Domain-Adversarial Transfer Learning for Fault Root Cause Identification in Cloud Computing Systems

Bruce Fang; Danyi Gao

REVIEW 5 major objections 5 minor 1 cited by

Domain-Adversarial Transfer Learning for Fault Root Cause Identification in Cloud Computing Systems

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 Transfer learning model identifies cloud fault root causes at 89.6% accuracy

desk verdict Routine combination of MMD, adversarial domain adaptation, and pseudo-labeling whose reported results rest on an unsupported label assumption for Alibaba Cluster Trace 2018; not ready for peer review. read the letter →

arxiv 2507.02233 v1 pith:2AMVDBHU submitted 2025-07-03 cs.DC

Bruce Fang , Danyi Gao This is my paper · ORCID

classification cs.DC

keywords domain-adversarialtransferlearningfaultrootcauseidentificationcloudcomputingpseudo-labelselectionMaximumMeanDiscrepancyclassimbalanceAlibabaClusterTrace2018domainadaptation

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 a transfer-learning method for identifying the root cause of faults in cloud computing systems when labeled fault data are scarce. It combines a shared feature extractor, an adversarial domain discriminator, a Maximum Mean Discrepancy term, and a pseudo-label selection step that feeds high-confidence target predictions back into training. On the Alibaba Cluster Trace 2018, the authors report that the method outperforms four domain-adaptation baselines, DANN, CDAN, FixBi, and ToAlign, reaching 89.6% accuracy, 85.4% F1-score, and 91.3% AUC. The claim matters because cloud operations teams often have rich data from one group of nodes but only partial or no labels from another, and the paper argues that its combination of mechanisms transfers diagnostic knowledge across such heterogeneous node groups.

What carries the argument

The machinery is a domain-adversarial transfer architecture: a shared feature extractor $F(x;\theta_f)$ produces representations from both source and target inputs; a classifier $C$ minimizes cross-entropy on labeled source data; a domain discriminator $D$ is trained adversarially so that $F$'s features become domain-indistinguishable, while an MMD term $L_{mmd}$ explicitly pulls the source and target feature distributions together; and a pseudo-label selector adds target samples whose predicted class probability exceeds $\delta$ to the training set. The total loss is $L_{total}=L_s+\lambda_1 L_{mmd}+\lambda_2 L_{adv}$. The pseudo-label threshold is what lets the model exploit unlabeled target data, and the adversarial term is what the paper credits for robustness on heterogeneous nodes.

What would settle it

A concrete test would be to run the same architecture on a cloud fault dataset with publicly available root-cause labels, or to publish the label construction for the Alibaba trace, and check whether the reported 89.6% accuracy, 85.4% F1, and 91.3% AUC can be reproduced under the same 10%-label, imbalanced, and heterogeneous-node protocols.

Watch

Extended reading notes

Core claim

The central claim is that a joint training objective—combining source-domain classification loss, an MMD distribution-matching term, and an adversarial loss between a feature extractor and a domain discriminator—plus a confidence-threshold pseudo-label mechanism produces a fault root-cause classifier that generalizes across nodes and time periods. In the paper's experiments, this is the only method that reaches 89.6% accuracy, 85.4% F1, and 91.3% AUC on the Alibaba data, and it maintains an accuracy of 0.81 and an AUC of 0.86 with only 10% labeled target data. Under worsening class imbalance the model degrades, as the authors acknowledge, but it remains the strongest of the compared methods, and on heterogeneous node types (CPU-, memory-, I/O-, and mixed-load nodes) the adversarial mechanism keeps performance high.

Load-bearing premise

The load-bearing premise is that the Alibaba Cluster Trace 2018 source domain actually contains a large number of labeled fault events with clearly defined fault types; if those labels do not exist in the data, the supervised loss, the pseudo-label selection, and the reported accuracy, F1, and AUC have no reproducible basis.

Editorial extensions

If this is right

If the reported metrics are correct, operations teams could train a fault root-cause classifier on one labeled group of nodes and apply it to another group with only 10% labeled data, retaining roughly 0.81 accuracy and 0.86 AUC.
The paper's joint objective indicates that combining distribution matching (MMD) with adversarial domain discrimination and pseudo-labels yields better cross-node generalization than any of the four baseline domain-adaptation methods it compares against.
Because the model degrades as fault-class imbalance grows, the paper implies that transfer alignment alone is not enough for rare-fault detection; reweighting or better pseudo-label selection would be needed in deployment.
On heterogeneous node types (CPU-, memory-, I/O-, and mixed-load), the domain-adversarial mechanism is reported to keep all metrics high, suggesting the approach is usable across resource pools with different workload profiles.

