The figures are difficult to read in the provided manuscript; please ensure all labels, especially the arrows and text boxes, are legible and consistent with the method descriptions.
  5. [Section VI.C] The relationship between the reward scores in Figures 5-6 and Eq. (6) is not fully defined; please specify how F1_gain is computed per batch and how the aggregate reward over the eight runs is derived.
  6. [Throughout] There are typos and inconsistent names, e.g., "PageHinkey" vs. "Page-Hinkley", "ustilises" in the Fig. 3 caption, and "origila" in the same caption; please proofread the manuscript.

Circularity Check

1 steps flagged · score 4.0 of 10

The central drift-detection F1 claims are not reproducible as stated because the alarm threshold converting utility U to binary retraining decisions is never specified; the reported confusion-matrix numbers therefore depend on an unstated, and potentially test-fitted, decision rule. No definitional equivalence or load-bearing self-citation is present in the detector derivation itself.

  1. fitted input called prediction [Section VI.C (Performance metrics) and Tables I-II; utility defined in Eq. (4)-(5).]
    "Since the methods proposed in this work use a utility function that designates whether the M0 model should be retrained or not, it is difficult to compare them with existing solutions. This happens because each method uses a different prediction confidence threshold that is uniquely tailored for itself."

    The alarm decision is implicitly st = 1 iff U > theta, but theta is never reported anywhere in the paper. The paper then records whether each method 'raises an alarm' and computes precision, recall, and F1 from those alarms. Because theta is a free parameter described as 'uniquely tailored' to each method, the reported F1 values (0.88-0.94 and 0.80-1.00) are not determined by the method as specified; they can be reproduced only by choosing theta on the evaluated batches. If theta was selected using the test batches, the headline 'prediction' is a fit to the evaluation ground truth rather than a label-free prediction. This is a partial circularity in the evaluation, although not a definitional identity in the detector itself.

full rationale

The proposed detectors are not circular by construction: U in Eq. (4)-(5) is a heuristic dissimilarity between macro-F1 curves during training and retraining, and the ground-truth drift labels come from the external dataset construction (new classes introduced in D_i), not from U. CFPT's use of M0 pseudo-labels is a self-referential modeling choice, but it is a stated assumption rather than an equivalence to the target; TabAutoDrift does not use M0 at all. No load-bearing self-citations, uniqueness theorems, or ansatz-smuggling citations appear; references [7] and [8] are external datasets. The principal deficiency is evaluation-level: the threshold theta that converts U into binary retraining alarms is never specified, despite the paper's admission that thresholds are 'uniquely tailored' per method. This makes the headline F1 margin unreproducible and potentially test-fitted, which is a fitted-input-called-prediction concern. Additionally, Table I's ADWIN row is internally inconsistent (TP=6, FP=0 forces precision=1.00 and F1=0.75, not the printed 0.60/0.60), further undercutting the reported comparison. These issues warrant a moderate circularity score, but they do not establish that the detector derivation itself is definitionally circular.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on several unstated or hand-tuned parameters, including the pseudo-label confidence threshold, the utility alarm threshold, and epoch counts. The methods also rely on the heuristic assumption that macro-F1 differences reflect drift, which is not derived from theory.

free parameters (5)
  • CFPT pseudo-label confidence threshold = not specified
    The method keeps the 'highest confidence scores' from M0's forward pass but does not state the fraction or threshold used to select pseudo-labels; this choice affects the quality of L1 and hence the utility U.
  • CFPT training and retraining epochs = 5
    Chosen by hand from an ablation on two drift datasets; the ablation shows sensitivity, so this is a tuned hyperparameter.
  • TabAutoDrift training epochs = 5
    Set to 5 based on the same ablation logic; the architecture is inherited from TabNet defaults.
  • Utility alarm threshold = not reported
    A threshold on U must convert the continuous utility to a binary retrain alarm, but the paper does not state how it is set; if set using test data, the F1 results reflect fitting.
  • Ts reward constant = 0.1
    User-set constant in the evaluation reward function; not part of the detectors but affects reported reward comparisons.
assumptions (3)
  • domain assumption Macro F1 on training data is a meaningful measure of convergence and learnability
    Both methods interpret epochwise macro-F1 differences as evidence of distribution mismatch; no theoretical justification is given (Section V-A, Eq. 4-5).
  • ad hoc to paper High-confidence pseudo-labels from M0 are informative about whether D1 comes from a new distribution
    CFPT's logic assumes that when drift occurs, M0 will produce pseudo-labels that are hard to fit, which is not guaranteed (Section V-A).
  • domain assumption The synthetic drift construction used in the fingerprinting dataset resembles real wireless concept drift
    The evaluation uses dataset splits where every third dataset introduces a new class, which may not reflect the gradual or subtle drifts of real deployments (Section VI-A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Label-Free Concept Drift Assessment for Reliable AI in Emerging Wireless Applications." pith.science (2026). https://pith.science/paper/J26HMC2F

