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

Chang Liu; Heyi Wang; Qiyuan Wu; Wenxuan Zhu; Xuehui Quan; Yang Wang

REVIEW 4 major objections 5 minor 5 cited by

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

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 a GRU plus attention and feedforward classifier predicts distributed-system faults from telemetry, beating four transformer baselines on accuracy, F1, and AUC.

desk verdict A standard GRU+attention stack on Azure telemetry; the evaluation protocol is too thin to trust the headline numbers. read the letter →

arxiv 2505.20705 v1 pith:FTWKPRO7 submitted 2025-05-27 cs.DC cs.LG

Yang Wang , Wenxuan Zhu , Xuehui Quan , Heyi Wang , Chang Liu , Qiyuan Wu This is my paper · ORCID

classification cs.DCcs.LG

keywords DistributedsystemsFaultpredictionTimeseriesmodelingAttentionmechanismGatedrecurrentunitTelemetrydataDeeplearning

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 aims to show that a relatively lightweight sequence model — a gated recurrent unit, a learned attention layer, and a feedforward classifier — can predict impending faults in a large distributed system from five-minute performance telemetry. The authors report Accuracy 95.2%, F1-Score 93.7%, and AUC 94.1% on a public cloud-telemetry dataset, beating four transformer-based baselines on every metric. Their ablation shows that adding attention to the GRU raises accuracy from 91.8% to 95.2% and F1 from 89.5% to 93.7%, which they read as evidence that focusing on key time segments drives the improvement. A stable loss decline that flattens after about ten epochs is offered as evidence of dependable convergence. If true, this would give operators a practical, proactive alternative to reactive alerting.

What carries the argument

The carrying mechanism is the attention-weighted context vector $s = \sum_{t=1}^{T}\alpha_t h_t$ built from GRU hidden states. The GRU update $h_t = \mathrm{GRU}(x_t, h_{t-1})$ encodes the evolving system state from the multivariate telemetry window; the attention weights $\alpha_t = \frac{\exp(q^{T}W_a h_t)}{\sum_{i}\exp(q^{T}W_a h_i)}$ select the time steps most indicative of an emerging fault; and a feedforward network maps $s$ to a binary fault decision with a weighted cross-entropy loss to counter class imbalance. This three-stage pipeline — temporal encoding, selective aggregation, classification — is what the paper credits for the reported gains, with the ablation attributing most of the improvement to the attention stage.

What would settle it

Recompute the same model and baselines with a strict chronological split — train on the first part of the month, test on a later non-overlapping period, and exclude any node seen in training from the test set. If the gaps over FEDformer shrink substantially or the absolute metrics drop, the reported superiority is an artifact of temporal overlap; if the numbers hold, the claim is confirmed.

Watch

Extended reading notes

Core claim

The paper's central claim is that temporal modeling with attention is enough to outperform large transformer-based time-series models on a real fault-prediction task. The model reads a multivariate window of CPU, memory, disk, and network metrics, compresses the sequence through a GRU, re-weights the hidden states by learned attention, and feeds the weighted context into a feedforward network that makes a binary fault call. On the cloud telemetry dataset, the authors report ACC 95.2%, F1 93.7%, and AUC 94.1%, compared with FEDformer's ACC 93.3%, F1 91.6%, and AUC 90.9% as the strongest baseline. The attention ablation, which lifts accuracy from 91.8% to 95.2% and F1 from 89.5% to 93.7%, is presented as evidence that the attention component is the principal source of the gain.

Load-bearing premise

The reported numbers are meaningful only if the evaluation windows are separated in time and by node from the training windows; the paper never specifies the train/test split, so autocorrelated five-minute telemetry could appear on both sides and inflate every metric.

Editorial extensions

If this is right

A lightweight GRU plus attention can beat four transformer baselines on cloud fault prediction, so heavy self-attention stacks are not necessary for this task.
Attention is the main driver of the gain: adding it raises accuracy from 91.8% to 95.2% and F1 from 89.5% to 93.7%.
The loss curves converge smoothly and plateau around ten epochs, suggesting the model can be trained stably on a realistic telemetry workload.
High accuracy together with high AUC means the model both classifies correctly at a decision threshold and ranks fault candidates well, which supports alert triage and prioritization.