Reading between the lines

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

Editorial inference: the paper never ablates MMD, adversarial training, and pseudo-labeling separately, so which component carries the gain is untested; a component-removal study would settle it.
Editorial inference: because the public Alibaba Cluster Trace 2018 is a workload and scheduling trace, reproducing the paper's numbers requires either a label construction step the paper does not describe or access to a differently-annotated version of the data.
Editorial inference: the pseudo-label confidence threshold $\delta$ is a likely sensitivity point; testing how the reported numbers move as $\delta$ varies would give a practical operating curve for deployment.

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.

Desk Editor's Note

read the letter

Quick take: don't spend referee time on this. The method is a standard stack — shared encoder, MMD (Eq. 2), adversarial domain discriminator (Eq. 3), and confidence-threshold pseudo-labels. The application to cloud root-cause identification is legitimate and the paper frames a real problem, but that is the extent of the novelty.

What it does well: the problem is important, and the authors pick the right baselines to compare against — DANN, CDAN, FixBi, and ToAlign. The method section describes the three loss components and the joint optimization cleanly enough. Credit also for the Figure 3 discussion, where they admit that transfer mechanisms alone degrade under class imbalance. That is more honest than many papers in this space.

The soft spots are not minor. There is no evaluation protocol: no data split, no network architecture, no hyperparameters, no optimizer, no preprocessing, and no variance. Table 1 reports single numbers with no error bars. That alone would make the empirical claim weak.

The load-bearing problem is the dataset. Section IV.A says the source domain in Alibaba Cluster Trace 2018 contains a large number of labeled fault events and clearly defined fault types. The public trace is a workload and scheduling trace; it does not document fault root-cause categories. The paper cites no label source and gives no construction procedure. If the labels do not exist, Table 1 and all three figures cannot have been produced as described. This is not a reproducibility gap; it is an unsupported foundation. The stress-test note is correct.

The novelty assessment holds too: each component is established, and the paper itself cites DANN, FixBi, and ToAlign as antecedents. The only new thing is the combination and the application domain. That might support a short paper if the experiments were fully specified and the labels were real, but not as it stands.

For peer review: desk reject. There is nothing a referee can check without code, labels, and a protocol. If the authors rebuild the evaluation on a dataset with documented fault labels and submit with code and full hyperparameters, the combination could be worth a second look.

Referee Report

5 major / 5 minor

Summary. The paper proposes a transfer-learning method for cloud fault root cause identification, combining a shared feature extractor, MMD-based distribution alignment, a domain-adversarial discriminator, and a confidence-threshold pseudo-label selection mechanism. The evaluation is claimed on the Alibaba Cluster Trace 2018 dataset under label scarcity, class imbalance, and heterogeneous node conditions. The central claim is that the proposed method reaches 89.6% accuracy, 85.4% F1-score, and 91.3% AUC, outperforming DANN, CDAN, FixBi, and ToAlign.

Significance. If the empirical claim were supported, the work would address a practically relevant problem: fault root cause identification with limited labeled data in heterogeneous cloud environments. The method itself is a fairly standard combination of existing transfer-learning components (MMD, adversarial domain adaptation, pseudo-labeling), so the novelty is incremental. The paper provides no code, no architecture details, no hyperparameter settings, and no statistical variance information. More fundamentally, the dataset description asserts the existence of labeled fault root-cause events in the Alibaba Cluster Trace 2018 that the public trace does not document, which places the entire empirical evaluation on an unsupported foundation. The claimed contribution therefore cannot currently be verified, and the manuscript falls short of the reproducibility and evidence standards expected for a systems or applied-ML publication.

major comments (5)

