Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Anomaly Detection and Early Warning Mechanism for Intelligent Monitoring Systems in Multi-Cloud Environments Based on LLM

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding LLM-generated contextual embeddings to a CNN-LSTM feature extractor, classified by a Deep SVM, detects anomalies in multi-cloud telemetry more accurately and with lower latency than VAE, GAN, GNN, and TCN…

desk verdict The model is a standard assembly and the 'significantly better' claim is unsupported by any quantitative evidence; the garbled Bayes rule is a serious mathematical error. read the letter →

arxiv 2506.07407 v1 pith:VBOJU7QF submitted 2025-06-09 cs.LG cs.AI

Yihong Jin , Ze Yang , Juntian Liu , Xinhe Xu This is my paper
classification cs.LGcs.AI
keywords anomalydetectionmulti-cloudmonitoringlargelanguagemodelsearlywarningsystemfeatureextractiondeepSVMcloudtelemetrycontextualembeddings
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 proposes an anomaly-detection and early-warning mechanism for monitoring software that runs across multiple cloud providers. The central claim is that a pipeline which combines spatial and temporal deep features with contextual embeddings produced by a large language model, and then classifies them with a deep support vector machine, detects abnormal cloud behavior more accurately and with less delay than four established baselines. The authors argue that LLM semantic understanding lets the system adapt to different cloud vendors without hand-written rules, because it exploits log lines, event traces, and condition strings alongside numeric telemetry. The payoff of the claim, if true, is real-time alerts that catch sudden failures and volatile behavior in complex multi-cloud deployments.

What carries the argument

The load-bearing object is a multi-level feature-extraction front end that fuses three views of monitoring data: a three-layer 1D CNN (kernels of size 3, 5, and 7) for spatial correlations among CPU, memory, and network I/O metrics; a two-layer bidirectional LSTM with 128 hidden units per direction over ten-step time windows for temporal dynamics; and an LLM semantic encoder that turns template-mined log lines, event traces, and condition strings into contextual embeddings $E_{\text{LLM}} \in \mathbb{R}^{n \times d}$. A self-attention weighting mechanism combines these representations, and the fused feature vector is fed to a Deep SVM with an RBF kernel for classification. A Bayesian confidence calculation on the SVM output is what drives the early-warning decision, with the claimed effect of tuning alert confidence to the situation.

What would settle it

Run the proposed pipeline on the same telemetry dataset twice: once with the LLM encoder producing contextual embeddings and once with those embeddings removed while CNN-LSTM and Deep SVM remain unchanged. If accuracy and latency are unchanged or better without the LLM, the paper's central claim fails. A second check is to report quantitative precision, recall, F1, and end-to-end detection latency instead of the qualitative plot, so the claimed superiority over VAE, GAN, GNN, and TCN can be measured directly.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that introducing an LLM as a high-level semantic encoder improves multi-cloud anomaly detection when combined with CNN and LSTM feature extraction. Logs and event traces are tokenized and turned into contextual embeddings; these are fused with temporally and spatially extracted telemetry features through a self-attention mechanism, and a Deep SVM with an RBF kernel makes the anomaly decision. A Bayesian early-warning component is added to produce confidence-weighted alerts. The paper reports, through qualitative comparisons, that this model identifies true anomalies more consistently than variational autoencoders, GANs, graph neural networks, and temporal convolutional networks, especially in volatile periods, and that its detection latency remains favorable as hidden-layer size grows.

Load-bearing premise

The central claim stands on the untested assumption that the LLM's contextual embeddings, rather than the CNN-LSTM features alone, are what improve detection accuracy and latency; the paper provides no ablation, feature-importance analysis, or quantitative metric isolating that contribution, and the assumption enters when the LLM is introduced as a semantic encoder in Section III-A.

Editorial extensions

If this is right

  • If the claim holds, cloud operators can detect anomalies during high-volatility periods that rule-based thresholds miss, because the model learns from context rather than fixed limits.
  • The same architecture can be ported across cloud providers without rewriting detection rules, since the LLM encodes provider-specific logs into a common semantic space.
  • Confidence-weighted alarms from the Bayesian module would let operations teams prioritize responses by estimated reliability rather than treating every alert equally.
  • The reported latency trend implies the model can be tuned to balance accuracy against response speed by choosing hidden-layer size.