Reading between the lines

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

The paper does not test prediction horizons, so a natural extension is training the same architecture to flag faults 15, 30, and 60 minutes ahead; that would reveal whether attention is capturing early precursors or only near-failure signatures.
Because the architecture is much lighter than transformer baselines, it is plausible it could run at the edge or per node, though the paper does not measure inference cost or memory footprint.
A matched-capacity GRU without attention would isolate whether the gain comes from the attention mechanism itself or simply from the extra parameters the module adds.

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

Read this one if you want a benchmark data point, not if you want new machinery. The model is GRU plus softmax attention plus a feedforward classifier, every component standard. The contribution is an evaluation on the Azure telemetry dataset, and the claim is that this simple stack beats FEDformer, Autoformer, Informer, and Transformer on ACC/F1/AUC.

What the paper does well: it is honest in scope, uses a real public dataset, and the ablation of the attention module is a reasonable check. There is no circularity; the comparison is against external baselines. The loss curve is not evidence of much, but it is harmless.

The soft spot is the evaluation protocol. Section III.A never says how the data is split. Train/test could be temporal, random by window, or disjoint by node; the paper does not say. With five-minute telemetry from thousands of nodes, consecutive windows from the same node are strongly autocorrelated. If training and test windows overlap or share nodes, the reported margins—largest gain over FEDformer is 1.9 ACC, 2.1 F1, 3.2 AUC—could evaporate. There are no error bars, no repeated runs, no hyperparameter settings, and no description of how the forecasting baselines were adapted to binary classification. Without these, Table 1 is not interpretable as a comparison. The references are also sloppy: 'Transformer' is cited as a power-system CNN paper, not the original architecture.

None of this is fatal. The architecture is sound, the experiment is real, and the missing pieces are reportable rather than conceptual. I would not send this to a serious referee as is; I would ask the authors for a clean split description, uncertainty estimates, and either code or a data link first. For a reading group, it is a useful example of how much evaluation detail matters in time-series ML. I would cite it only if I needed a quick instance of GRU+attention for fault prediction, and even then I would check the numbers first.

Referee Report

4 major / 5 minor

Summary. The paper proposes a fault-prediction method for distributed systems that combines a Gated Recurrent Unit (GRU), a temporal attention mechanism, and a feedforward classification network. The input is multivariate performance telemetry, and the output is a binary early-warning signal for system failures. The method is evaluated on the Microsoft Azure 'Telemetry Data from a Large-Scale Cloud System' dataset. The authors report that their model achieves Accuracy 95.2%, F1-Score 93.7%, and AUC 94.1%, outperforming Transformer, Informer, Autoformer, and FEDformer baselines on all metrics. An ablation analyzing the contribution of the attention module and a loss-convergence plot are also presented. The central claim is that this lightweight temporal architecture surpasses mainstream time-series models for fault prediction.

Significance. If the experimental claim were fully supported, the paper would provide a useful data point: a relatively simple GRU+attention model beating recent Transformer-based forecasting architectures on a real-world cloud-telemetry fault-prediction task. The use of a public Azure dataset, the handling of class imbalance via weighted cross-entropy, and the explicit attention-mechanism ablation are strengths. There is no circularity in the method: the model is trained on labeled telemetry and compared against independent baselines. However, the current evidence base is missing key experimental-protocol details (train/test split, hyperparameters, repeated runs, baseline adaptations), so the reported performance gains are not yet established. The significance of the contribution is therefore conditional on a major revision that supplies these details.

major comments (4)

