Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Research on Cloud Platform Network Traffic Monitoring and Anomaly Detection System based on Large Language Models

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims a hybrid LLM, self-attention, and autoencoder system detects cloud network attacks more accurately and with fewer false positives than classical baselines on CICIDS 2017.

desk verdict Under-specified system sketch: the LLM is never defined and no accuracy numbers appear, so the central performance claim is untestable. read the letter →

arxiv 2504.17807 v1 pith:TUOB65OY submitted 2025-04-22 cs.NI cs.AIcs.LG

Ze Yang , Yihong Jin , Juntian Liu , Xinhe Xu , Yihan Zhang , Shuyang Ji This is my paper · ORCID
classification cs.NIcs.AIcs.LG
keywords cloudnetworktrafficanomalydetectionlargelanguagemodelsself-attentionautoencodertransferlearningCICIDS2017falsepositiverate
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 argues that a hybrid architecture—a pre-trained large language model paired with Transformer self-attention, an autoencoder-based anomaly layer, and transfer learning—can monitor cloud network traffic and detect anomalies more accurately and with fewer false positives than classical baselines such as autoencoder, random forest, SVM, LSTM, and GNN. The motivation is that cloud traffic is high-dimensional, strongly time-correlated, and subject to novel or shifting attack patterns that rule-based or static deep-learning detectors miss. The authors report that their model detects zero-day attacks and congestion patterns on the CICIDS 2017 dataset while keeping a favorable training-to-inference speed trade-off. If the claim holds, cloud operators would gain a single model that both predicts expected traffic sequences and flags deviations, adapting to new environments via transfer learning without large labeled datasets.

What carries the argument

The load-bearing mechanism is the self-attention matrix $A_{tt'} = \mathrm{Softmax}((X_t W_q)(X_{t'}^{\phantom{'}} W_k)^{\top}/\sqrt{d_k})$, computed over normalized traffic features described as bandwidth, latency, and packet loss; it encodes pairwise influence between time points, and the weighted sum $Z_t = \sum_{t'} A_{tt'} X_{t'}$ gives each time point a context-aware representation. Anomaly detection then minimizes $L_{\text{detect}} = \sum_{t=1}^{T} ( \|Z_t - \hat{Z}_t\|_2^2 + \lambda_t \|A_t\|_1 )$, combining reconstruction error with a sparsity penalty on the attention matrix. Transfer learning is added through $L_{\text{transfer}} = L_{\text{source}} + \sum_i \lambda_i L_{\text{target},i}$, so the model can adapt to new network environments without full retraining.

What would settle it

Re-run the CICIDS 2017 70/15/15 split with the paper's equations (1)–(5) but replace the claimed pre-trained language model with a randomly initialized self-attention encoder of comparable size; if accuracy and false-positive rate match the paper's reported 'Ours' numbers, the LLM component is not load-bearing. A second check: reproduce the 10-run computation-cost comparison; if 'Ours' costs close to 12.5 seconds while a genuine LLM forward pass would dominate inference, the claimed efficiency would not hold.

Watch

Extended reading notes

Core claim

The central claim is that representing normalized cloud-traffic feature vectors as a temporal sequence and processing them through a Transformer-style self-attention layer, with an autoencoder reconstruction loss and a transfer-learning objective stacked on top, outperforms traditional anomaly detectors in both accuracy and computational cost. The paper additionally claims that the pre-trained language model analyzes and predicts the probable network traffic, while the anomaly layer accounts for temporality and context, and that the resulting system can identify zero-day attacks and traffic congestion with a significantly reduced false-positive rate. Experimental results are reported against autoencoder, random forest, SVM, LSTM, and GNN baselines on CICIDS 2017.

Load-bearing premise

The load-bearing premise is that a pre-trained large language model can meaningfully process the standardized numeric traffic features (bandwidth, latency, packet loss) as if they were sequence tokens, and that language pretraining transfers to network anomaly detection; the paper does not specify the LLM, the tokenization, or whether the experimental 'Ours' model actually contains a pre-trained language model rather than just a small self-attention layer.