[IV.A] The evaluation's load-bearing premise is that the Alibaba Cluster Trace 2018 contains labeled fault root-cause events. Section IV.A states that the source domain 'contains a large number of labeled fault events and clearly defined fault types,' but the public Alibaba Cluster Trace 2018 is a workload and scheduling trace whose documented schema contains resource usage, task, and instance data, not labeled root-cause categories. The paper neither cites a label source, nor describes how fault labels were constructed, nor states how many fault classes exist, nor specifies how source and target node groups were selected. Without such labels, Equations (1) and (3) cannot be optimized, the pseudo-label mechanism cannot be validated, and every number in Table 1 is unsupported. This is not a minor clarity omission; it prevents any independent check of the central claim.
[III, Eqs. (2)-(4)] The method section is not reproducible as written. Equation (2) and Equation (3) contain malformed notation: the sums and indices are garbled, the MMD kernel is not defined, and the domain-discriminator loss in Equation (3) does not clearly specify the target-domain label convention. Equation (4) defines the total loss as L_total = L_s + lambda_1 * L_mmd + lambda_2 * L_adv, but the text says 'A and B are the weight coefficients' without giving values or a selection procedure. No network architecture, feature dimensionality, optimizer, learning rate, batch size, or number of training epochs is stated. These omissions make the method impossible to reimplement from the manuscript.
[IV.B, Table 1] Table 1 reports a single accuracy, F1-score, and AUC value per method, with no error bars, no repeated runs, no standard deviation, and no statistical significance test. The claim that the proposed method 'consistently outperforms' DANN, CDAN, FixBi, and ToAlign across all metrics is therefore not established. Moreover, the baseline implementations are not described, so it is unclear whether the baselines were tuned fairly or even run on the same fault-label construction. Given the missing label-construction details, the reported margins (e.g., 2.5 points in accuracy over ToAlign) cannot be interpreted as meaningful.
[IV, Figures 2-4] The robustness analyses in Figures 2, 3, and 4 are described only in prose; the manuscript provides no axis definitions, no quantitative data tables, no confidence intervals, and no number of trials. For example, the claim that 'with only 10% labeled data, the model achieves an accuracy of 0.81 and an AUC of 0.86' is reported without explaining what the 10% refers to, how many samples are in the labeled subset, or how the subset was sampled. These figures therefore do not provide a verifiable basis for the conclusions about label scarcity, class imbalance, or heterogeneous-node adaptability.
[III, pseudo-label mechanism] The pseudo-label selection uses the model's own high-confidence predictions on target data as training labels, which introduces a self-training circularity. The manuscript does not quantify the accuracy of the selected pseudo-labels, does not describe how the confidence threshold delta was chosen, and does not analyze the risk of confirmation bias, particularly for minority fault classes under class imbalance. Since the paper explicitly claims improved minority-class recognition, the absence of any pseudo-label quality analysis is a load-bearing gap in the evaluation.

minor comments (5)

[Abstract and Keywords] The keyword 'domain confrontation' is an unnatural translation; it should be 'domain adversarial' or 'domain-adversarial learning'.
[III, Eq. (4)] The text says 'A and B are the weight coefficients,' but the equation uses lambda_1 and lambda_2; the notation should be made consistent.
[III and IV.A] Section III defines the target domain as unlabeled data, while Section IV.A says the target domain contains 'partially labeled or unlabeled data.' This inconsistency should be resolved, and the exact label availability in the experiments should be stated.
[References] Several references have incomplete bibliographic information, including [4], [13], [16], and [18], where venue or page details are missing. In addition, several cited works concern unrelated topics (e.g., building energy prediction, cache management, lesion segmentation), which weakens the related-work framing.
[IV.B, Table 1] The table does not report the number of fault classes, the class-imbalance ratio, the number of source and target samples, or the train/validation split. Adding these details is essential for interpreting the reported metrics.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the method is a standard transfer-learning combination compared against external baselines; the main weakness is an unsupported dataset-labeling assumption, which is a validity issue, not circularity.

full rationale

The core derivation chain is self-contained and not circular. Equations (1)-(4) define a standard supervised classification loss on the source domain, an MMD distributional alignment term, an adversarial domain-discrimination term, and a joint objective; none of these equations defines its target quantity in terms of the reported outcome. The pseudo-label mechanism is explicitly described as a heuristic: high-confidence target predictions, thresholded by a confidence value, are added to training (Section III). This is a self-training loop, but it is not hidden or used to derive the empirical claim by construction; the reported Table 1 results are compared against external baselines (DANN, CDAN, FixBi, ToAlign) on accuracy, F1-score, and AUC. No self-citation is load-bearing, no uniqueness theorem is imported, and no known result is merely renamed. The significant weakness is in Section IV.A: the claim that Alibaba Cluster Trace 2018 'contains a large number of labeled fault events and clearly defined fault types' is unsupported and, if false, would invalidate the experiments. That is a reproducibility and external-validity concern rather than a circularity, because the paper does not define the target labels as the model's own outputs in the evaluation. The pseudo-label selection could become circular only if the same pseudo-labeled samples were used in both training and evaluation, but the paper does not specify such an overlap, so no specific reduction can be exhibited.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two unverified domain assumptions (shared label space and the existence of labeled fault categories in the trace) plus unspecified tuning parameters. No new physical or mathematical entities are introduced.