[III.A and Table 1] The manuscript never specifies how the dataset is split into training, validation, and test sets. It does not state whether the split is temporal (train on an early period, test on a later period), random by window, or disjoint by node identifier. Because the data are 5-minute-interval telemetry from thousands of nodes over a month, consecutive windows from the same node are strongly autocorrelated. Without a time-based or node-disjoint split, the reported metrics (ACC 95.2, F1 93.7, AUC 94.1) can be inflated by temporal leakage, making the comparison in Table 1 uninterpretable. This is load-bearing for the central claim, and the omission must be fixed by specifying the exact split protocol.
[Table 1 and Section III.B] Table 1 reports single-point metrics with no error bars, confidence intervals, or number of independent runs. The margins over the best baseline, FEDformer, are small (1.9 ACC, 2.1 F1, 3.2 AUC), and no statistical significance test is provided. Additionally, Section III.B does not report the hyperparameters (learning rate, number of epochs, sequence window length T, GRU hidden dimension, attention query dimension, class weights, optimizer), which prevents reproducibility and makes it impossible to judge whether the baselines were tuned fairly. These omissions directly affect the validity of the claimed outperformance.
[Section III.B] The four baselines are listed with citations but the paper does not describe how they were adapted to binary fault classification. FEDformer, Informer, Autoformer, and the cited Transformer are forecasting architectures; it is not stated what classification head, loss function, or output decoding they used, nor whether they were retrained from scratch or fine-tuned. Without this information, the reader cannot rule out an unfair comparison, and the claim of outperforming 'mainstream time-series models' is not substantiated.
[Figure 2 and ablation text] The attention ablation reports Accuracy increasing from 91.8% to 95.2% and F1-Score from 89.5% to 93.7%, but the text never states the AUC values for the model without attention, despite asserting that the AUC also improved. The AUC improvement is therefore unsupported by any reported number. Please provide the full set of metrics for both conditions.

minor comments (5)

[Section II] The mathematical notation is garbled in several places (e.g., the GRU update equation and the attention-weight equation), making the architecture description difficult to parse. Please typeset all equations clearly.
[Section III.A] The dataset description says 'part of the data is annotated with known fault events' but does not define the prediction horizon (how far in advance a fault must be predicted) or the class ratio after windowing. These details affect both the task definition and the interpretation of the metrics.
[Figure 3] The loss-convergence plot is described as showing training and validation loss, but the axes and the epoch scale are not described in the text; the claim of a plateau after approximately 10 epochs cannot be verified without axis labels.
[References [19]-[21]] The citations for Transformer, Informer, and Autoformer do not appear to point to the canonical time-series forecasting papers: [19] is a CNN-based transformer for power-system fault detection, [20] is a semi-supervised Informer for robot fault diagnosis, and [21] combines contrastive learning with Autoformer. Please either cite the original architectures or clarify the exact baseline implementations used.
[Section III.B] The sentence 'Compared to baseline Transformer models, the proposed approach enhances accuracy by approximately 4 percentage points' uses 95.2 - 91.4 = 3.8; consider saying 'approximately 4' or giving the exact difference.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the GRU+attention+FFN classifier is trained and evaluated on an external benchmark; Table 1 reports independent outcomes, not definitions or fitted predictions.

full rationale

The paper contains no derivation chain whose conclusion is presupposed by its premises. The proposed model is a supervised classifier trained on labeled Azure telemetry, and the reported Accuracy, F1-Score, and AUC values in Table 1 are outcomes of an external benchmark comparison against Transformer, Informer, Autoformer, and FEDformer; they are not used to define the model, set its parameters, or construct its inputs. The architectural components (GRU, attention, feedforward classifier, weighted cross-entropy loss) are presented with standard equations and cited as general techniques, not as results established by the present authors. No cited work shares the current authors, so no self-citation is load-bearing. The absence of a documented train/test split, lack of error bars, and unclear baseline adaptation are experimental-validity concerns about whether the benchmark is trustworthy, but they are not instances of circular reasoning: the comparison remains an empirical measurement rather than an analytic identity or a fitted input renamed as a prediction. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The model is a standard supervised classifier: its learned weights come from training on the dataset, which is normal practice and not circular. The items above are the manually chosen, unreported choices and domain assumptions that the reported performance depends on. No new physical or conceptual entities are introduced.

free parameters (6)

Sequence window length T
Input windows are constructed from telemetry sequences in Section III.A, but the window length is not reported; all metrics depend on it.
GRU hidden state dimension
The number of GRU units is chosen by hand and not reported; it controls model capacity and the reported accuracy.
Attention query vector q and weight matrix W_a
Learnable parameters in the attention formula in Section II, fitted to training data, with no initialization or regularization described.
Weighted cross-entropy class weights
Section II says weighted cross-entropy handles class imbalance, but the weights are not given.
Learning rate and optimizer
Training details are absent, including the optimizer, learning rate, batch size, and stopping criterion.
Number of training epochs = ~10 (loss plateau)
Section III.B says the loss plateaus after approximately 10 epochs, but the exact epoch count and convergence threshold are not stated.