Reading between the lines

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

  • The paper leaves the LLM's contribution unmeasured: an ablation that removes the LLM embeddings while keeping the CNN-LSTM and Deep SVM fixed would test whether the claimed gain comes from the language model or from the deep feature stack itself.
  • If LLM embeddings prove to be the active ingredient, a natural extension is cross-cloud transfer: embeddings could let a detector trained on one provider's logs recognize anomalies on another provider with little or no fine-tuning.
  • The Bayesian update printed in the paper is circular as written, so the early-warning confidence calculation would need a corrected posterior formulation before the mechanism could be deployed as described.
  • A quantitative comparison (precision, recall, F1, and end-to-end latency) on the same dataset would turn the qualitative plot into a measurable claim that other systems could be checked against.
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 / 5 minor

Summary. The manuscript proposes a hybrid anomaly detection and early warning framework for multi-cloud monitoring systems. The architecture combines CNN and bidirectional LSTM feature extraction with an LLM-based semantic encoder, a self-attention mechanism, a Deep SVM classifier, and a Bayesian confidence update for early warning decisions. The authors evaluate the model on the IBM Console Telemetry Dataset against VAE, GAN, GNN, and TCN baselines and claim that it is significantly better in detection accuracy and latency. The paper includes an architecture figure and two qualitative comparison figures, but it reports no quantitative metrics, no ablations, no error bars, and no reproducibility artifacts such as code or data release.

Significance. If supported, the direction of combining LLM contextual embeddings with classical time-series models for cloud anomaly detection is timely and potentially useful. However, the manuscript provides no quantitative evidence for any of its headline claims: no accuracy, precision, recall, F1, false-positive/negative rates, or latency statistics are reported, and the one mathematical component specific to the early-warning mechanism (Eq. (8)) is incorrect as written. Given these gaps, the paper in its current form does not establish its central contribution, and the experimental section would need to be substantially rebuilt to support the conclusions.

major comments (4)
  1. [IV-B] The central claim that the proposed model is significantly better than traditional anomaly detection in accuracy and latency is unsupported. Figures 2 and 3 are qualitative plots with no axis labels, units, error bars, confidence intervals, or numerical values, and the text reports no precision, recall, F1-score, false-positive rate, false-negative rate, or latency distributions. The abstract and conclusion assert superiority on the basis of these figures alone, so this is a load-bearing gap that cannot be repaired by rephrasing.
  2. [III-B, Eq. (8)] Equation (8) is not a valid Bayesian update. It states p(z_i) = p(ŷ_i)p(ŷ_i)/p(z_i), which is dimensionally inconsistent and uses the same symbol p(ŷ_i) for both the prior and the likelihood. The correct confidence update would be p(ŷ_i|z_i) = p(z_i|ŷ_i)p(ŷ_i)/p(z_i). As written, the early-warning confidence mechanism is undefined and cannot be implemented from the text.
  3. [III-A] The role of the LLM is not specified at the level needed to reproduce or test the claimed contribution. The text states that the LLM outputs contextual embeddings E_LLM in R^{n×d} and that context vectors c_i in R^k are used with self-attention, but it never defines which LLM is used, how the embeddings are generated, how c_i is derived from E_LLM, or how the LLM features are fused with the concatenated CNN/LSTM features before the Deep SVM classifier. Without these details, the benefit of the LLM cannot be isolated or ablated.
  4. [IV-A] The experimental setup is not reproducible. The dataset description, 'more than 39,000 rows and 117,000 columns of data,' is implausible as stated, and no train/test split, preprocessing pipeline, label definition, class balance, or anomaly ratio is provided. Baseline implementations and hyperparameters are absent, and no LLM configuration is reported. Consequently, the comparisons depicted in Figures 2 and 3 cannot be interpreted or verified.
minor comments (5)
  1. [IV-A] The list of baselines is introduced as 'four thresholds-based anomaly detection methods,' but VAE, GAN, GNN, and TCN are not threshold-based methods; the characterization is inconsistent with the descriptions that follow.
  2. [V] The conclusion refers to a 'Massive Language Model (LLM),' which is inconsistent with the 'Large-Scale Language Model' terminology used in the abstract and introduction; the terminology should be unified.
  3. [Figures 2 and 3] Both figures lack axis labels and units, and their captions are too terse to indicate what quantity is plotted or which time period is shown; the figures should be made self-contained.
  4. [References] Reference [2] is cited to support statements about cost reduction and avoiding vendor lock-in, but the cited paper concerns hardware-accelerated decoding for LLM speculation and does not appear to support that claim; the citation should be corrected.
  5. [III-B] The text contains formatting artifacts, including 'leftmargin=1.5em' in the baseline list and the duplicated phrase 'GNN captures' in the GNN bullet; these should be cleaned up.