Editorial extensions

If this is right

  • Cloud monitoring systems could use one model that both predicts expected traffic and flags deviations, with inference speed suitable for real-time deployment.
  • New or mutated attack types could be absorbed through transfer learning from a source task, reducing the need for large labeled datasets in each new environment.
  • Combining reconstruction error with an attention-sparsity penalty could lower false-positive rates relative to threshold-based or autoencoder-only detectors.
  • The reported training-versus-inference trade-off suggests the hybrid is more deployable than LSTM-based detectors in latency-sensitive cloud settings.

Reading between the lines

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

  • The paper never names the pre-trained LLM or the tokenization scheme; a plausible reading is that the 'LLM' is a small Transformer encoder, in which case the language-model framing may be incidental and the contribution reduces to a self-attention autoencoder.
  • The zero-day claim is testable by holding out one attack class during training and measuring detection on it; the paper does not report such an experiment, so this remains an open inference.
  • The same hybrid loss could transfer to other streaming telemetry domains, such as server metrics or financial transactions, wherever rare anomalies sit inside strongly time-correlated sequences.
  • If the reported computation cost is accurate, the hybrid's 12.5-second total is suspiciously close to the GNN baseline of 12.3 seconds; a reader should check whether the LLM is actually invoked at inference time.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 manuscript proposes a hybrid system for cloud network traffic monitoring and anomaly detection that combines a pre-trained large language model (LLM) with self-attention, an autoencoder-style reconstruction loss, transfer learning, contrastive learning, and Bayesian optimization. The method is evaluated on the CICIDS 2017 dataset against Autoencoder, Random Forest, SVM, LSTM, and GNN baselines, and the authors claim superior accuracy, lower false positive rate, and better computational efficiency, including adaptability to zero-day attacks. The paper provides no quantitative performance numbers; Figures 1-3 are presented only as captions, Table I contains a single cost comparison, and the LLM component is never concretely specified.

Significance. The problem addressed is important and the design idea of using pretrained sequence models for traffic anomaly detection is plausible. If the claimed results were properly evidenced, the work could be of interest to the network-intrusion-detection community. However, as submitted, the contribution is not assessable: the architecture is under-specified, the experimental evidence consists of uninterpretable figure references and a single table without variance, and the central claims are therefore unsupported. The paper also correctly identifies limitations of rule-based methods and the need to model temporal dependencies, which are reasonable motivations. No reproducible code, dataset-processing details, or falsifiable quantitative predictions are provided.

