Graph Neural Network and Transformer Integration for Unsupervised System Anomaly Discovery

Ming Gong; Nia Qi; Yingnan Deng; Yujun Zou; Yun Zi; Zhihao Xue

REVIEW 4 major objections 6 minor 1 cited by

Graph Neural Network and Transformer Integration for Unsupervised System Anomaly Discovery

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Unsupervised GTF-Net outperforms GDN, Anomaly Transformer, DONUT+, and MO-GAAL on Alibaba cluster traces by fusing graph structure with temporal attention.

desk verdict A sketch of a plausible GCN+Transformer fusion that omits the training objective, so the claimed SOTA results are unverifiable and the paper should be desk rejected. read the letter →

arxiv 2508.09401 v1 pith:AHVBAPT5 submitted 2025-08-13 cs.LG

Yun Zi , Ming Gong , Zhihao Xue , Yujun Zou , Nia Qi , Yingnan Deng This is my paper · ORCID

classification cs.LG

keywords graphneuralnetworkTransformerunsupervisedanomalydetectiondistributedbackendservicedynamicfeaturefusionAlibabaclustertracescore

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 GTF-Net, an unsupervised anomaly detector for distributed backend services that combines graph convolution over service-invocation graphs with a Transformer encoder over per-node metric sequences. The central claim is that a learnable fusion of structural and behavioral representations lets one model catch anomalies that graph-only or sequence-only methods miss, including cascading propagation and cross-node drift. On the Alibaba Ali-CEP cluster trace data, GTF-Net is reported to reach F1 0.889, AUC 0.942, and KS 0.741, above GDN, Anomaly Transformer, DONUT+, and MO-GAAL. The paper presents this as a label-free route to end-to-end anomaly detection in real cloud operations.

What carries the argument

The dynamic service graph $G^{(t)}=(V^{(t)},E^{(t)})$ with adjacency $\tilde{A}=A+I$ and degree matrix $\tilde{D}$; the graph-convolution update $H^{(l+1)}=\sigma(\tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2}H^{(l)}W^{(l)})$; the Transformer encoder $Z_i=\mathrm{TransformerEncoder}(X_i+P)$ with attention $\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(QK^T/\sqrt{d_k})V$; and the learnable fusion $U_i=\gamma H_i+(1-\gamma)Z_i^{\mathrm{final}}$ with MLP scoring $s_i=f_{\mathrm{MLP}}(U_i)$. Together they carry the argument: graph layers capture multi-hop propagation, the Transformer captures long-range temporal dependencies and local mutations, and the fusion weight balances the two before a nonlinear

What would settle it

Re-run GTF-Net on the Alibaba Ali-CEP dataset with the architecture as described and any explicit unsupervised loss (reconstruction error, contrastive deviation, or similar); if no such loss reproduces Table 1's F1 0.889, AUC 0.942, and KS 0.741 with a three-layer Transformer, the central performance claim fails as stated.

Watch

Extended reading notes

Core claim

GTF-Net models the backend system at each time step as a dynamic graph whose nodes are services and whose edges are invocation/dependency relations. Multi-layer graph convolution with self-loops and symmetric degree normalization produces structural node embeddings, while a Transformer with positional encoding and scaled dot-product attention models each node's temporal behavior. The two representations are combined as $U_i=\gamma H_i+(1-\gamma)Z_i^{\mathrm{final}}$ with a learnable $\gamma\in[0,1]$, and an MLP maps $U_i$ to an anomaly score $s_i$. The paper's discovery claim is that this joint structural–temporal representation, trained without labels by measuring deviation in the represent

Load-bearing premise

The load-bearing premise is that an unsupervised training objective for the anomaly-scoring MLP exists; Section III states only that training is unsupervised and that anomalies are identified by deviation in the representation space, without specifying a loss, deviation metric, or optimization procedure.