assumptions (5)

standard math GRU state update h_t = GRU(h_{t-1}, x_t) is a standard architecture that captures temporal dependencies.
Invoked in Section II without proof; correctness is assumed from prior published work on gated recurrent units.
standard math Softmax attention with a learnable query vector produces a valid weighted context representation.
Used in Section II for weighted fusion of hidden states; the standard attention formulation is assumed from the literature.
domain assumption The Microsoft Azure telemetry dataset annotations correctly identify true fault events.
Section III.A relies on dataset labels for supervised training and evaluation; noisy or inconsistent labels would change all metrics.
domain assumption The evaluation split is non-overlapping and temporally independent.
Section III.B reports results without describing the train/test split; without independence, temporal autocorrelation can inflate metrics.
domain assumption Linear interpolation of missing values preserves the temporal patterns needed for fault prediction.
Section III.A states missing values are linearly interpolated; long gaps or frequent missing values could introduce artifacts.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Time-Series Learning for Proactive Fault Prediction in Distributed Systems with Deep Neural Structures." pith.science (2026). https://pith.science/paper/FTWKPRO7

@misc{pith2026250520705,
  author       = {Pith},
  title        = {Pith review of: Time-Series Learning for Proactive Fault Prediction in Distributed Systems with Deep Neural Structures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTWKPRO7}},
  note         = {Machine review of arXiv:2505.20705}
}

read the original abstract

This paper addresses the challenges of fault prediction and delayed response in distributed systems by proposing an intelligent prediction method based on temporal feature learning. The method takes multi-dimensional performance metric sequences as input. We use a Gated Recurrent Unit (GRU) to model the evolution of system states over time. An attention mechanism is then applied to enhance key temporal segments, improving the model's ability to identify potential faults. On this basis, a feedforward neural network is designed to perform the final classification, enabling early warning of system failures. To validate the effectiveness of the proposed approach, comparative experiments and ablation analyses were conducted using data from a large-scale real-world cloud system. The experimental results show that the model outperforms various mainstream time-series models in terms of Accuracy, F1-Score, and AUC. This demonstrates strong prediction capability and stability. Furthermore, the loss function curve confirms the convergence and reliability of the training process. It indicates that the proposed method effectively learns system behavior patterns and achieves efficient fault detection.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

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

Graph Neural Network and Transformer Integration for Unsupervised System Anomaly Discovery
cs.LG 2025-08 reject novelty 3.0 of 10

GTF-Net, a GCN-Transformer fusion, reports F1 0.889 and AUC 0.942 for anomaly detection on Alibaba cluster data, but omits the training objective and code.
Domain-Adversarial Transfer Learning for Fault Root Cause Identification in Cloud Computing Systems
cs.DC 2025-07 reject novelty 3.0 of 10

The paper claims a transfer-learning model with adversarial alignment and pseudo-labels outperforms four existing methods for cloud fault root cause identification, but the experiments are not reproducible.
Collaborative Multi-Agent Reinforcement Learning Approach for Elastic Cloud Resource Scaling
cs.DC 2025-07 reject novelty 3.0 of 10

A coordinated multi-agent autoscaling scheme with workload prediction is claimed to outperform prior controllers, but the method and evaluation are underspecified to the point that the claim cannot be verified.
Multi-Level Service Performance Forecasting via Spatiotemporal Graph Neural Networks
cs.LG 2025-08 reject novelty 2.0 of 10

A GCN-plus-GRU spatiotemporal forecasting model is proposed for service performance, claiming SOTA on Alibaba Cluster Trace 2018, but the novelty is minimal and the experimental reporting is insufficient.
Autonomous Resource Management in Microservice Systems via Reinforcement Learning
cs.DC 2025-07 reject novelty 1.0 of 10

A generic reinforcement learning scheduler for microservices is claimed to improve response time, throughput, and resource utilization, but the evidence is an unverifiable table with no error bars.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages · cited by 5 Pith papers

[1]

Fault tolerance in distributed systems using deep learning approaches,

B. Assiri and A. Sheneamer, "Fault tolerance in distributed systems using deep learning approaches," PLOS One, vol. 20, no. 1, 2025

work page 2025
[2]

Modeling multi- hop semantic paths for recommendation in heterogeneous information networks,

