Multi-Level Service Performance Forecasting via Spatiotemporal Graph Neural Networks

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

REVIEW 4 major objections 6 minor 29 references

Multi-Level Service Performance Forecasting via Spatiotemporal Graph Neural Networks

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 A spatiotemporal graph neural network that couples graph convolution over service call graphs with gated recurrent temporal modeling can forecast multi-level backend service performance more accurately than three existing baselines, the pap

desk verdict A standard GCN-GRU with an unspecified time encoding and unreproducible experiments; the empirical superiority claim is untestable. read the letter →

arxiv 2508.07122 v1 pith:L5YSIT2A submitted 2025-08-09 cs.LG

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

classification cs.LG

keywords spatiotemporalgraphneuralnetworkserviceperformancepredictionbackendservicescallconvolutionalgatedrecurrentunittimeencodingclustertrace

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

Backend services fail in cascades, so predicting their performance means modeling both the call graph and how it changes over time. This paper proposes a spatiotemporal graph neural network that turns each time slice of a distributed system into a directed service call graph, extracts structural dependencies with graph convolution, and tracks their evolution with a gated recurrent unit. Trained end-to-end to regress future performance metrics, the model reports lower MAE and RMSE and higher R² than ASTGCN, DGCRN, and Graph WaveNet on a large public cluster trace. The paper also finds a 10-minute aggregation window is the sweet spot and that forecast quality remains above R² = 0.90 through high concurrency. If these results hold, the approach gives operations teams more accurate lead time for capacity planning and anomaly response.

What carries the argument

The central object is the time-indexed service call graph $G_t=(V_t,E_t)$ with node feature vectors $x_i^t \in \mathbb{R}^d$ and directed, weighted edges representing call frequency or dependency strength. The load-bearing mechanism is the GCN layer $H^{(l+1)}=\sigma(\tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2}H^{(l)}W^{(l)})$ with self-looped adjacency $\tilde{A}=A+I$, which propagates high-order dependency information across the topology; the GRU then carries these structural representations forward in time, and an MLP regression output produces $\hat{y}_i^{t+\Delta t}=f(h_i^t;\theta)$. This GCN-GRU-time-encoding stack is what unifies structural and temporal modeling.

What would settle it

Run the model on the same public cluster trace with an explicitly documented preprocessing pipeline and fixed train/test split; if the claimed MAE of 0.123, RMSE of 0.197, and R² of 0.941 are not reproduced within a reasonable tolerance, or if Graph WaveNet performs at least as well under identical conditions, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a backend service system can be represented as a sequence of directed graphs—one per time slice, with service nodes carrying runtime feature vectors and edges carrying call relationships—and that applying graph convolution to each slice, feeding the resulting structural representations through a gated recurrent unit, and regressing with a multilayer perceptron yields state-of-the-art forecasts of future service performance metrics. The model is trained end-to-end by minimizing mean squared error over predicted versus observed future metrics. On a large public production cluster trace, the paper reports that this design achieves MAE 0.123, RMSE 0.197, and R²

Load-bearing premise

The load-bearing premise is that the service call graph reconstructed from the cluster trace—its nodes, edges, edge weights, and per-node features—faithfully represents the system's real dependency structure; the paper does not specify how that reconstruction is done.

Editorial extensions

If this is right

Operators could use the model to anticipate service response-time degradation roughly 10 minutes ahead, before it propagates down the call chain.
The best prediction quality is achieved at a 10-minute temporal window; shorter windows lose context and longer windows blur non-stationary dynamics.
Under low-to-high concurrency (up to about 5,000 requests per second), the model keeps R² above 0.90, so routine production workloads are within its reliable range.
At extreme concurrency beyond 8,000 rps, errors rise but the model still tracks overall trends, so it remains usable as an early-warning signal during bursts.
An end-to-end MSE-trained GCN-GRU stack is enough to beat the three reported baselines on the multi-level service forecasting benchmark.