free parameters (4)

λ1 (MMD loss weight)
Equation (4) defines total loss with λ1 multiplying the MMD loss; the value is not reported, so any reported gain could depend on tuning it.
λ2 (adversarial loss weight)
Equation (4) defines total loss with λ2 multiplying the adversarial loss; the value is not reported.
δ (pseudo-label confidence threshold)
Section III selects target samples with prediction confidence at least δ; the threshold is not specified.
Network architecture hyperparameters
No layer sizes, optimizer, learning rate, batch size, or training epochs are given; these choices affect all reported metrics.

assumptions (3)

domain assumption Source and target domains share the same feature and label spaces, so MMD and adversarial alignment can transfer fault knowledge.
Invoked in Section III around Equations (2)-(4); no evidence is given that the label sets match or that distribution alignment improves target classification.
domain assumption Alibaba Cluster Trace 2018 contains labeled fault events and clearly defined fault types.
Section IV.A states this; the public trace is a workload and scheduling dataset and is not documented to contain root-cause fault labels, so the supervised setup is unverified.
domain assumption High-confidence target predictions are correct enough to use as pseudo-labels.
Section III's pseudo-label mechanism; the paper does not analyze threshold selection, error propagation, or confirmation bias.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Domain-Adversarial Transfer Learning for Fault Root Cause Identification in Cloud Computing Systems." pith.science (2026). https://pith.science/paper/2AMVDBHU

@misc{pith2026250702233,
  author       = {Pith},
  title        = {Pith review of: Domain-Adversarial Transfer Learning for Fault Root Cause Identification in Cloud Computing Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AMVDBHU}},
  note         = {Machine review of arXiv:2507.02233}
}

read the original abstract

This paper addresses the challenge of fault root cause identification in cloud computing environments. The difficulty arises from complex system structures, dense service coupling, and limited fault information. To solve this problem, an intelligent identification algorithm based on transfer learning is proposed. The method introduces a shared feature extraction module and a domain adversarial mechanism to enable effective knowledge transfer from the source domain to the target domain. This improves the model's discriminative ability and generalization performance in the target domain. The model incorporates a pseudo-label selection strategy. When labeled samples are lacking in the target domain, high-confidence predictions are used in training. This enhances the model's ability to recognize minority classes. To evaluate the stability and adaptability of the method in real-world scenarios, experiments are designed under three conditions: label scarcity, class imbalance, and heterogeneous node environments. Experimental results show that the proposed method outperforms existing mainstream approaches in several key metrics, including accuracy, F1-Score, and AUC. The model demonstrates stronger discriminative power and robustness. Notably, under extreme class imbalance and significant structural differences in the target domain, the model still maintains high performance. This validates the effectiveness and practical value of the proposed mechanisms in complex cloud computing systems.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Topology-Aware Graph Reinforcement Learning for Dynamic Routing in Cloud Networks
cs.LG 2025-09 reject novelty 3.0 of 10

A graph RL router with structure-aware state encoding and policy-driven edge rewiring reports improved throughput and latency on GEANT, based on single-run table comparisons.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages · cited by 1 Pith paper

[1]

A hybrid cloud and edge control strategy for demand responses using deep reinforcement learning and transfer learning,

T. Tao, J. Qiu, and S. Lai, "A hybrid cloud and edge control strategy for demand responses using deep reinforcement learning and transfer learning," IEEE Transactions on Cloud Computing, vol. 10, no. 1, pp. 56–71, 2021

work page 2021
[2]

A Deep Learning-Based Predictive Framework for Backend Latency Using AI-Augmented Structured Modeling,

Z. Fang, "A Deep Learning-Based Predictive Framework for Backend Latency Using AI-Augmented Structured Modeling," Journal of Computer Technology and Software, vol. 3, no. 7, 2024

work page 2024
[3]

Anomaly Detection in Microservice Environments via Conditional Multiscale GANs and Adaptive Temporal Autoencoders,

Y. Ma, "Anomaly Detection in Microservice Environments via Conditional Multiscale GANs and Adaptive Temporal Autoencoders," Transactions on Computational and Scientific Methods, vol. 4, no. 10, 2024

work page 2024
[4]