major comments (5)
  1. [Abstract; Section IV-B] The central claim that the proposed model outperforms traditional methods in detection accuracy and computational efficiency and significantly reduces the false positive rate is not supported by any reported quantitative metric. Section IV-B refers to Figures 1-3 only by caption, with qualitative statements such as "the Ours method was superior," and Table I reports a single set of computation costs without variance, hardware configuration, or statistical tests. No accuracy, precision, recall, F1, false-positive rate, or detection-latency numbers appear anywhere in the text, so the abstract's performance claims cannot be verified.
  2. [Section III-A; Table I] The manuscript never specifies which pre-trained LLM is used, how the normalized numeric feature vectors X'_t are tokenized or embedded into the sequence input, or what the model's parameter count and context length are. Without this, the reader cannot determine whether the reported "Ours" model is actually a pretrained LLM as claimed, and the asserted benefits of LLM pretraining for zero-day attack detection have no basis. Moreover, Table I lists "Ours" at 12.5 s versus 10.7 s for the Autoencoder, which directly contradicts the computational-efficiency claim; either a different, statistically grounded metric should be reported or this claim should be retracted.
  3. [Section III-B, Eq. (4)] The anomaly detection loss in Eq. (4) combines a reconstruction error with an L1 penalty on the attention matrix, but the paper does not specify how this loss is converted into an anomaly score or how a detection threshold is chosen. Figure 3 is said to show F1-score versus thresholds, but no threshold-selection procedure or actual F1 values are reported. Consequently, the claimed false-positive reduction cannot be evaluated or reproduced.
  4. [Section III-B, Eq. (5)] Eq. (5) defines L_transfer as a weighted sum of a source loss and m target losses. This is a standard multi-task and transfer objective and does not constitute a "novel transfer learning methodology" as claimed in the abstract; no mechanism is described for adapting to unknown network structures and adversarial conditions, and no experiment demonstrates adaptation with limited labeled data. This component should either be developed substantively or omitted from the contributions.
  5. [Section IV-A] The experimental setup omits essential details: the authors say they chose two types of data from CICIDS 2017 but do not state which attack types, the number of samples, the class balance, the feature dimensionality, or the time-series sharding scheme. In addition, Section IV-B lists Autoencoder, Random Forest, SVM, and LSTM as comparison methods, while Table I also includes GNN, whose definition and configuration are never given. These omissions prevent reproduction and make it impossible to assess whether the comparison is fair.
minor comments (5)
  1. [Section IV-B] The sentence "Model capable of handling time series data modeling but has low training efficiency and tends to overfit heterogeneous data" appears to be missing its subject; presumably "LSTM" is intended.
  2. [Abstract; throughout] The text contains numerous grammatical errors (e.g., "a better capture," "effectively identify") and inconsistent phrasing, and it would benefit from a careful editing pass.
  3. [Section II] Several cited papers are not clearly connected to the proposed method (e.g., [12], [16], [21]), and there is no discussion of prior LLM-based network intrusion detection systems, which makes the related-work positioning difficult.
  4. [Section III-B] The sentence "Then with contrastive learning applied, the feature discrimination ability will improve..." is a fragment, and no details are given on the contrastive loss formulation or on the Bayesian optimization procedure that is mentioned.
  5. [Figures] Figures 1-3 are not included in the manuscript body; only captions are present, so the claimed experimental comparisons cannot be inspected even visually.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity: 'novel transfer learning' is a renamed multi-task loss; the larger empirical claim is unsupported but not circular.

  1. renaming known result [Abstract and Section III-B, Equation (5)]
    "Moreover, we present a novel transfer learning-based methodology to enhance the model's effectiveness to quickly adapt to unknown network structures and adversarial conditions without requiring extensive labeled datasets. ... Ltransfer = Lsource + m X i=1 λiLtarget,i, (5)"

    Equation (5) is, by construction, the standard multi-task learning objective: a weighted sum of the source-task loss and the target-task losses. The paper labels this definitional weighting a 'novel transfer learning methodology' and credits it with quick adaptation, but no adaptation rule, parameter-sharing scheme, or additional mechanism is specified. The claimed methodological novelty therefore reduces to renaming a known construction; the adaptation benefit is asserted from the equation rather than derived from any independent content.

full rationale

Most of the paper's derivation consists of standard components—z-scoring (Eq. 1), self-attention (Eqs. 2–3), autoencoder reconstruction plus L1 penalty (Eq. 4)—none of which are fitted to the target results or derived from the paper's own claims. The central empirical claim of superior accuracy and efficiency on CICIDS 2017 is asserted from figures and a ten-run table without reported numbers, error bars, or a specification of the LLM; that is a lack of falsifiable evidence rather than circularity. The only step that reduces by construction is the 'novel transfer learning-based methodology' of Eq. 5, which is simply the multi-task objective Lsource + sum λi Ltarget,i; presenting it as a novel methodology is a renamed known formulation. This step is definitional but not load-bearing for the external baseline comparison, and the paper invokes no same-author uniqueness theorem or self-citation chain to force its architecture. Overall circularity is therefore minor.

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

The central claims depend on several unstated modeling assumptions: that a pre-trained LLM transfers to tabular traffic features, that CICIDS 2017 represents cloud zero-day conditions, that the reconstruction-plus-attention loss separates normal from abnormal traffic, and that multi-task transfer helps without labeled data. The paper also leaves hyperparameters unspecified, though Bayesian optimization is mentioned. No new physical or technical entities are introduced.