Reading between the lines

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

The paper's claimed advantage would be much more convincing if the graph-construction step—node identities, edge weights, feature alignment—were specified; an obvious next step is to publish that pipeline and run an ablation removing the graph convolution to isolate the GRU's contribution.
Because the model predicts per-node future metrics on a graph, the same architecture should transfer to other graph-structured telemetry, such as network traffic between routers, container orchestrations, or microservice meshes, whenever a directed dependency graph and time-series features are available.
The 10-minute optimal window suggests a sweet spot tied to the dataset's dynamics; on traces with slower or faster failure propagation, the optimal window would shift, so practitioners would need to re-tune it.
A natural test the paper leaves implicit is whether the reported gains persist under distribution shift, such as new service versions or traffic surges beyond the training range; the concurrency stability test is a first step in that direction.

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

Colleague,

I agree with the reader's assessment. The paper is a standard GCN-GRU stack with an unspecified "time encoding" that is the only claimed novelty, and the experimental section doesn't provide enough detail to evaluate the results. The problem—forecasting performance metrics across cascading backend services—is real and worth working on, and the authors do test across time windows and concurrency bands, which is more than many short papers bother with. The GCN and GRU equations are correct as far as they go. But that's where the good news ends.

The central claim is empirical superiority over ASTGCN, DGCRN, and Graph WaveNet on Alibaba 2018. That claim is untestable from the text. Section IV.A describes the dataset but never defines how it becomes a graph: we don't learn what the nodes, edges, edge weights, node features, or the prediction target are, nor the prediction horizon. The time encoding is mentioned in the abstract and Section III, then never described with a formula or even a sentence. That is the paper's own stated contribution, so this is not a minor detail. The experiments report a single table with no error bars, no number of seeds, no train/validation/test split, no hyperparameters. Figure 2 and 3 are referenced but not included in this text, so we can't see them. This is a textbook case of an unsupported empirical superiority claim.

The reference list also raises an eyebrow: many 2025 arXiv and conference papers in the related work have a "this inspires us" sentence but little substantive connection, and some affiliations look wrong. That suggests the related work was padded rather than engaged with.

Given all this, I would not send this to peer review as is. The instruction to "treat every part of the manuscript as in-scope evidence" makes it worse, because the missing pieces are load-bearing. A revised version that specifies the time encoding, graph construction, and split details, and includes error bars or multiple runs, might deserve a referee. This version does not.

Don't cite it; I wouldn't bring it to a reading group either. The paper's only value right now is as an example of how fast a sensible idea can be rendered useless by missing implementation details.

Referee Report

4 major / 6 minor

Summary. The paper proposes a spatiotemporal graph neural network (STGNN) for forecasting performance metrics of multi-level backend services. The method abstracts service call graphs over time, applies graph convolutional layers to capture structural dependencies, uses GRUs to model temporal dynamics, and includes an unspecified 'time encoding mechanism.' The model is trained end-to-end with an MSE loss. Experiments on the Alibaba Cluster Trace 2018 claim lower MAE/RMSE and higher R² than ASTGCN, DGCRN, and Graph WaveNet, along with additional experiments on time-window length and concurrency levels.

Significance. If properly validated, the paper would offer a practical application of existing spatiotemporal GNN components to service performance forecasting. The combination of GCN and GRU is standard, so the methodological novelty is limited. The main potential contribution is empirical: demonstrating that such a model outperforms strong baselines on a public cluster trace. However, that contribution is currently not assessable because the graph construction, time encoding, hyperparameters, and experimental protocol are not specified. No code or reproducibility artifacts are provided. The paper therefore cannot, as written, support its central claim of superiority.

major comments (4)