Multi-Scale Feature Integration and Spatial Attention for Accurate Lesion Segmentation,

Y. Wu, Y. Lin, T. Xu, X. Meng, H. Liu, and T. Kang, "Multi-Scale Feature Integration and Spatial Attention for Accurate Lesion Segmentation," 2025

work page 2025
[5]

Context-Aligned and Evidence-Based Detection of Hallucinations in Large Language Model Outputs,

Y. Peng, "Context-Aligned and Evidence-Based Detection of Hallucinations in Large Language Model Outputs," Transactions on Computational and Scientific Methods, vol. 5, no. 6, 2025

work page 2025
[6]

Exploiting transfer learning for emotion recognition under cloud-edge-client collaborations,

D. Wu, X. Han, Z. Yang, et al., "Exploiting transfer learning for emotion recognition under cloud-edge-client collaborations," IEEE Journal on Selected Areas in Communications, vol. 39, no. 2, pp. 479–490, 2020

work page 2020
[7]

Transfer learning approach to IDS on cloud IoT devices using optimized CNN,

O. D. Okey, D. C. Melgarejo, M. Saadi, et al., "Transfer learning approach to IDS on cloud IoT devices using optimized CNN," IEEE Access, vol. 11, pp. 1023–1038, 2023

work page 2023
[8]

Federated transfer learning for IIoT devices with low computing power based on blockchain and edge computing,

P. Zhang, H. Sun, J. Situ, et al., "Federated transfer learning for IIoT devices with low computing power based on blockchain and edge computing," IEEE Access, vol. 9, pp. 98630–98638, 2021

work page 2021

Show all 22 references

[9]

A survey on transfer learning,

S. Panigrahi, A. Nanda, and T. Swarnkar, "A survey on transfer learning," Intelligent and Cloud Computing: Proceedings of ICICC 2019, Volume 1, pp. 781–789, 2020

2019
[10]

A survey on deep transfer learning to edge computing for mitigating the COVID-19 pandemic,

A. Sufian, A. Ghosh, A. S. Sadiq, et al., "A survey on deep transfer learning to edge computing for mitigating the COVID-19 pandemic," Journal of Systems Architecture, vol. 108, p. 101830, 2020

2020
[11]

Forecasting workload in cloud computing: towards uncertainty-aware predictions and transfer learning,

A. Rossi, A. Visentin, D. Carraro, et al., "Forecasting workload in cloud computing: towards uncertainty-aware predictions and transfer learning," Cluster Computing, vol. 28, no. 4, p. 258, 2025

2025
[12]

A Meta-Learning Framework for Cross-Service Elastic Scaling in Cloud Environments,

T. Tang, "A Meta-Learning Framework for Cross-Service Elastic Scaling in Cloud Environments," Journal of Computer Technology and Software, vol. 3, no. 8, 2024

2024
[13]

Self-Attention-Based Modeling of Multi-Source Metrics for Performance Trend Prediction in Cloud Systems,

H. Xin and R. Pan, "Self-Attention-Based Modeling of Multi-Source Metrics for Performance Trend Prediction in Cloud Systems," Journal of Computer Technology and Software, vol. 4, no. 4, 2025

2025
[14]

Time-Series Learning for Proactive Fault Prediction in Distributed Systems with Deep Neural Structures,

Y. Wang, W. Zhu, X. Quan, H. Wang, C. Liu, and Q. Wu, "Time-Series Learning for Proactive Fault Prediction in Distributed Systems with Deep Neural Structures," arXiv preprint, arXiv:2505.20705, 2025

2025 arXiv
[15]

Continuous Control-Based Load Balancing for Distributed Systems Using TD3 Reinforcement Learning,

Y. Duan, "Continuous Control-Based Load Balancing for Distributed Systems Using TD3 Reinforcement Learning," Journal of Computer Technology and Software, vol. 3, no. 6, 2024

2024
[16]

A Deep Q-Network Approach to Intelligent Cache Management in Dynamic Backend Environments,

Y. Sun, R. Meng, R. Zhang, Q. Wu, and H. Wang, "A Deep Q-Network Approach to Intelligent Cache Management in Dynamic Backend Environments," 2025

2025
[17]

Bootstrapped Structural Prompting for Analogical Reasoning in Pretrained Language Models,

Y. Xing, "Bootstrapped Structural Prompting for Analogical Reasoning in Pretrained Language Models," Transactions on Computational and Scientific Methods, vol. 4, no. 11, 2024