Circularity Check

1 steps flagged · score 6.0 of 10

The only reducible step is §III-B Eq. (8), where the 'Bayesian' confidence p(z_i) is defined in terms of itself, so the early-warning posterior collapses to the prior; self-citations are background and not load-bearing.

  1. self definitional [Section III-B, Equation (8)]
    "Assuming that the output of our anomaly detection model is p(ˆyi), we utilize Bayesian inference to calculate the confidence of the anomaly p(ˆyi|zi), giving an early warning decision, as in Equation (8): p(zi) = p(ˆyi)p(ˆyi) / p(zi), where p(ˆyi) is the prior probability, p(ˆyi) is the likelihood function, and p(zi) is the marginal probability of the data."

    The quantity being computed, p(z_i), appears on both sides of Equation (8). Rearranging gives p(z_i)^2 = p(ˆy_i)^2, and with nonnegative probabilities p(z_i)=p(ˆy_i). Substituting back, the claimed posterior p(ˆy_i|z_i) equals the prior p(ˆy_i); no data or context updates the confidence. Thus the early-warning/confidence output is fixed by the equation itself rather than derived from the model output or observations. This is a self-definitional step: the 'Bayesian inference' result is equivalent to its own input (the prior) by construction, not to the likelihood it purports to use.

full rationale

The paper's empirical comparison is not circular: Figures 2 and 3 are qualitative, and the absence of metrics is an evidence gap, not a reduction of the headline claim to its inputs. The self-citations in the introduction and related work ([2], [5], [6], [12]) provide background motivation (multi-cloud complexity, LLM log processing) but are not used to derive the architecture or the experimental result, so they are not load-bearing. The single genuine circular step is Equation (8) in the early-warning mechanism: the left-hand side p(z_i) is repeated in the denominator of the right-hand side, so the equation has the trivial solution p(z_i)=p(ˆy_i) and the claimed Bayesian confidence update degenerates to the prior. Because this reduces a stated output of the proposed mechanism (the confidence-based early warning decision) to its own prior by construction, but does not force the overall anomaly-detection comparison, the circularity is partial rather than total.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the specific hyperparameter choices for the CNN, LSTM, and SVM, none of which are justified or varied. The key modeling assumptions are that the dataset is representative of multi-cloud environments and that LLM embeddings add signal, both unverified. Equation (8) is an incorrect statement of Bayes' theorem.

free parameters (9)
  • CNN kernel sizes = 3, 5, 7
    Hand-chosen design choices described in Section III-A.
  • LSTM hidden units = 128 per direction
    Hand-chosen design choice described in Section III-A.
  • LSTM window size = 10 steps
    Hand-chosen design choice described in Section III-A.
  • CNN output dimension = 64
    Hand-chosen design choice described in Section III-A.
  • LSTM output dimension = 256
    Hand-chosen design choice described in Section III-A.
  • SVM regularization parameter C = not specified
    Appears in Eq. (5) without a reported value.
  • Learning rate eta = not specified
    Appears in Eq. (7) without a reported value.
  • Attention key dimension d_k = not specified
    Appears in Eq. (4) without a reported value.
  • LLM embedding dimension d = not specified
    Mentioned in Section III-A without a reported value.
assumptions (4)
  • standard math Standard deep learning and SVM mathematics are valid.
    Equations (1)-(7) rely on standard CNN, LSTM, attention, and SVM formulations.
  • domain assumption The IBM Console Telemetry Dataset is representative of multi-cloud environments.
    Section IV-A states the dataset is suitable without evidence of representative coverage across cloud providers.
  • domain assumption LLM embeddings improve anomaly detection over numeric features alone.
    This is asserted in Section III-A but never tested with an ablation or feature importance analysis.
  • ad hoc to paper Equation (8) is a valid Bayesian update.
    Section III-B writes p(zi) = p(ŷi)p(ŷi)/p(zi), which is not a correct form of Bayes' theorem and cannot compute a confidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Anomaly Detection and Early Warning Mechanism for Intelligent Monitoring Systems in Multi-Cloud Environments Based on LLM." pith.science (2026). https://pith.science/paper/VBOJU7QF