[IV.A and III] The model input is never operationalized. Equations (1)-(4) assume a well-defined adjacency matrix A, degree matrix D, feature matrix H, and prediction target y^{t+Δt}, but the paper does not specify how V_t, E_t, edge weights, node features, or the target metric are derived from the Alibaba Cluster Trace 2018. The statement in IV.A that 'fine-grained parsing ... accurately reconstructed' is not supported by any concrete mapping from raw logs to graph objects. Without this, the empirical results in Table 1 cannot be reproduced or interpreted; the entire experimental claim rests on an unstated preprocessing pipeline.
[III and IV.B] The 'time encoding mechanism' is named but never defined. No formula, positional/periodic encoding choice, or integration with Eqs. (1)-(2) is provided, despite the abstract and Section III claiming it enhances non-stationary sequence modeling. The architecture is therefore incompletely specified. Similarly, Eq. (3) mentions an MLP regression head but the number of layers/units is omitted.
[IV.B, Table 1] No experimental protocol is reported: prediction horizon Δt, train/validation/test split, number of independent runs, seeds, hyperparameters (GCN layers, hidden sizes, learning rate, dropout, optimizer), or input scaling are absent. Table 1 gives point estimates only; without error bars or multiple seeds, differences such as MAE 0.123 vs 0.142 may be noise. The baseline configurations are also unspecified, so the claimed superiority over ASTGCN, DGCRN, and Graph WaveNet is unsupported as presented.
[IV.B, Figures 2-3] The time-window and concurrency experiments are reported qualitatively. Figure 2 is not shown in the submitted text, and Figure 3 lacks axis labels, units, and error bars. The text states the model is 'optimal' at 10 minutes, and that R² exceeds 0.90 for low-to-high concurrency, but no quantitative values are given. Moreover, the 10-minute window appears to be selected after inspecting test performance, with no held-out validation procedure described, risking selection on the test set.

minor comments (6)

[Section III, Eq. (1)] The equation is garbled by font issues; please rewrite cleanly and define all symbols.
[Section III, Eq. (2)] The input z_i^t is not defined precisely; specify how GCN outputs are transformed into GRU inputs.
[Section IV.A] The dataset section does not state the temporal resolution, the number of services/nodes, or the chosen prediction target (CPU, response time, memory, etc.). Add these details.
[Section IV.B] Report units for MAE/RMSE and clarify which performance metric is being predicted.
[References] Several references appear tangential or self-referential; please verify each citation supports the sentence it anchors and prune irrelevant entries.
[Figure 1] Figure 1 is referenced but not present in the text; include the architecture diagram.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model is a standard trainable GCN+GRU+MLP pipeline whose empirical claims rest on external benchmark experiments, not on self-referential derivations.

full rationale

The paper's derivation chain is: represent the system as a graph G_t=(V_t,E_t); apply graph convolution (Eq. 1) to obtain structural features; apply GRU (Eq. 2) for temporal dynamics; apply an MLP regressor (Eq. 3) to predict future performance y_i^{t+Δt}; and train end-to-end with MSE loss (Eq. 4). Each step is a function of the input features and learned weights; the target is an actual future observation, not a quantity already used as a fitted input. I find no fitted parameter that is later renamed a prediction, no equation in which the claimed result is equal by construction to an input, no uniqueness theorem imported from the authors' prior work, and no load-bearing self-citation. The related-work citations are used only as design inspiration (e.g., 'This inspires our use of graph convolutions...'), not as evidence that the proposed model works; the empirical claim rests on Table 1 and Figures 2-3. The paper's central weakness is reproducibility: the trace-to-graph reconstruction, time encoding mechanism, hyperparameters, splits, and prediction horizon are underspecified, so Table 1 cannot be independently checked. That is a rigor and correctness concern, not a circularity concern. Per the reviewing rules, lack of standard consensus or lack of experimental detail is not itself circularity. Therefore the honest finding is no significant circularity, score 0.

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

The paper's model is a composition of standard GCN and GRU blocks, so the central claim rests on the assumption that the public dataset can be faithfully turned into graphs and that the chosen hyperparameters are valid. No new theoretical entities are introduced.

free parameters (3)

Time window length (T) = 10 minutes (selected as optimal)
The paper tests multiple time windows (5 to 30+ min) and selects 10 minutes based on the best metrics; no validation set is described to justify this choice.
Concurrency level thresholds = Low (<=1000 rps), Medium (1001-2500), High (2501-5000), Very High (5001-8000), Extreme (>8000)
These bin boundaries are arbitrary and affect the stability analysis; no sensitivity analysis is given.
Model hyperparameters (GCN layers, hidden sizes, learning rate, dropout, etc.)
None are reported, making the model impossible to replicate and its reported performance unverifiable.