2024
[18]

Unified Instruction Encoding and Gradient Coordination for Multi-Task Language Models,

W. Zhang, Z. Xu, Y. Tian, Y. Wu, M. Wang, and X. Meng, "Unified Instruction Encoding and Gradient Coordination for Multi-Task Language Models," 2025

2025
[19]

A general multi-source ensemble transfer learning framework integrate of LSTM-DANN and similarity metric for building energy prediction,

X. Fang, G. Gong, G. Li, et al., "A general multi-source ensemble transfer learning framework integrate of LSTM-DANN and similarity metric for building energy prediction," Energy and Buildings, vol. 252, p. 111435, 2021

2021
[20]

Deep transfer learning strategy based on TimesBlock-CDAN for predicting thermal environment and air conditioner energy consumption in residential buildings,

L. Sun, Z. Hu, M. Mae, et al., "Deep transfer learning strategy based on TimesBlock-CDAN for predicting thermal environment and air conditioner energy consumption in residential buildings," Applied Energy, vol. 381, p. 125188, 2025

2025
[21]

Fixbi: Bridging domain spaces for unsupervised domain adaptation,

J. Na, H. Jung, H. J. Chang, et al., "Fixbi: Bridging domain spaces for unsupervised domain adaptation," Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1094– 1103, 2021

2021
[22]

Toalign: Task-oriented alignment for unsupervised domain adaptation,

G. Wei, C. Lan, W. Zeng, et al., "Toalign: Task-oriented alignment for unsupervised domain adaptation," Advances in Neural Information Processing Systems, vol. 34, pp. 13834–13846, 2021

2021

Pith tools

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

[1] [1]

A hybrid cloud and edge control strategy for demand responses using deep reinforcement learning and transfer learning,

T. Tao, J. Qiu, and S. Lai, "A hybrid cloud and edge control strategy for demand responses using deep reinforcement learning and transfer learning," IEEE Transactions on Cloud Computing, vol. 10, no. 1, pp. 56–71, 2021

work page 2021

[2] [2]

A Deep Learning-Based Predictive Framework for Backend Latency Using AI-Augmented Structured Modeling,

Z. Fang, "A Deep Learning-Based Predictive Framework for Backend Latency Using AI-Augmented Structured Modeling," Journal of Computer Technology and Software, vol. 3, no. 7, 2024

work page 2024

[3] [3]

Anomaly Detection in Microservice Environments via Conditional Multiscale GANs and Adaptive Temporal Autoencoders,

Y. Ma, "Anomaly Detection in Microservice Environments via Conditional Multiscale GANs and Adaptive Temporal Autoencoders," Transactions on Computational and Scientific Methods, vol. 4, no. 10, 2024

work page 2024

[4] [4]

Multi-Scale Feature Integration and Spatial Attention for Accurate Lesion Segmentation,

Y. Wu, Y. Lin, T. Xu, X. Meng, H. Liu, and T. Kang, "Multi-Scale Feature Integration and Spatial Attention for Accurate Lesion Segmentation," 2025

work page 2025

[5] [5]

Context-Aligned and Evidence-Based Detection of Hallucinations in Large Language Model Outputs,

Y. Peng, "Context-Aligned and Evidence-Based Detection of Hallucinations in Large Language Model Outputs," Transactions on Computational and Scientific Methods, vol. 5, no. 6, 2025

work page 2025

[6] [6]

Exploiting transfer learning for emotion recognition under cloud-edge-client collaborations,

D. Wu, X. Han, Z. Yang, et al., "Exploiting transfer learning for emotion recognition under cloud-edge-client collaborations," IEEE Journal on Selected Areas in Communications, vol. 39, no. 2, pp. 479–490, 2020

work page 2020

[7] [7]

Transfer learning approach to IDS on cloud IoT devices using optimized CNN,

O. D. Okey, D. C. Melgarejo, M. Saadi, et al., "Transfer learning approach to IDS on cloud IoT devices using optimized CNN," IEEE Access, vol. 11, pp. 1023–1038, 2023

work page 2023

[8] [8]

Federated transfer learning for IIoT devices with low computing power based on blockchain and edge computing,

P. Zhang, H. Sun, J. Situ, et al., "Federated transfer learning for IIoT devices with low computing power based on blockchain and edge computing," IEEE Access, vol. 9, pp. 98630–98638, 2021

work page 2021