Editorial extensions

If this is right

If the reported numbers hold, GTF-Net offers a label-free detector for cloud backend systems that tracks anomaly propagation paths across service dependencies instead of flagging only single-metric spikes.
A single learnable fusion weight $\gamma$ makes the model self-adjust the balance between topology and behavior, which could reduce threshold tuning in operations.
The sensitivity analysis suggests three Transformer layers are enough; deeper stacks add cost and dilute attention, so deployment can use a compact encoder.
At 19.6 ms per inference, the method is fast enough to be considered for real-time or near-real-time monitoring in large clusters.
Because both modules are general, the same structure-plus-behavior design transfers to other graph-structured temporal systems, which the paper names as industrial IoT, financial risk, and edge computing.

Reading between the lines

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

The paper keeps $\gamma$ static; a natural extension is a time-dependent or attention-based fusion weight, which could adapt to shifts in whether structure or behavior is the stronger anomaly signal.
Detection quality likely depends on how faithfully the service-invocation graph is constructed; comparing hand-built dependency graphs against automatically inferred graphs from traces would be a testable next step.
The same architecture could be evaluated on other public multivariate time-series benchmarks to see whether the structural component helps outside cloud traces; the paper only reports Ali-CEP.

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

The paper proposes GTF-Net, a graph convolutional network fused with a Transformer encoder to detect anomalies in distributed service systems. The architecture is straightforward: GCN layers propagate structural information, a Transformer encodes each node's time series, a learned scalar gamma fuses the two representations, and an MLP maps the fusion to an anomaly score. The motivation is sensible, and the Alibaba cluster trace is a reasonable testbed. The equations for graph convolution and self-attention are standard, and nothing in the design is wrong per se. It just isn't a complete paper. Section III never defines the training objective. There is no loss function, no reconstruction target, no contrastive objective, and no optimization rule. The sentence about 'degree of deviation in the representation space' is a statement of intent, not an algorithm. Without a loss, the learnable parameters (GCN weights, attention projections, gamma, and the MLP) have no update rule, and the anomaly scores are arbitrary. The comparison in Table 1 therefore has no well-defined system under test. This is a load-bearing omission, not a minor gap. The empirical section is also thin: no hyperparameters, no data preprocessing details, no error bars or repeated runs, and the sensitivity analysis mentioned in the text (Figure 2) is missing entirely. The 'dynamic graph' is just a sequence of static snapshots, with no edge-time modeling, so the temporal aspect is handled only by the Transformer. The references are another red flag. Several citations are clearly unrelated to the topic (organ segmentation, clinical NLP, capsule networks, LLM distillation) and appear to pad the bibliography. That pattern undercuts confidence in the related-work understanding. In short, the central idea is plausible but the paper is an incomplete sketch. It cannot be reproduced, audited, or even implemented as written. The reader's low-confidence reject is spot on; if anything, the confidence could be higher because the omissions are unmistakable. A serious editor should desk reject this rather than spend referee hours on a manuscript that fails to specify its own training procedure. If the authors supply the missing objective, full experimental details, and clean up the references, it might be worth a workshop submission, but in its current form it is not ready for peer review.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes GTF-Net, an unsupervised anomaly detection method for distributed backend services that combines a GCN-based structural encoder with a Transformer-based temporal encoder, fuses the two representations via a learnable convex combination, and maps the fused embedding to an anomaly score through an MLP. The authors report experiments on Alibaba Cluster Trace data, claiming F1 0.889, AUC 0.942, and KS 0.741, outperforming GDN, Anomaly Transformer, DONUT+, and MO-GAAL, together with a sensitivity analysis of Transformer depth. The central problem is that Section III never defines a training objective, loss function, deviation metric, or optimization procedure; consequently the model as described is not trainable, and the reported metrics cannot be reproduced or audited.