free parameters (4)
  • Regularization factor lambda_t in anomaly detection loss (Eq. 4) = Not reported; described as auto-tuned via Bayesian optimization
    Controls the weight of the attention-based anomaly term; no value or search range is given, yet the F1 threshold curves in Fig. 3 depend on this kind of weighting.
  • Transfer-learning task weights lambda_i (Eq. 5) = Not reported
    Weights for each target task in the multi-task transfer loss; no values or tuning procedure are reported.
  • Anomaly threshold = Not reported
    Fig. 3 shows F1 varies with threshold and implies threshold selection, but the chosen operating threshold is never stated.
  • Transformer architecture hyperparameters (layers, heads, hidden size, sequence length) = Not reported
    The method is described only by Eq. 2 and Eq. 3; no architecture size or pretrained checkpoint is identified, so the 'LLM' part is unspecified.
assumptions (4)
  • domain assumption CICIDS 2017 is a valid proxy for cloud-platform network traffic and contains zero-day-style anomalies.
    Section IV-A uses CICIDS 2017 to support claims about real cloud traffic and zero-day attacks, but the dataset is a labeled benchmark with known attack classes, not an unlabeled zero-day deployment.
  • ad hoc to paper A pre-trained LLM can process standardized numerical traffic feature vectors as sequence tokens and transfer its pretraining to anomaly detection.
    Section III-A feeds normalized features X'_t into the self-attention mechanism; no tokenization, pretraining corpus, or evidence for cross-domain transfer is provided.
  • ad hoc to paper The multi-task loss in Eq. 5 improves adaptation to new network environments without extensive labeled data.
    The paper asserts transfer-learning benefits but reports no transfer experiment, such as training on one network and testing on another.
  • domain assumption Reconstruction error in Eq. 4 is a sufficient anomaly signal for the supervised detection task.
    The loss combines reconstruction and attention regularization; no derivation or ablation shows that this score separates normal from abnormal traffic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Research on Cloud Platform Network Traffic Monitoring and Anomaly Detection System based on Large Language Models." pith.science (2026). https://pith.science/paper/TUOB65OY

@misc{pith2026250417807,
  author       = {Pith},
  title        = {Pith review of: Research on Cloud Platform Network Traffic Monitoring and Anomaly Detection System based on Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TUOB65OY}},
  note         = {Machine review of arXiv:2504.17807}
}
read the original abstract

The rapidly evolving cloud platforms and the escalating complexity of network traffic demand proper network traffic monitoring and anomaly detection to ensure network security and performance. This paper introduces a large language model (LLM)-based network traffic monitoring and anomaly detection system. In addition to existing models such as autoencoders and decision trees, we harness the power of large language models for processing sequence data from network traffic, which allows us a better capture of underlying complex patterns, as well as slight fluctuations in the dataset. We show for a given detection task, the need for a hybrid model that incorporates the attention mechanism of the transformer architecture into a supervised learning framework in order to achieve better accuracy. A pre-trained large language model analyzes and predicts the probable network traffic, and an anomaly detection layer that considers temporality and context is added. Moreover, we present a novel transfer learning-based methodology to enhance the model's effectiveness to quickly adapt to unknown network structures and adversarial conditions without requiring extensive labeled datasets. Actual results show that the designed model outperforms traditional methods in detection accuracy and computational efficiency, effectively identify various network anomalies such as zero-day attacks and traffic congestion pattern, and significantly reduce the false positive rate.

Figures

Figures reproduced from arXiv: 2504.17807 by the authors.

Figure 1
Figure 1. Comparison of Accuracy for Different Methods [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. F1-Score with Different Thresholds Additionally, we compare our proposed model with existing advanced and traditional model through computing the detec￾tion process 10 times Methods Auto encoder Random Forest SVM LSTM GNN Ours Computation Cost (second) 10.7 24.5 17.4 18.3 12.3 12.5 TABLE I: Computation Cost Comparison V. CONCLUSION The work presented here provides insight into the strengths and weaknesses of a varie… view at source ↗
Figure 2
Figure 2. Comparison Results in Training Time and Inference [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Federated Learning-Based Data Collaboration Method for Enhancing Edge Cloud AI System Security Using Large Language Models

    cs.CR 2025-06 reject novelty 3.0 of 10

    The paper combines LLM-guided update weighting, selective SMC, and adversarial training for federated learning, claiming a 15% robustness gain without presenting the underlying experiment.