@misc{pith2026250607407,
  author       = {Pith},
  title        = {Pith review of: Anomaly Detection and Early Warning Mechanism for Intelligent Monitoring Systems in Multi-Cloud Environments Based on LLM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VBOJU7QF}},
  note         = {Machine review of arXiv:2506.07407}
}
read the original abstract

With the rapid development of multi-cloud environments, it is increasingly important to ensure the security and reliability of intelligent monitoring systems. In this paper, we propose an anomaly detection and early warning mechanism for intelligent monitoring system in multi-cloud environment based on Large-Scale Language Model (LLM). On the basis of the existing monitoring framework, the proposed model innovatively introduces a multi-level feature extraction method, which combines the natural language processing ability of LLM with traditional machine learning methods to enhance the accuracy of anomaly detection and improve the real-time response efficiency. By introducing the contextual understanding capabilities of LLMs, the model dynamically adapts to different cloud service providers and environments, so as to more effectively detect abnormal patterns and predict potential failures. Experimental results show that the proposed model is significantly better than the traditional anomaly detection system in terms of detection accuracy and latency, and significantly improves the resilience and active management ability of cloud infrastructure.

Figures

Figures reproduced from arXiv: 2506.07407 by the authors.

Figure 1
Figure 1. Architecture of Multi-level Feature Extraction and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Anomaly Detection: Comparison of Methods. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of Detection Delay Across Models. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 19 canonical work pages

  1. [1]

    Artificial intelligence-aided digital twin design: A systematic review,

    N. Hao, Y . Li, K. Liu, S. Liu, Y . Lu, B. Xu, C. Li, J. Chen, L. Yue, T. Fu et al., “Artificial intelligence-aided digital twin design: A systematic review,”Preprints, 2024

    work page 2024

  2. [2]

    Hades: Hardware accelerated decoding for efficient speculation in large language models,

    Z. Yang, Y . Jin, and X. Xu, “Hades: Hardware accelerated decoding for efficient speculation in large language models,”arXiv preprint arXiv:2412.19925, 2024

    arXiv 2024

  3. [3]

    Autonomous anomaly detection in cloud security: Ai and dspm with machine learning,

    F. Hussain and M. Daniel, “Autonomous anomaly detection in cloud security: Ai and dspm with machine learning,” 2024

    work page 2024

  4. [4]

    Proactive cloud security with dspm: Ai and machine learning-based anomaly detection,

    W. Sadeq and J. Yang, “Proactive cloud security with dspm: Ai and machine learning-based anomaly detection,” 2022

    work page 2022

  5. [5]

    Advances in appfl: A comprehensive and extensible federated learning framework,

    Z. Li, S. He, Z. Yang, M. Ryu, K. Kim, and R. Madduri, “Advances in appfl: A comprehensive and extensible federated learning framework,” arXiv preprint arXiv:2409.11585, 2024

    arXiv 2024

  6. [6]

    Research on large language model cross-cloud privacy protection and collaborative training based on federated learning,

    Z. Yang, Y . Jin, Y . Zhang, J. Liu, and X. Xu, “Research on large language model cross-cloud privacy protection and collaborative training based on federated learning,”arXiv preprint arXiv:2503.12226, 2025

    arXiv 2025

  7. [7]

    Cloud-based ai systems: Leveraging large language models for intelligent fault detection and autonomous self-healing,