@misc{pith2026250800042,
  author       = {Pith},
  title        = {Pith review of: Label-Free Concept Drift Assessment for Reliable AI in Emerging Wireless Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J26HMC2F}},
  note         = {Machine review of arXiv:2508.00042}
}
read the original abstract

Machine learning models deployed in non-stationary environments degrade silently, since as the input distribution drifts their accuracy decays without an error signal and without labels to reveal it. Sustaining reliable AI therefore requires a concept-drift detector that acts as an external observer of the deployed model, monitoring it using unlabeled operational data alone, so that an MLOps actuator triggers retraining and redeployment only when it is warranted. This paper contributes two concept drift detectors, namely Confidence-Filtered Pseudo-Label Transfer (CFPT) and TabAutoDrift, which combine representation learning with statistical testing to compute an expected utility score that signals whether a deployed model should be retrained, without requiring ground-truth labels after deployment. The detectors are evaluated on two emerging, label-scarce wireless application domains in which post deployment ground truth is effectively unavailable, namely outdoor fingerprinting-based localization and link-anomaly detection. They outperform the classical detectors ADWIN, DDM, and CUSUM, attaining a drift-detection F1-score between 0.88 and 0.94 in the fingerprinting use case and between 0.80 and 1.00 in the link-anomaly use case, up to 0.24 higher than the strongest classical detector. Interpreted as reliability decisions, this precision indicates that the proposed detectors signal retraining more dependably.

Figures

Figures reproduced from arXiv: 2508.00042 by the authors.

Figure 1
Figure 1. An example of concept drift between two different [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An example of concept drift between normal and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The proposed CFPT methods that leverages a semi [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The proposed TabAutoDrift method where the TabNet [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The performance of different concept drift detection [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The performance of different concept drift detection [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: The effect of different number of training epochs to [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 7
Figure 7. Figure 7: The effect of different number of training and re [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [1]

    The role of ai enablers in overcoming impairments in 6g networks,

    M. Saimler, S. ˙Ickin, G. Bernini, N. Toumi, M. Diamanti, S. Papavas- siliou, M. Zivkovic, O. U. Akgul, and B. M. Khorsandi, “The role of ai enablers in overcoming impairments in 6g networks,” in 2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit), 2024, pp. 913–918

    work page 2024

  2. [2]

    Machine learning and wi-fi: Unveiling the path toward ai/ml-native ieee 802.11 networks,

    F. Wilhelmi, S. Szott, K. Kosek-Szott, and B. Bellalta, “Machine learning and wi-fi: Unveiling the path toward ai/ml-native ieee 802.11 networks,” IEEE Communications Magazine , pp. 1–7, 2024

    work page 2024

  3. [3]

    Operationalizing ai/ml in future net- works: A bird’s eye view from the system perspective,

    Q. Liu, T. Zhang, M. Hemmatpour, H. Qiu, D. Zhang, C. S. Chen, M. Mellia, and A. Aghasaryan, “Operationalizing ai/ml in future net- works: A bird’s eye view from the system perspective,” IEEE Commu- nications Magazine, vol. 63, no. 4, pp. 176–182, 2025

    work page 2025

  4. [4]

    Misconfig- uration in o-ran: Analysis of the impact of ai/ml,

    N. M. Yungaicela-Naula, V . Sharma, and S. Scott-Hayward, “Misconfig- uration in o-ran: Analysis of the impact of ai/ml,” Computer Networks, p. 110455, 2024

    work page 2024

  5. [5]