  2. Cross-Cloud Data Privacy Protection: Optimizing Collaborative Mechanisms of AI Systems by Integrating Federated Learning and LLMs

    cs.CR 2025-05 reject novelty 2.0 of 10

    The paper asserts that combining federated learning, large language model features, and homomorphic encryption delivers the best cross-cloud privacy and training performance, but reports no quantitative evidence.

Reference graph

Works this paper leans on

24 extracted references · 18 canonical work pages · cited by 2 Pith papers

  1. [1]

    Combating the challenges of false positives in ai- driven anomaly detection systems and enhancing data security in the cloud,

    O. Olateju, S. U. Okon, U. Igwenagu, A. A. Salami, T. O. Oladoyinbo, and O. O. Olaniyi, “Combating the challenges of false positives in ai- driven anomaly detection systems and enhancing data security in the cloud,” Available at SSRN 4859958 , 2024

    work page 2024

  2. [2]

    Intrusion detection in cloud computing based on time series anomalies utilizing machine learning,

    A.-R. Al-Ghuwairi, Y . Sharrab, D. Al-Fraihat, M. AlElaimat, A. Al- sarhan, and A. Algarni, “Intrusion detection in cloud computing based on time series anomalies utilizing machine learning,” Journal of Cloud Computing, vol. 12, no. 1, p. 127, 2023

    work page 2023

  3. [3]

    A deep intrusion detection system in lambda architecture based on edge cloud computing for iot,

    R. Alghamdi and M. Bellaiche, “A deep intrusion detection system in lambda architecture based on edge cloud computing for iot,” in 2021 4th International conference on artificial intelligence and big data (ICAIBD). IEEE, 2021, pp. 561–566

    work page 2021

  4. [4]

    Enhanced detection of anomalous network behavior in cloud- driven big data systems using deep learning models,

    Y . Lin, “Enhanced detection of anomalous network behavior in cloud- driven big data systems using deep learning models,” Journal of Theory and Practice of Engineering Science , vol. 4, no. 08, pp. 1–11, 2024

    work page 2024

  5. [5]

    An adaptable and unsupervised tinyml anomaly detection system for extreme industrial environments,

    M. Antonini, M. Pincheira, M. Vecchio, and F. Antonelli, “An adaptable and unsupervised tinyml anomaly detection system for extreme industrial environments,” Sensors, vol. 23, no. 4, p. 2344, 2023

    work page 2023

  6. [6]

    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,” in 2024 5th International Conference on Information Science, Parallel and Distributed Systems (ISPDS). IEEE, 2024, pp. 263–267

    2024

  7. [7]

    Towards real-time and personalized code generation,

    H. Xu, X. Wang, and H. Chen, “Towards real-time and personalized code generation,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , 2024, pp. 5568–5569

    work page 2024

  8. [8]

    Harnessing earnings reports for stock predictions: A qlora-enhanced llm approach,

    H. Ni, S. Meng, X. Chen, Z. Zhao, A. Chen, P. Li, S. Zhang, Q. Yin, Y . Wang, and Y . Chan, “Harnessing earnings reports for stock predictions: A qlora-enhanced llm approach,” in 2024 6th International Conference on Data-driven Optimization of Complex Systems (DOCS) . IEEE, 2024, pp. 909–915

    work page 2024

Show all 24 references
  1. [9]

    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,” in 2024 6th International Conference on Electronic Engineering and Informatics (EEI) . IEEE, 2024

    2024

  2. [10]

    Deep learning for cross-border transaction anomaly detection in anti-money laundering systems,

    Q. Yu, Z. Xu, and Z. Ke, “Deep learning for cross-border transaction anomaly detection in anti-money laundering systems,” in 2024 6th International Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI). IEEE, 2024, pp. 244–248

    2024

  3. [11]

    Monilog: An automated log-based anomaly detection system for cloud computing infrastructures,

    A. Vervaet, “Monilog: An automated log-based anomaly detection system for cloud computing infrastructures,” in 2021 IEEE 37th Inter- national Conference on Data Engineering (ICDE) . IEEE, 2021, pp. 2739–2743

    2021

  4. [12]

    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,” in Proceedings of the 23rd Workshop o...