    C. Ji and H. Luo, “Cloud-based ai systems: Leveraging large language models for intelligent fault detection and autonomous self-healing,” arXiv preprint arXiv:2505.11743, 2025

    arXiv 2025

  8. [8]

    Nlp-adbench: Nlp anomaly detection benchmark,

    Y . Li, J. Li, Z. Xiao, T. Yang, Y . Nian, X. Hu, and Y . Zhao, “Nlp-adbench: Nlp anomaly detection benchmark,”arXiv preprint arXiv:2412.04784, 2024

    arXiv 2024

Show all 44 references
  1. [9]

    A decision-language model (dlm) for dynamic restless multi-armed bandit tasks in public health,

    N. Behari, E. Zhang, Y . ZHAO, A. Taneja, D. M. Nagaraj, and M. Tambe, “A decision-language model (dlm) for dynamic restless multi-armed bandit tasks in public health,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems

  2. [10]

    Scalable neural network kernels,

    A. Sehanobish, K. M. Choromanski, Y . ZHAO, K. A. Dubey, and V . Likhosherstov, “Scalable neural network kernels,” inThe Twelfth International Conference on Learning Representations

  3. [11]

    RAG-RLRC-LaySum at BioLaySumm: Integrating retrieval-augmented generation and readability control for layman summarization of biomedical texts,

    Y . Ji, Z. Li, R. Meng, S. Sivarajkumar, Y . Wang, Z. Yu, H. Ji, Y . Han, H. Zeng, and D. He, “RAG-RLRC-LaySum at BioLaySumm: Integrating retrieval-augmented generation and readability control for layman summarization of biomedical texts,” inProceedings of the 23rd Workshop on...

    2024

  4. [12]

    Adaptive fault tolerance mechanisms of large language models in cloud computing environ- ments,

    Y . Jin, Z. Yang, X. Xu, Y . Zhang, and S. Ji, “Adaptive fault tolerance mechanisms of large language models in cloud computing environ- ments,”arXiv preprint arXiv:2503.12228, 2025

    2025 arXiv

  5. [13]

    Time series modeling for heart rate prediction: From arima to transformers,

    H. Ni, S. Meng, X. Geng, P. Li, Z. Li, X. Chen, X. Wang, and S. Zhang, “Time series modeling for heart rate prediction: From arima to transformers,” in2024 6th International Conference on Electronic Engineering and Informatics (EEI). IEEE, 2024, pp. 584–589

    2024

  6. [14]

    Cross-cloud data privacy protection: Optimizing collaborative mechanisms of ai systems by integrating federated learning and llms,

    H. Luo and C. Ji, “Cross-cloud data privacy protection: Optimizing collaborative mechanisms of ai systems by integrating federated learning and llms,”arXiv preprint arXiv:2505.13292, 2025

    2025 arXiv

  7. [15]

    Proactive security in multi-cloud environments: A blockchain integrated real-time anomaly detection and mitigation framework,

    J. M. Lakshmi, K. Krishna Prasad, and G. Viswanath, “Proactive security in multi-cloud environments: A blockchain integrated real-time anomaly detection and mitigation framework,”Cuestiones de Fisioter- apia, vol. 54, no. 2, pp. 392–417, 2025

    2025

  8. [16]

    Enhancing network security in multi-cloud environments through adaptive threat detection

    K. O. Chauke, T. Muchenje, and N. Makondo, “Enhancing network security in multi-cloud environments through adaptive threat detection.”

  9. [17]

    Security challenges in multi-cloud environments: Solutions and best practices,

    M. Madanan, P. Patel, P. Agrawal, P. Mudholkar, M. Mudholkar, and V . Jaganraja, “Security challenges in multi-cloud environments: Solutions and best practices,” in2024 7th International Conference on Contemporary Computing and Informatics (IC3I), vol. 7. IEEE, 2024, pp. 1608–1614

    2024

  10. [18]

    Ai-driven anomaly detection: Strengthening multi-cloud security,

    D. Neol, I. Jacob, and S. Andrewson, “Ai-driven anomaly detection: Strengthening multi-cloud security,” 2025

    2025

  11. [19]