Significance. If the reported results were adequately supported, the paper would present a straightforward integration of structural and temporal modeling for unsupervised system anomaly detection on a public cloud benchmark. The choice of metrics (F1, AUC, KS) and the use of the Alibaba trace are appropriate. However, as written, the scientific contribution is not assessable: the architecture equations are standard textbook formulas, the unsupervised learning mechanism is unspecified, and the experimental protocol is incomplete. No code, hyperparameters, error bars, or data splits are provided. The paper therefore offers no falsifiable, reproducible evidence for its central claim. On the positive side, the paper addresses a practical problem and names relevant baseline families, but this does not compensate for the missing specification of the system under test.

major comments (4)

[Section III, Eq. (5)] No training objective is defined. Equations (1)-(5) specify graph convolution, attention, fusion U_i = gamma H_i + (1-gamma) Z_i, and score s_i = MLP(U_i), but there is no loss function, reconstruction target, contrastive objective, deviation metric, or parameter update rule. The sentence 'trained under unsupervised conditions and identifies abnormal behaviors based on the degree of deviation in the representation space' is a statement of intent, not an algorithm. Without an objective, W^(l), attention projections, gamma, and the MLP weights have no update rule, so the anomaly score is arbitrary and Table 1 cannot be reproduced. This is the primary gap and must be fixed for any evaluation to be meaningful.
[Section IV-B, Table 1] The experimental protocol is incomplete. There is no data split (train/validation/test), no preprocessing or normalization, no description of how the dynamic graph is constructed (window size, edge thresholds), no hyperparameter values (GCN layers, Transformer layers/dimensions/heads, MLP structure, learning rate, batch size, epochs), no threshold selection procedure for F1/KS, and no variance/error bars across runs. The time cost of 19.6 ms is not interpretable without hardware and batch-size details. These omissions make the numbers in Table 1 unverifiable and prevent comparison with baselines.
[Section IV-A] The evaluation uses weak labels without explaining how. The text states that 'a small number of system alerts and scheduling anomalies' are used as 'weak labels for auxiliary evaluation' but then says the core detection task remains unsupervised. It is not specified which labels define the ground truth for F1, AUC, and KS, how the anomaly-score threshold is chosen, or whether these weak labels influence model selection or training. This ambiguity obscures the link between the unsupervised score and the reported metrics.
[Section III] The architecture is described only as a sequence of standard formulas: symmetric normalized adjacency with self-loops (Eq. (1)), scaled dot-product attention (Eq. (3)), convex fusion (Eq. (4)), and an unspecified MLP. The number of GCN layers, hidden dimensions, activation functions, positional encoding scheme, sequence length, fusion-weight initialization, and MLP architecture are all omitted. Without these details, the method cannot be implemented from the text, and it is unclear how GTF-Net differs from existing GNN-Transformer anomaly detectors.

minor comments (6)

[Section IV-B, Figure 2] Figure 2 is referenced in the text but not present in the manuscript; the sensitivity-analysis results are described only in prose.
[References] Citation mismatches: Table 1 labels 'GDN [19]' but Reference [19] is a log-anomaly paper, not the Graph Deviation Network; 'Anomaly Transformer [20]' points to MemTo, not the Anomaly Transformer (which is Reference [5]).
[References] Several citations appear unrelated to the technical content, e.g., medical image segmentation [2], clinical NLP [3], and LLM distillation [16]. The paper should cite the actual GNN/Transformer anomaly-detection literature and verify each reference.
[Section IV-A] The dataset is referred to as 'Ali-CEP'; the standard name is Alibaba Cluster Trace Program. The exact version and how to access it should be provided.
[Section IV] The 'Conclusion' section heading is missing; the text runs directly from Section IV into the conclusion paragraph.
[Section III, Eq. (2)] Equation (2) is garbled: 'PXrEncoderTransformeZ' should read 'Z_i = TransformerEncoder(X_i + P)'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the missing training objective is an under-specification/verifiability gap, not a circular derivation.