assumptions (4)

standard math GCN propagation rule (Equation 1) from Kipf and Welling is adopted as the graph convolution operator.
The equation is the standard first-order approximation of spectral graph convolution.
standard math GRU update equations (Equation 2) are used for temporal modeling.
The GRU is a widely known recurrent architecture; no modification is presented.
domain assumption The Alibaba Cluster Trace 2018 can be transformed into a series of service call graphs with reliable per-node features.
Section IV.A makes this assumption but gives no mapping from raw cluster logs to graph nodes/edges/features.
domain assumption The service call structure within each time window is effectively static for GCN processing.
The paper models each time slice as a graph G_t, but the GCN is applied on a snapshot; changes in topology are not dynamically incorporated beyond re-computing adjacency per time step.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Level Service Performance Forecasting via Spatiotemporal Graph Neural Networks." pith.science (2026). https://pith.science/paper/L5YSIT2A

@misc{pith2026250807122,
  author       = {Pith},
  title        = {Pith review of: Multi-Level Service Performance Forecasting via Spatiotemporal Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5YSIT2A}},
  note         = {Machine review of arXiv:2508.07122}
}

read the original abstract

This paper proposes a spatiotemporal graph neural network-based performance prediction algorithm to address the challenge of forecasting performance fluctuations in distributed backend systems with multi-level service call structures. The method abstracts system states at different time slices into a sequence of graph structures. It integrates the runtime features of service nodes with the invocation relationships among services to construct a unified spatiotemporal modeling framework. The model first applies a graph convolutional network to extract high-order dependency information from the service topology. Then it uses a gated recurrent network to capture the dynamic evolution of performance metrics over time. A time encoding mechanism is also introduced to enhance the model's ability to represent non-stationary temporal sequences. The architecture is trained in an end-to-end manner, optimizing the multi-layer nested structure to achieve high-precision regression of future service performance metrics. To validate the effectiveness of the proposed method, a large-scale public cluster dataset is used. A series of multi-dimensional experiments are designed, including variations in time windows and concurrent load levels. These experiments comprehensively evaluate the model's predictive performance and stability. The experimental results show that the proposed model outperforms existing representative methods across key metrics such as MAE, RMSE, and R2. It maintains strong robustness under varying load intensities and structural complexities. These results demonstrate the model's practical potential for backend service performance management tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

[1]

Micro frontend based performance improvement and prediction for microservices using machine learning,

N. Kaushik, H. Kumar, and V. Raj, "Micro frontend based performance improvement and prediction for microservices using machine learning," Journal of Grid Computing, vol. 22, no. 2, 44, 2024

work page 2024
[2]

Suanming: Explainable prediction of performance degradations in microservice applications,

J. Grohmann, M. Straesser, A. Chalbani, et al., "Suanming: Explainable prediction of performance degradations in microservice applications," Proceedings of the ACM/SPEC International Conference on Performance Engineering, pp. 165-176, 2021

work page 2021
[3]

Intelligent performance prediction: The use case of a Hadoop cluster,

D. Uzunidis, P. Karkazis, C. Roussou, et al., "Intelligent performance prediction: The use case of a Hadoop cluster," Electronics, vol. 10, no. 21, 2690, 2021

work page 2021
[4]

Leveraging convolutional neural network-transformer synergy for predictive modeling in risk- based applications,

Y. Wang, Z. Xu, Y. Yao, J. Liu, and J. Lin, "Leveraging convolutional neural network-transformer synergy for predictive modeling in risk- based applications," Proceedings of the 2024 4th International Conference on Electronic Information Engineering and Computer Communication, pp. 1565-1570, 2024

work page 2024
[5]

Market turbulence prediction and risk control with improved A3C reinforcement learning,

J. Liu, X. Gu, H. Feng, Z. Yang, Q. Bao, and Z. Xu, "Market turbulence prediction and risk control with improved A3C reinforcement learning," Proceedings of the 2025 8th International Conference on Advanced Algorithms and Control Engineering, pp. 2634-2638, 2025

work page 2025
[6]

AI back-end as a service for learning switching of mobile apps between the fog and the cloud,

D. Athanasopoulos and D. Liu, "AI back-end as a service for learning switching of mobile apps between the fog and the cloud," IEEE Transactions on Services Computing, vol. 15, no. 2, pp. 656-668, 2021

work page 2021
[7]

Entity boundary detection in social texts using BiLSTM-CRF with integrated social features,

Y. Zhao, W. Zhang, Y. Cheng, Z. Xu, Y. Tian, and Z. Wei, "Entity boundary detection in social texts using BiLSTM-CRF with integrated social features," 2025

work page 2025
[8]