    Learning under concept drift: A review,

    J. Lu, A. Liu, F. Dong, F. Gu, J. Gama, and G. Zhang, “Learning under concept drift: A review,” IEEE Transactions on Knowledge and Data Engineering, vol. 31, no. 12, pp. 2346–2363, 2019

    2019

  6. [6]

    Leaf: Navigating concept drift in cellular networks,

    S. Liu, F. Bronzino, P. Schmitt, A. N. Bhagoji, N. Feamster, H. G. Crespo, T. Coyle, and B. Ward, “Leaf: Navigating concept drift in cellular networks,” Proceedings of the ACM on Networking , vol. 1, no. CoNEXT2, pp. 1–24, 2023

    work page 2023

  7. [7]

    Log-a-tec testbed outdoor localization using ble beacons,

    B. Bertalani ˇc, G. Morano, and G. Cerar, “Log-a-tec testbed outdoor localization using ble beacons,” in2022 International Balkan Conference on Communications and Networking (BalkanCom) , 2022, pp. 115–119

    work page 2022

  8. [8]

    Resource-aware time series imaging classification for wireless link layer anomalies,

    B. Bertalani ˇc, M. Me ˇza, and C. Fortuna, “Resource-aware time series imaging classification for wireless link layer anomalies,” IEEE Trans- actions on Neural Networks and Learning Systems , vol. 34, no. 10, pp. 8031–8043, 2023

    work page 2023

Show all 31 references
  1. [9]

    A unifying view on dataset shift in classification,

    J. G. Moreno-Torres, T. Raeder, R. Alaiz-Rodr ´ıguez, N. V . Chawla, and F. Herrera, “A unifying view on dataset shift in classification,” Pattern recognition, vol. 45, no. 1, pp. 521–530, 2012

    2012

  2. [10]

    Learning under concept drift: A review,

    J. Lu, A. Liu, F. Dong, F. Gu, J. Gama, and G. Zhang, “Learning under concept drift: A review,” IEEE transactions on knowledge and data engineering, vol. 31, no. 12, pp. 2346–2363, 2018

    2018

  3. [11]

    Learning with drift detection,

    J. Gama, P. Medas, G. Castillo, and P. Rodrigues, “Learning with drift detection,” in Advances in Artificial Intelligence – SBIA 2004 , A. L. C. Bazzan and S. Labidi, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2004, pp. 286–295

    2004

  4. [12]

    Continuous inspection schemes,

    E. S. Page, “Continuous inspection schemes,” Biometrika, vol. 41, no. 1/2, pp. 100–115, 1954. [Online]. Available: http://www.jstor.org/ stable/2333009

    1954

  5. [13]

    Bifet and R

    A. Bifet and R. Gavald `a, Learning from Time-Changing Data with Adaptive Windowing , pp. 443–448. [Online]. Available: https: //epubs.siam.org/doi/abs/10.1137/1.9781611972771.42

    doi

  6. [14]

    Detecting concept drift using statistical testing,

    K. Nishida and K. Yamauchi, “Detecting concept drift using statistical testing,” in Discovery Science, V . Corruble, M. Takeda, and E. Suzuki, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2007, pp. 264– 269

    2007

  7. [15]

    Concept drift detection using autoencoders in data streams processing,

    M. Jaworski, L. Rutkowski, and P. Angelov, “Concept drift detection using autoencoders in data streams processing,” in International Con- ference on Artificial Intelligence and Soft Computing . Springer, 2020, pp. 124–133

    2020

  8. [16]

    Recent advances in concept drift adaptation methods for deep learning

    L. Yuan, H. Li, B. Xia, C. Gao, M. Liu, W. Yuan, and X. You, “Recent advances in concept drift adaptation methods for deep learning.” in IJCAI, 2022, pp. 5654–5661

    2022

  9. [17]

    Evolving cybersecurity frontiers: A comprehensive survey on concept drift and feature dynamics aware machine and deep learning in intrusion detection systems,

    M. A. Shyaa, N. F. Ibrahim, Z. Zainol, R. Abdullah, M. Anbar, and L. Alzubaidi, “Evolving cybersecurity frontiers: A comprehensive survey on concept drift and feature dynamics aware machine and deep learning in intrusion detection systems,” Engineering Applications of Artifici...