full rationale

Walked the claimed derivation chain: dynamic graph snapshot, GCN updates (Eq. 1), Transformer positional encoding and attention (Eqs. 2-3), fusion U_i = gamma H_i + (1-gamma) Z_i, and anomaly score s_i = MLP(U_i). None of these equations defines the evaluation metrics or is fitted to the reported F1/AUC/KS values. The fusion weight gamma is a learnable parameter, not a fit to test labels; the MLP score is an output of the model, not constructed from the benchmark. No equation reduces to its own input. No uniqueness theorem or prior result by the present authors is invoked to force the architecture, and the related-work citations are contextual rather than load-bearing. The paper's central weakness is that Section III states the method is 'trained under unsupervised conditions and identifies abnormal behaviors based on the degree of deviation in the representation space' but never defines a loss function, deviation metric, or optimization procedure. That is a reproducibility and completeness failure, not a circularity: an underspecified algorithm cannot be audited, but the paper also does not define its predictions as algebraic consequences of the evaluation data. Thus no circular step can be quoted, and the circularity score is 0.

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

The central claim rests on standard deep learning building blocks, but the training objective and evaluation setup are unstated. The only explicit learnable quantity is the fusion weight gamma, with all other parameters left unspecified. No new physical or conceptual entities are introduced.

free parameters (3)

Fusion weight gamma = unknown
Learned scalar balancing structural and temporal embeddings; no value or update rule given in paper.
GCN weight matrices W^(l) = unknown
Standard learned weights for graph convolution, no initialization or training details.
Transformer and MLP parameters = unknown
All network weights are unspecified; no hyperparameters (layers, heads, dimensions) are reported.

assumptions (3)

domain assumption The service system can be represented as a snapshot graph with invocation edges that adequately captures anomaly propagation
Section III defines G(t) as a graph snapshot; assumes invocation topology is sufficient for detecting anomalies, but no validation of this assumption.
ad hoc to paper Unsupervised anomaly detection can be achieved by measuring deviation in the fused representation space
Section III states the model is trained under unsupervised conditions and identifies abnormalities via representation deviation, but no precise deviation metric or training objective is given.
domain assumption The Ali-CEP dataset contains reliable anomaly labels or weak labels for evaluation
Section IV-A claims weak labels exist but gives no provenance or labeling process.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Graph Neural Network and Transformer Integration for Unsupervised System Anomaly Discovery." pith.science (2026). https://pith.science/paper/AHVBAPT5

@misc{pith2026250809401,
  author       = {Pith},
  title        = {Pith review of: Graph Neural Network and Transformer Integration for Unsupervised System Anomaly Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHVBAPT5}},
  note         = {Machine review of arXiv:2508.09401}
}

read the original abstract

This study proposes an unsupervised anomaly detection method for distributed backend service systems, addressing practical challenges such as complex structural dependencies, diverse behavioral evolution, and the absence of labeled data. The method constructs a dynamic graph based on service invocation relationships and applies graph convolution to extract high-order structural representations from multi-hop topologies. A Transformer is used to model the temporal behavior of each node, capturing long-term dependencies and local fluctuations. During the feature fusion stage, a learnable joint embedding mechanism integrates structural and behavioral representations into a unified anomaly vector. A nonlinear mapping is then applied to compute anomaly scores, enabling an end-to-end detection process without supervision. Experiments on real-world cloud monitoring data include sensitivity analyses across different graph depths, sequence lengths, and data perturbations. Results show that the proposed method outperforms existing models on several key metrics, demonstrating stronger expressiveness and stability in capturing anomaly propagation paths and modeling dynamic behavior sequences, with high potential for practical deployment.

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 · 16 canonical work pages · cited by 1 Pith paper

[1]

Graph neural network-based anomaly detection in multivariate time series,

A. Deng and B. Hooi, "Graph neural network-based anomaly detection in multivariate time series," Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 5, pp. 4027-4035, 2021

work page 2021
[2]

Domain-adaptive organ segmentation through SegFormer architecture in clinical imaging,

X. Zhang and X. Wang, "Domain-adaptive organ segmentation through SegFormer architecture in clinical imaging," Transactions on Computational and Scientific Methods, vol. 5, no. 7, 2025. doi: https://doi.org/10.5281/zenodo.16513017

work page doi:10.5281/zenodo.16513017 2025
[3]