    Research on the improvement of cyber- security early warning systems based on cloud computing platforms,

    W. Kou, G. Zhang, and C. Yan, “Research on the improvement of cyber- security early warning systems based on cloud computing platforms,” in 2024 IEEE 6th International Conference on Civil Aviation Safety and Information Technology (ICCASIT). IEEE, 2024, pp. 1115–1121

    2024

  12. [20]

    Multi-cloud resource scheduling intelligent system with endogenous security

    N. Cai and G. He, “Multi-cloud resource scheduling intelligent system with endogenous security.”Electronic Research Archive, vol. 32, no. 2, 2024

    2024

  13. [21]

    Give: Structured reasoning of large language models with knowledge graph inspired veracity extrapolation,

    J. He, M. D. Ma, J. Fan, D. Roth, W. Wang, and A. Ribeiro, “Give: Structured reasoning of large language models with knowledge graph inspired veracity extrapolation,” 2025. [Online]. Available: https://arxiv.org/abs/2410.08475

    2025 arXiv

  14. [22]

    Self-give: Associative thinking from limited structured knowledge for enhanced large language model reasoning,

    J. He, J. Fan, B. Jiang, I. Houine, D. Roth, and A. Ribeiro, “Self-give: Associative thinking from limited structured knowledge for enhanced large language model reasoning,” 2025. [Online]. Available: https://arxiv.org/abs/2505.15062

    2025

  15. [23]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”arXiv e-prints, pp. arXiv–1312, 2013

    2013

  16. [24]

    Fine-grained control of generative data aug- mentation in iot sensing,

    T. Wang, Q. Yang, R. Wang, D. Sun, J. Li, Y . Chen, Y . Hu, C. Yang, T. Kimura, D. Karaet al., “Fine-grained control of generative data aug- mentation in iot sensing,”Advances in Neural Information Processing Systems, vol. 37, pp. 32 787–32 812, 2024

    2024

  17. [25]

    Exploring diverse methods in visual question answering,

    P. Li, Q. Yang, X. Geng, W. Zhou, Z. Ding, and Y . Nian, “Exploring diverse methods in visual question answering,” in2024 5th International Conference on Electronic Communication and Artificial Intelligence (ICECAI). IEEE, 2024, pp. 681–685

    2024

  18. [26]

    Contextual hourglass network for semantic segmentation of high resolution aerial imagery,

    P. Li, Y . Lin, and E. Schultz-Fellenz, “Contextual hourglass network for semantic segmentation of high resolution aerial imagery,” in2024 5th International Conference on Electronic Communication and Artificial Intelligence (ICECAI). IEEE, 2024, pp. 15–18

    2024

  19. [27]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

    2020

  20. [28]

    Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks,

    P. Li, M. Abouelenien, R. Mihalcea, Z. Ding, Q. Yang, and Y . Zhou, “Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks,” in2024 5th International Conference on Information Science, Parallel and Distributed Systems (I...

    2024

  21. [29]

    A comparative study on enhancing prediction in social network advertisement through data augmentation,

    Q. Yang, P. Li, X. Xu, Z. Ding, W. Zhou, and Y . Nian, “A comparative study on enhancing prediction in social network advertisement through data augmentation,” in2024 4th International Conference on Machine Learning and Intelligent Systems Engineering (MLISE). IEEE, 2024, pp. 214–218

    2024

  22. [30]

    Style transfer: From stitching to neural networks,

    X. Xu, Z. Wang, Y . Zhang, Y . Liu, Z. Wang, Z. Xu, M. Zhao, and H. Luo, “Style transfer: From stitching to neural networks,” in2024 5th International Conference on Big Data & Artificial Intelligence & Software Engineering (ICBASE). IEEE, 2024, pp. 526–530

    2024

  23. [31]

    Graph neural networks: A review of methods and applications,

    J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M. Sun, “Graph neural networks: A review of methods and applications,” AI open, vol. 1, pp. 57–81, 2020

    2020

  24. [32]

    Handling missing sensors in topology-aware iot applications with gated graph neural network,

    S. Liu, S. Yao, Y . Huang, D. Liu, H. Shao, Y . Zhao, J. Li, T. Wang, R. Wang, C. Yanget al., “Handling missing sensors in topology-aware iot applications with gated graph neural network,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol....