Unsupervised detection of fraudulent transactions in e-commerce using contrastive learning,

X. Li, Y. Peng, X. Sun, Y. Duan, Z. Fang, and T. Tang, "Unsupervised detection of fraudulent transactions in e-commerce using contrastive learning," Proceedings of the 2025 4th International Symposium on Computer Applications and Information Technology, pp. 1663-1667, 2025

work page 2025

Show all 29 references

[9]

Regression analysis of predictions and forecasts of cloud data center KPIs using the boosted decision tree algorithm,

T. W. Gyeera, A. J. H. Simons, and M. Stannett, "Regression analysis of predictions and forecasts of cloud data center KPIs using the boosted decision tree algorithm," IEEE Transactions on Big Data, vol. 9, no. 4, pp. 1071-1085, 2022

2022
[10]

Microservice-oriented workload prediction using deep learning,

S. Ştefan and V. Niculescu, "Microservice-oriented workload prediction using deep learning," e-Informatica Software Engineering Journal, vol. 16, no. 1, 2022

2022
[11]

PERT-GNN: Latency prediction for microservice-based cloud-native applications via graph neural networks,

D. S. H. Tam, Y. Liu, H. Xu, et al., "PERT-GNN: Latency prediction for microservice-based cloud-native applications via graph neural networks," Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 2155-2165, 2023

2023
[12]

Cross-scale attention and multi- layer feature fusion YOLOv8 for skin disease target detection in medical images,

T. Xu, Y. Xiang, J. Du, and H. Zhang, "Cross-scale attention and multi- layer feature fusion YOLOv8 for skin disease target detection in medical images," Journal of Computer Technology and Software, vol. 4, no. 2, 2025

2025
[13]

RT-DETR-based multimodal detection with modality attention and feature alignment,

Y. Lou, "RT-DETR-based multimodal detection with modality attention and feature alignment," Journal of Computer Technology and Software, vol. 3, no. 5, 2024

2024
[14]

Integrating system state into spatio temporal graph neural network for microservice workload prediction,

Y. Luo, M. Gao, Z. Yu, et al., "Integrating system state into spatio temporal graph neural network for microservice workload prediction," Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 5521-5531, 2024

2024
[15]

A hybrid recommendation approach integrating matrix decomposition and deep neural networks for enhanced accuracy and generalization,

R. Wang, Y. Luo, X. Li, Z. Zhang, J. Hu, and W. Liu, "A hybrid recommendation approach integrating matrix decomposition and deep neural networks for enhanced accuracy and generalization," Proceedings of the 2025 5th International Conference on Neural Networks, Information and ...

2025
[16]

Context-aware rule mining using a dynamic transformer- based framework,

J. Liu, et al., "Context-aware rule mining using a dynamic transformer- based framework," Proceedings of the 2025 8th International Conference on Advanced Algorithms and Control Engineering, 2025

2025
[17]

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
[18]

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," Proceedings of the 2025 8th International Conference on Advanced Algorithms and Control Engineering, pp. 1492-...

2025
[19]

Temporal-spatial deep learning for memory usage forecasting in cloud servers,

K. Aidi and D. Gao, "Temporal-spatial deep learning for memory usage forecasting in cloud servers," 2025

2025
[20]

Deep probabilistic modeling of user behavior for anomaly detection via mixture density networks,

L. Dai, W. Zhu, X. Quan, R. Meng, S. Chai, and Y. Wang, "Deep probabilistic modeling of user behavior for anomaly detection via mixture density networks," arXiv preprint arXiv:2505.08220, 2025

2025 arXiv
[21]

Knowledge-informed policy structuring for multi-agent collaboration using language models,

Y. Ma, G. Cai, F. Guo, Z. Fang, and X. Wang, "Knowledge-informed policy structuring for multi-agent collaboration using language models," Journal of Computer Science and Software Applications, vol. 5, no. 5, 2025

2025
[22]

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

2025 arXiv
[23]

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
[24]

Self-supervised credit scoring with masked autoencoders: Addressing data gaps and noise robustly,

Y. Yao, "Self-supervised credit scoring with masked autoencoders: Addressing data gaps and noise robustly," Journal of Computer Technology and Software, vol. 3, no. 8, 2024