    2024

  10. [18]

    A survey on machine learning for recurring concept drifting data streams,

    A. L. Su ´arez-Cetrulo, D. Quintana, and A. Cervantes, “A survey on machine learning for recurring concept drifting data streams,” Expert Systems with Applications , vol. 213, p. 118934, 2023

    2023

  11. [19]

    Insomnia: Towards concept-drift robustness in network intrusion detection,

    G. Andresini, F. Pendlebury, F. Pierazzi, C. Loglisci, A. Appice, and L. Cavallaro, “Insomnia: Towards concept-drift robustness in network intrusion detection,” in Proceedings of the 14th ACM workshop on artificial intelligence and security , 2021, pp. 111–122

    2021

  12. [20]

    Class imbalance and concept drift invariant online botnet threat detection framework for heterogeneous iot edge,

    A. Nitish, J. Hanumanthappa, S. P. SP et al. , “Class imbalance and concept drift invariant online botnet threat detection framework for heterogeneous iot edge,” Computers & Security , vol. 141, p. 103820, 2024

    2024

  13. [21]

    Intrusion detection in the iot data streams using concept drift localization,

    R. Chu, P. Jin, H. Qiao, and Q. Feng, “Intrusion detection in the iot data streams using concept drift localization,” AIMS mathematics, vol. 9, no. 1, pp. 1535–1561, 2024

    2024

  14. [22]

    A multi-agent adaptive deep learning framework for online intrusion detection,

    M. Soltani, K. Khajavi, M. Jafari Siavoshani, and A. H. Jahangir, “A multi-agent adaptive deep learning framework for online intrusion detection,” Cybersecurity, vol. 7, no. 1, p. 9, 2024

    2024

  15. [23]

    23when training and test sets are different: Characterizing learning transfer,

    S. Amos, “23when training and test sets are different: Characterizing learning transfer,” in Dataset Shift in Machine Learning . The MIT Press, 12 2008. [Online]. Available: https://doi.org/10.7551/mitpress/ 9780262170055.003.0001

    2008 doi

  16. [24]

    On the impact of industrial delays when mitigating distribution drifts: an empirical study on real-world financial systems,

    T. Simonetto, M. Cordy, S. Ghamizi, Y . Le Traon, C. Lefebvre, A. Boys- tov, and A. Goujon, “On the impact of industrial delays when mitigating distribution drifts: an empirical study on real-world financial systems,” KDD Workshop, 2024

    2024

  17. [25]

    Detection of data drift and outliers affecting machine learning model performance over time,

    S. Ackerman, E. Farchi, O. Raz, M. Zalmanovici, and P. Dube, “Detection of data drift and outliers affecting machine learning model performance over time,” 2022. [Online]. Available: https: //arxiv.org/abs/2012.09258

    2022 arXiv

  18. [26]

    Why the pseudo label based semi- supervised learning algorithm is effective?

    Z. Min, Q. Ge, and C. Tai, “Why the pseudo label based semi- supervised learning algorithm is effective?” 2023. [Online]. Available: https://arxiv.org/abs/2211.10039

    2023 arXiv

  19. [27]

    Pseudo-labeling and confirmation bias in deep semi-supervised learning,

    E. Arazo, D. Ortego, P. Albert, N. E. O’Connor, and K. McGuin- ness, “Pseudo-labeling and confirmation bias in deep semi-supervised learning,” in 2020 International Joint Conference on Neural Networks (IJCNN), 2020, pp. 1–8

    2020

  20. [28]

    In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning,

    M. N. Rizve, K. Duarte, Y . S. Rawat, and M. Shah, “In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning,” 2021. [Online]. Available: https://arxiv.org/abs/2101.06329

    2021 arXiv

  21. [29]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , ser. KDD ’16. New York, NY , USA: Association for Computing Machinery, 2016, p. 785–794. [Online]. Avail...

    2016

  22. [30]

    Tabnet: Attentive interpretable tabular learning,

    S. O. Arik and T. Pfister, “Tabnet: Attentive interpretable tabular learning,” 2020. [Online]. Available: https://arxiv.org/abs/1908.07442

    2020 arXiv

  23. [31]

    Analysis of descriptors of concept drift and their impacts,

    A. Costa, R. Giusti, and E. M. dos Santos, “Analysis of descriptors of concept drift and their impacts,” in Informatics, vol. 12, no. 1. MDPI, 2025, p. 13

    2025

Pith tools

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