    2020

  25. [33]

    Influence pathway discovery on social media,

    X. Liu, R. Wang, D. Sun, J. Li, C. Youn, Y . Lyu, J. Zhan, D. Wu, X. Xu, M. Liuet al., “Influence pathway discovery on social media,” in 2023 IEEE 9th International Conference on Collaboration and Internet Computing (CIC). IEEE, 2023, pp. 105–109

    2023

  26. [34]

    Revisiting over- smoothing in deep gcns,

    C. Yang, R. Wang, S. Yao, S. Liu, and T. Abdelzaher, “Revisiting over- smoothing in deep gcns,”arXiv preprint arXiv:2003.13663, 2020

    2003 arXiv

  27. [35]

    T-GAE: Transferable Graph Autoencoder for Network Alignment,

    J. He, C. I. Kanatsoulis, and A. Ribeiro, “T-GAE: Transferable Graph Autoencoder for Network Alignment,”arXiv e-prints, p. arXiv:2310.03272, Oct. 2023

    2023 arXiv

  28. [36]

    Graphsnapshot: Graph machine learning acceleration with fast storage and retrieval,

    D. Liu, R. Waleffe, M. Jiang, and S. Venkataraman, “Graphsnapshot: Graph machine learning acceleration with fast storage and retrieval,” arXiv preprint arXiv:2406.17918, 2024

    2024 arXiv

  29. [37]

    Drtr: Distance-aware graph representation learning,

    D. Liu and Y . Yu, “Drtr: Distance-aware graph representation learning,” inAI4X 2025 International Conference

    2025

  30. [38]

    Llmeasyquant: Scalable quantization for parallel and distributed llm inference,

    ——, “Llmeasyquant: Scalable quantization for parallel and distributed llm inference,” inAI4X 2025 International Conference

    2025

  31. [39]

    Contemporary model compression on large language models inference,

    D. Liu, “Contemporary model compression on large language models inference,”arXiv preprint arXiv:2409.01990, 2024

    2024 arXiv

  32. [40]

    Fast- cache: Fast caching for diffusion transformer through learnable linear approximation,

    D. Liu, J. Zhang, Y . Li, Y . Yu, B. Lengerich, and Y . N. Wu, “Fast- cache: Fast caching for diffusion transformer through learnable linear approximation,”arXiv preprint arXiv:2505.20353, 2025

    2025

  33. [41]

    Mitigating the risk of health inequity exacerbated by large language models,

    Y . Ji, W. Ma, S. Sivarajkumar, H. Zhang, E. M. Sadhu, Z. Li, X. Wu, S. Visweswaran, and Y . Wang, “Mitigating the risk of health inequity exacerbated by large language models,”npj Digital Medicine, vol. 8, no. 1, p. 246, 2025. [Online]. Available: https://doi.org/10.1038/s417...

    2025 doi

  34. [42]

    Unveiling markov heads in pretrained language models for offline re- inforcement learning,

    W. Zhao, Q. Xu, L. Xu, L. Song, J. Wang, C. Zhou, and J. Bian, “Unveiling markov heads in pretrained language models for offline re- inforcement learning,” inProceedings of the Forty-second International Conference on Machine Learning, 2025

    2025

  35. [43]

    A stochastic gda method with backtracking for solving nonconvex (strongly) concave minimax problems,

    Q. Xu, X. Zhang, N. S. Aybat, and M. G ¨urb¨uzbalaban, “A stochastic gda method with backtracking for solving nonconvex (strongly) concave minimax problems,”arXiv preprint arXiv:2403.07806, 2024

    2024 arXiv

  36. [44]

    Agda+: Proximal alternating gradient descent ascent method with a nonmonotone adaptive step-size search for nonconvex minimax problems,

    X. Zhang, Q. Xu, and N. S. Aybat, “Agda+: Proximal alternating gradient descent ascent method with a nonmonotone adaptive step-size search for nonconvex minimax problems,”arXiv preprint arXiv:2406.14371, 2024

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.