Clinical NLP with attention-based deep learning for multi-disease prediction,

T. Xu, X. Deng, X. Meng, H. Yang, and Y. Wu, "Clinical NLP with attention-based deep learning for multi-disease prediction," arXiv preprint arXiv:2507.01437, 2025

arXiv 2025
[4]

Capsule Network-Based Semantic Intent Modeling for Human-Computer Interaction

S. Wang, Y. Zhuang, R. Zhang, and Z. Song, "Capsule network-based semantic intent modeling for human-computer interaction," arXiv preprint arXiv:2507.00540, 2025

work page Pith review arXiv 2025
[5]

Anomaly transformer: Time series anomaly detection with association discrepancy,

J. Xu, H. Wu, J. Wang, et al., "Anomaly transformer: Time series anomaly detection with association discrepancy," arXiv preprint arXiv:2110.02642, 2021

arXiv 2021
[6]

Causal discriminative modeling for robust cloud service fault detection,

H. Wang, “Causal discriminative modeling for robust cloud service fault detection,” Journal of Computer Technology and Software, vol. 3, no. 7, 2024.doi: https://doi.org/10.5281/zenodo.15851629

work page doi:10.5281/zenodo.15851629 2024
[7]

Behavioral anomaly detection in distributed systems via federated contrastive learning,

R. Meng, H. Wang, Y. Sun, Q. Wu, L. Lian and R. Zhang, “Behavioral anomaly detection in distributed systems via federated contrastive learning,” arXiv preprint arXiv:2506.19246, 2025

arXiv 2025
[8]

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.doi: https://doi.org/10.5281/zenodo.15559874

work page doi:10.5281/zenodo.15559874 2025

Show all 22 references

[9]

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. doi: https://doi.org/10.5281/zenodo.15661629

2024 doi
[10]

Modeling audit workflow dynamics with deep Q- learning for intelligent decision-making,

Z. Liu and Z. Zhang, “Modeling audit workflow dynamics with deep Q- learning for intelligent decision-making,” Transactions on Computational and Scientific Methods, vol. 4, no. 12, 2024.doi: https://doi.org/10.5281/zenodo.15725060

2024 doi
[11]

Selective noise injection and feature scoring for unsupervised request anomaly detection,

Y. Cheng, “Selective noise injection and feature scoring for unsupervised request anomaly detection,” Journal of Computer Technology and Software, vol. 3, no. 9, 2024.doi: https://doi.org/10.5281/zenodo.15735364

2024 doi
[12]

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
[13]

Deep learning for root cause detection in distributed systems with structural encoding and multi-modal attention,

Y. Ren, “Deep learning for root cause detection in distributed systems with structural encoding and multi-modal attention,” Journal of Computer Technology and Software, vol. 3, no. 5, 2024.doi: https://doi.org/10.5281/zenodo.15851611

2024 doi
[14]

Collaborative multi-agent reinforcement learning approach for elastic cloud resource scaling,

B. Fang and D. Gao, “Collaborative multi-agent reinforcement learning approach for elastic cloud resource scaling,” arXiv preprint arXiv:2507.00550, 2025

2025 arXiv
[15]

Fusion- based retrieval-augmented generation for complex question answering with LLMs,

Y. Sun, R. Zhang, R. Meng, L. Lian, H. Wang and X. Quan, “Fusion- based retrieval-augmented generation for complex question answering with LLMs,” 2025. doi: 10.20944/preprints202507.0826.v1

2025
[16]

Collaborative distillation strategies for parameter-efficient language model deployment,

X. Meng, Y. Wu, Y. Tian, X. Hu, T. Kang and J. Du, “Collaborative distillation strategies for parameter-efficient language model deployment,” arXiv preprint arXiv:2507.15198, 2025

2025 arXiv
[17]