H. Zheng, Y. Xing, L. Zhu, X. Han, J. Du and W. Cui, "Modeling multi- hop semantic paths for recommendation in heterogeneous information networks," arXiv preprint arXiv:2505.05989, 2025

arXiv 2025
[3]

Analyzing data augmentation techniques for contrastive learning in recommender models,

M. Wei, H. Xin, Y. Qi, Y. Xing, Y. Ren and T. Yang, "Analyzing data augmentation techniques for contrastive learning in recommender models," 2025

work page 2025
[4]

Structured preference modeling for reinforcement learning-based fine-tuning of large models,

L. Zhu, F. Guo, G. Cai and Y. Ma, "Structured preference modeling for reinforcement learning-based fine-tuning of large models," Journal of Computer Technology and Software, vol. 4, no. 4, 2025

work page 2025
[5]

Context- guided dynamic retrieval for improving generation quality in RAG models,

J. He, G. Liu, B. Zhu, H. Zhang, H. Zheng and X. Wang, "Context- guided dynamic retrieval for improving generation quality in RAG models," arXiv preprint arXiv:2504.19436, 2025

arXiv 2025
[6]

Joint semantic detection and dissemination control of phishing attacks on social media via LLama-based modeling,

R. Wang, "Joint semantic detection and dissemination control of phishing attacks on social media via LLama-based modeling," 2025

work page 2025
[7]

Deep learning-based fault prediction in wireless sensor network embedded cyber-physical systems for industrial processes,

H. Ruan, Y. Hu, C. Zhang, J. Shi and J. Zhang, "Deep learning-based fault prediction in wireless sensor network embedded cyber-physical systems for industrial processes," IEEE Access, vol. 10, pp. 10867– 10879, 2022

work page 2022
[8]

Industrial artificial intelligence based energy management system: Integrated framework for electricity load forecasting and fault prediction,

Y. Hu, Q. Li, Y. Ruan, J. Zhang and X. Wang, "Industrial artificial intelligence based energy management system: Integrated framework for electricity load forecasting and fault prediction," Energy, vol. 244, pp. 123195, 2022

work page 2022

Show all 22 references

[9]

Multivariate time series forecasting through automated feature extraction and transformer-based modeling,

Y. Cheng, "Multivariate time series forecasting through automated feature extraction and transformer-based modeling," Journal of Computer Science and Software Applications, vol. 5, no. 5, 2025

2025
[10]

Federated learning for cross-domain data privacy: A distributed approach to secure collaboration,

Y. Zhang, J. Liu, J. Wang, L. Dai, F. Guo and G. Cai, "Federated learning for cross-domain data privacy: A distributed approach to secure collaboration," arXiv preprint arXiv:2504.00282, 2025

2025 arXiv
[11]

Topology-aware decision making in distributed scheduling via multi-agent reinforcement learning,

B. Wang, "Topology-aware decision making in distributed scheduling via multi-agent reinforcement learning," Transactions on Computational and Scientific Methods, vol. 5, no. 4, 2025

2025
[12]

Towards robust few- shot text classification using transformer architectures and dual loss strategies,

X. Han, Y. Sun, W. Huang, H. Zheng and J. Du, "Towards robust few- shot text classification using transformer architectures and dual loss strategies," arXiv preprint arXiv:2505.06145, 2025

2025 arXiv
[13]

Diffusion-transformer framework for deep mining of high-dimensional sparse data,

W. Cui and A. Liang, "Diffusion-transformer framework for deep mining of high-dimensional sparse data," Journal of Computer Technology and Software, vol. 4, no. 4, 2025

2025
[14]

Dynamic operating system scheduling using double DQN: A reinforcement learning approach to task optimization,

X. Sun, Y. Duan, Y. Deng, F. Guo, G. Cai and Y. Peng, "Dynamic operating system scheduling using double DQN: A reinforcement learning approach to task optimization," arXiv preprint arXiv:2503.23659, 2025

2025 arXiv
[15]

Distributed network traffic scheduling via trust-constrained policy learning mechanisms,

Y. Ren, M. Wei, H. Xin, T. Yang and Y. Qi, "Distributed network traffic scheduling via trust-constrained policy learning mechanisms," Transactions on Computational and Scientific Methods, vol. 5, no. 4, 2025

2025
[16]

A reinforcement learning approach to traffic scheduling in complex data center topologies,