    2024

  5. [13]

    Apply machine learning techniques to detect malicious network traffic in cloud computing,

    A. Alshammari and A. Aldribi, “Apply machine learning techniques to detect malicious network traffic in cloud computing,” Journal of Big Data, vol. 8, no. 1, p. 90, 2021

    2021

  6. [14]

    Ai-enhanced cybersecurity: Machine learn- ing for anomaly detection in cloud computing,

    P. Thapa and T. Arjunan, “Ai-enhanced cybersecurity: Machine learn- ing for anomaly detection in cloud computing,” Quarterly Journal of Emerging Technologies and Innovations, vol. 9, no. 1, pp. 25–37, 2024

    2024

  7. [15]

    Practical autoencoder based anomaly detection by using vector reconstruction error,

    H. Torabi, S. L. Mirtaheri, and S. Greco, “Practical autoencoder based anomaly detection by using vector reconstruction error,” Cybersecurity, vol. 6, no. 1, p. 1, 2023

    2023

  8. [16]

    A pre-trained zero-shot sequential recommendation framework via popularity dynamics,

    J. Wang, P. Rathi, and H. Sundaram, “A pre-trained zero-shot sequential recommendation framework via popularity dynamics,” in Proceedings of the 18th ACM Conference on Recommender Systems, ser. RecSys ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 433–4...

    2024

  9. [17]

    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,” in 2024 5th International Conference on Electronic Communication and Artificial Intelligence (ICECAI). IEEE, 2024, pp. 681–685

    2024

  10. [18]

    Enhancing exchange rate forecasting with explainable deep learning models,

    S. Meng, A. Chen, C. Wang, M. Zheng, F. Wu, X. Chen, H. Ni, and P. Li, “Enhancing exchange rate forecasting with explainable deep learning models,” in 2024 4th International Conference on Electronic Information Engineering and Computer Science (EIECS). IEEE, 2024, pp. 892–896

    2024

  11. [19]

    Beyond localized graph neural networks: An attributed motif regularization framework,

    A. Sankar, J. Wang, A. Krishnan, and H. Sundaram, “Beyond localized graph neural networks: An attributed motif regularization framework,” in 2020 IEEE International Conference on Data Mining (ICDM) , 2020, pp. 472–481

    2020

  12. [20]

    Protocf: Prototypical collaborative filtering for few-shot recommendation,

    ——, “Protocf: Prototypical collaborative filtering for few-shot recommendation,” in Proceedings of the 15th ACM Conference on Recommender Systems , ser. RecSys ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 166–175. [Online]. Available: https://doi.org/...

    2021

  13. [21]

    Cot: an efficient and accurate method for detecting marker genes among many subtypes,

    Y . Lu, C.-T. Wu, S. J. Parker, Z. Cheng, G. Saylor, J. E. Van Eyk, G. Yu, R. Clarke, D. M. Herrington, and Y . Wang, “Cot: an efficient and accurate method for detecting marker genes among many subtypes,” Bioinformatics Advances, vol. 2, no. 1, p. vbac037, 2022

    2022

  14. [22]

    T-GAE: Transferable graph autoencoder for network alignment,

    J. HE, C. Kanatsoulis, and A. Ribeiro, “T-GAE: Transferable graph autoencoder for network alignment,” in The Third Learning on Graphs Conference, 2024. [Online]. Available: https://openreview.net/forum? id=Lm48V5zrzh

    2024

  15. [23]

    Embracing the informative missingness and silent gene in analyzing biologically diverse samples,

    D. Du, S. Bhardwaj, Y . Lu, Y . Wang, S. J. Parker, Z. Zhang, J. E. Van Eyk, G. Yu, R. Clarke, D. M. Herrington et al. , “Embracing the informative missingness and silent gene in analyzing biologically diverse samples,” Scientific reports, vol. 14, no. 1, p. 28265, 2024

    2024

  16. [24]

    Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning,

    J. Liu, F. Shang, Y . Liu, H. Liu, Y . Li, and Y . Gong, “Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning,” in Proceedings of the 32nd ACM International Conference on Multimedia , 2024, pp. 2955–2963

    2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.