2024
[25]

Temporal graph representation learning for evolving user behavior in transactional networks,

X. Liu, Q. Xu, K. Ma, Y. Qin, and Z. Xu, "Temporal graph representation learning for evolving user behavior in transactional networks," 2025

2025
[26]

MPDP: A probabilistic architecture for microservice performance diagnosis and prediction,

T. H. Noor, "MPDP: A probabilistic architecture for microservice performance diagnosis and prediction," Computer Systems Science & Engineering, vol. 48, no. 5, 2024

2024
[27]

AST-GCN: Attribute-augmented spatiotemporal graph convolutional network for traffic forecasting,

J. Zhu, Q. Wang, C. Tao, et al., "AST-GCN: Attribute-augmented spatiotemporal graph convolutional network for traffic forecasting," IEEE Access, vol. 9, pp. 35973-35983, 2021

2021
[28]

Hierarchical dynamic graph convolutional network for spatio-temporal forecasting,

S. Li, B. Xu, and X. Wang, "Hierarchical dynamic graph convolutional network for spatio-temporal forecasting," Proceedings of the 2024 International Joint Conference on Neural Networks, pp. 1-8, 2024

2024
[29]

Graph WaveNet for deep spatial- temporal graph modeling,

Z. Wu, S. Pan, G. Long, et al., "Graph WaveNet for deep spatial- temporal graph modeling," arXiv preprint arXiv:1906.00121, 2019

1906 arXiv

Pith tools

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

[1] [1]

Micro frontend based performance improvement and prediction for microservices using machine learning,

N. Kaushik, H. Kumar, and V. Raj, "Micro frontend based performance improvement and prediction for microservices using machine learning," Journal of Grid Computing, vol. 22, no. 2, 44, 2024

work page 2024

[2] [2]

Suanming: Explainable prediction of performance degradations in microservice applications,

J. Grohmann, M. Straesser, A. Chalbani, et al., "Suanming: Explainable prediction of performance degradations in microservice applications," Proceedings of the ACM/SPEC International Conference on Performance Engineering, pp. 165-176, 2021

work page 2021

[3] [3]

Intelligent performance prediction: The use case of a Hadoop cluster,

D. Uzunidis, P. Karkazis, C. Roussou, et al., "Intelligent performance prediction: The use case of a Hadoop cluster," Electronics, vol. 10, no. 21, 2690, 2021

work page 2021

[4] [4]

Leveraging convolutional neural network-transformer synergy for predictive modeling in risk- based applications,

Y. Wang, Z. Xu, Y. Yao, J. Liu, and J. Lin, "Leveraging convolutional neural network-transformer synergy for predictive modeling in risk- based applications," Proceedings of the 2024 4th International Conference on Electronic Information Engineering and Computer Communication, pp. 1565-1570, 2024

work page 2024

[5] [5]

Market turbulence prediction and risk control with improved A3C reinforcement learning,

J. Liu, X. Gu, H. Feng, Z. Yang, Q. Bao, and Z. Xu, "Market turbulence prediction and risk control with improved A3C reinforcement learning," Proceedings of the 2025 8th International Conference on Advanced Algorithms and Control Engineering, pp. 2634-2638, 2025

work page 2025

[6] [6]

AI back-end as a service for learning switching of mobile apps between the fog and the cloud,

D. Athanasopoulos and D. Liu, "AI back-end as a service for learning switching of mobile apps between the fog and the cloud," IEEE Transactions on Services Computing, vol. 15, no. 2, pp. 656-668, 2021

work page 2021

[7] [7]

Entity boundary detection in social texts using BiLSTM-CRF with integrated social features,

Y. Zhao, W. Zhang, Y. Cheng, Z. Xu, Y. Tian, and Z. Wei, "Entity boundary detection in social texts using BiLSTM-CRF with integrated social features," 2025

work page 2025

[8] [8]

Unsupervised detection of fraudulent transactions in e-commerce using contrastive learning,

X. Li, Y. Peng, X. Sun, Y. Duan, Z. Fang, and T. Tang, "Unsupervised detection of fraudulent transactions in e-commerce using contrastive learning," Proceedings of the 2025 4th International Symposium on Computer Applications and Information Technology, pp. 1663-1667, 2025

work page 2025