Y. Deng, "A reinforcement learning approach to traffic scheduling in complex data center topologies," Journal of Computer Technology and Software, vol. 4, no. 3, 2025

2025
[17]

Fault detection, isolation and service restoration in modern power distribution systems: A review,

I. Srivastava, H. Patel, K. Ghosh, S. Mishra and R. Verma, "Fault detection, isolation and service restoration in modern power distribution systems: A review," Energies, vol. 15, no. 19, pp. 7264, 2022

2022
[18]

A survey of graph- based deep learning for anomaly detection in distributed systems,

A. D. Pazho, H. Wu, M. S. Obaidat and A. Nayak, "A survey of graph- based deep learning for anomaly detection in distributed systems," IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 1, pp. 1– 20, 2023

2023
[19]

CNN-based transformer model for fault detection in power system networks,

J. B. Thomas, S. G. Chaudhari and N. K. Verma, "CNN-based transformer model for fault detection in power system networks," IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–10, 2023

2023
[20]

Semi-supervised Informer for the compound fault diagnosis of industrial robots,

C. Deng, Y. Wang, R. Li, Y. Li and W. Zhang, "Semi-supervised Informer for the compound fault diagnosis of industrial robots," Sensors, vol. 24, no. 12, pp. 3732, 2024

2024
[21]

Anomaly detection for multivariate time series based on contrastive learning and Autoformer,

X. Shang, W. Zhang, Z. Liu, Q. Chen and Y. Gao, "Anomaly detection for multivariate time series based on contrastive learning and Autoformer," Proceedings of the 2024 27th International Conference on Computer Supported Cooperative Work in Design, 2024

2024
[22]

Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting,

T. Zhou, Z. Ma, S. Wang and Y. Wang, "Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting," Proceedings of the 2022 International Conference on Machine Learning, 2022

2022

Pith tools

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

[1] [1]

Fault tolerance in distributed systems using deep learning approaches,

B. Assiri and A. Sheneamer, "Fault tolerance in distributed systems using deep learning approaches," PLOS One, vol. 20, no. 1, 2025

work page 2025

[2] [2]

Modeling multi- hop semantic paths for recommendation in heterogeneous information networks,

H. Zheng, Y. Xing, L. Zhu, X. Han, J. Du and W. Cui, "Modeling multi- hop semantic paths for recommendation in heterogeneous information networks," arXiv preprint arXiv:2505.05989, 2025

arXiv 2025

[3] [3]

Analyzing data augmentation techniques for contrastive learning in recommender models,

M. Wei, H. Xin, Y. Qi, Y. Xing, Y. Ren and T. Yang, "Analyzing data augmentation techniques for contrastive learning in recommender models," 2025

work page 2025

[4] [4]

Structured preference modeling for reinforcement learning-based fine-tuning of large models,

L. Zhu, F. Guo, G. Cai and Y. Ma, "Structured preference modeling for reinforcement learning-based fine-tuning of large models," Journal of Computer Technology and Software, vol. 4, no. 4, 2025

work page 2025

[5] [5]

Context- guided dynamic retrieval for improving generation quality in RAG models,

J. He, G. Liu, B. Zhu, H. Zhang, H. Zheng and X. Wang, "Context- guided dynamic retrieval for improving generation quality in RAG models," arXiv preprint arXiv:2504.19436, 2025

arXiv 2025

[6] [6]

Joint semantic detection and dissemination control of phishing attacks on social media via LLama-based modeling,

R. Wang, "Joint semantic detection and dissemination control of phishing attacks on social media via LLama-based modeling," 2025

work page 2025

[7] [7]

Deep learning-based fault prediction in wireless sensor network embedded cyber-physical systems for industrial processes,

H. Ruan, Y. Hu, C. Zhang, J. Shi and J. Zhang, "Deep learning-based fault prediction in wireless sensor network embedded cyber-physical systems for industrial processes," IEEE Access, vol. 10, pp. 10867– 10879, 2022

work page 2022

[8] [8]

Industrial artificial intelligence based energy management system: Integrated framework for electricity load forecasting and fault prediction,

Y. Hu, Q. Li, Y. Ruan, J. Zhang and X. Wang, "Industrial artificial intelligence based energy management system: Integrated framework for electricity load forecasting and fault prediction," Energy, vol. 244, pp. 123195, 2022

work page 2022