Multivariate time-series anomaly detection via graph attention network,

H. Zhao, Y. Wang, J. Duan, et al., "Multivariate time-series anomaly detection via graph attention network," Proceedings of the 2020 IEEE International Conference on Data Mining, pp. 841-850, 2020

2020
[18]

TranAD: Deep transformer networks for anomaly detection in multivariate time series data,

S. Tuli, G. Casale, and N. R. Jennings, "TranAD: Deep transformer networks for anomaly detection in multivariate time series data," arXiv preprint arXiv:2201.07284, 2022

2022 arXiv
[19]

Graph neural networks based log anomaly detection and explanation,

Z. Li, J. Shi, and M. Van Leeuwen, "Graph neural networks based log anomaly detection and explanation," Proceedings of the 2024 IEEE/ACM 46th International Conference on Software Engineering: Companion Proceedings, pp. 306-307, 2024

2024
[20]

Memto: Memory-guided transformer for multivariate time series anomaly detection,

J. Song, K. Kim, J. Oh, et al., "Memto: Memory-guided transformer for multivariate time series anomaly detection," Advances in Neural Information Processing Systems, vol. 36, pp. 57947-57963, 2023

2023
[21]

Time-series anomaly detection service at Microsoft,

H. Ren, B. Xu, Y. Wang, et al., "Time-series anomaly detection service at Microsoft," Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 3009-3017, 2019

2019
[22]

Generative adversarial active learning for unsupervised outlier detection,

Y. Liu, Z. Li, C. Zhou, et al., "Generative adversarial active learning for unsupervised outlier detection," IEEE Transactions on Knowledge and Data Engineering, vol. 32, no. 8, pp. 1517-1528, 2019

2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Graph neural network-based anomaly detection in multivariate time series,

A. Deng and B. Hooi, "Graph neural network-based anomaly detection in multivariate time series," Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 5, pp. 4027-4035, 2021

work page 2021

[2] [2]

Domain-adaptive organ segmentation through SegFormer architecture in clinical imaging,

X. Zhang and X. Wang, "Domain-adaptive organ segmentation through SegFormer architecture in clinical imaging," Transactions on Computational and Scientific Methods, vol. 5, no. 7, 2025. doi: https://doi.org/10.5281/zenodo.16513017

work page doi:10.5281/zenodo.16513017 2025

[3] [3]

Clinical NLP with attention-based deep learning for multi-disease prediction,

T. Xu, X. Deng, X. Meng, H. Yang, and Y. Wu, "Clinical NLP with attention-based deep learning for multi-disease prediction," arXiv preprint arXiv:2507.01437, 2025

arXiv 2025

[4] [4]

Capsule Network-Based Semantic Intent Modeling for Human-Computer Interaction

S. Wang, Y. Zhuang, R. Zhang, and Z. Song, "Capsule network-based semantic intent modeling for human-computer interaction," arXiv preprint arXiv:2507.00540, 2025

work page Pith review arXiv 2025

[5] [5]

Anomaly transformer: Time series anomaly detection with association discrepancy,

J. Xu, H. Wu, J. Wang, et al., "Anomaly transformer: Time series anomaly detection with association discrepancy," arXiv preprint arXiv:2110.02642, 2021

arXiv 2021

[6] [6]

Causal discriminative modeling for robust cloud service fault detection,

H. Wang, “Causal discriminative modeling for robust cloud service fault detection,” Journal of Computer Technology and Software, vol. 3, no. 7, 2024.doi: https://doi.org/10.5281/zenodo.15851629

work page doi:10.5281/zenodo.15851629 2024

[7] [7]

Behavioral anomaly detection in distributed systems via federated contrastive learning,

R. Meng, H. Wang, Y. Sun, Q. Wu, L. Lian and R. Zhang, “Behavioral anomaly detection in distributed systems via federated contrastive learning,” arXiv preprint arXiv:2506.19246, 2025

arXiv 2025

[8] [8]

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.doi: https://doi.org/10.5281/zenodo.15559874

work page doi:10.5281/zenodo.15559874 2025