Utilizing Graph Neural Networks for Effective Link Prediction in Microservice Architectures

Alireza Ezaz; Ghazal Khodabandeh; Majid Babaei; Naser Ezzati-Jivan

REVIEW 5 major objections 5 minor 28 references

Utilizing Graph Neural Networks for Effective Link Prediction in Microservice Architectures

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A graph attention network predicts future microservice call links with 0.91 accuracy and 0.92 F1 on a real cluster trace.

desk verdict Reasonable application paper, but the evaluation makes the headline performance claim unsupported; worth a round of major revision, not rejection. read the letter →

arxiv 2501.15019 v1 pith:RRAAOJD5 submitted 2025-01-25 cs.LG

Ghazal Khodabandeh , Alireza Ezaz , Majid Babaei , Naser Ezzati-Jivan This is my paper · ORCID

classification cs.LG

keywords linkpredictionmicroservicecallgraphsgraphattentionnetworktemporalsegmentationnegativesamplingneuralnetworksdistributedsystemsadaptivemonitoring

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 tries to establish that a graph attention network can reliably forecast which microservices will call which other microservices in the near future. It argues that microservice call graphs are denser and more time-sensitive than social networks, so link prediction needs a model that blends structural and temporal information. If correct, the approach would let operators spot emerging dependencies and intervene before a failure cascades through a distributed system. The paper reports accuracy of 0.91, F1 of 0.92, precision of 0.89, and recall of 0.96 on a real-world microservice trace, outperforming simpler graph, LSTM, and similarity-based baselines.

What carries the argument

The central object is a Graph Attention Network (GAT) operating on time-windowed directed call graphs. Each window produces its own graph; nodes are microservices and edges are caller-to-callee interactions. The GAT computes attention coefficients between nodes, aggregates neighbor features with those coefficients, and outputs embeddings. Link probability is the sigmoid of the dot product of two node embeddings. Advanced negative sampling selects non-edges with probability proportional to node degree raised to a tunable power, so negative examples concentrate near central hubs while explicitly excluding existing edges. Training uses binary cross-entropy loss and a threshold on the link probability for classification.

What would settle it

Look in the full trace for pairs labeled as negatives during testing and check whether they appear as positive calls anywhere outside the sampled windows; if many do, recompute precision, recall, and AUC using only pairs whose non-existence is confirmed in the complete trace. If accuracy drops materially below 0.91, the central claim fails.

Watch

Extended reading notes

Core claim

This paper claims that a Graph Attention Network combined with fixed-window temporal segmentation and degree-weighted negative sampling predicts future edges in a microservice call graph with accuracy 0.91 and F1 0.92 on the real-world trace dataset. The model builds a directed graph for each time window, uses identity node features and timestamp edge features, and learns link probabilities through a dot-product score passed through a sigmoid. Its attention mechanism assigns different weights to different neighbors, which the authors argue is essential when certain service interactions matter more than others. The reported results beat the NodeSim, adjusted NodeSim, LSTM, simple GNN, and simple temporal GNN baselines on the same data, and the attention heatmaps show which connections the model prioritizes as training converges.

Load-bearing premise

The evaluation treats every caller-callee pair not seen in the sampled windows as a true negative; if the underlying trace is incomplete and some of those pairs actually call each other, the negative set is contaminated and the reported metrics are not a true measure of predictive quality.

Editorial extensions

If this is right

Operators could use predicted future call edges to enable adaptive monitoring, tracing, and logging, catching performance bottlenecks before they escalate.
Attention weights identify which inter-service connections the model considers influential, offering a way to prioritize debugging and resource allocation.
The temporal-window-plus-GAT recipe could transfer to other dense, time-sensitive networks such as fraud detection, recommendation systems, or social networks, as the paper itself suggests.
The contrast with static similarity methods indicates that temporal structure is a major source of predictive signal in microservice environments.
The model's strong recall suggests it is well suited for alerting systems where missing a real call is more costly than a false alarm.

Reading between the lines

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

Editorial inference: because node features are only identity vectors and edge features are just timestamps, the reported performance mostly reflects graph structure and temporal windowing; adding latency, error-rate, or resource-usage features is a cheap, testable way to push accuracy further.
Editorial inference: the paper's Table 2 shows the simple GNN achieving a higher AUC (0.94) than the proposed approach (0.89), so the claim of superiority is metric-dependent; a reader comparing methods should weigh all metrics rather than fixating on F1.
Editorial inference: on a trace whose completeness is unverified, the reported precision and recall may be optimistic; a stronger validation would rerun the experiment on a dataset where non-edges are confirmed by independent ground truth.
Editorial inference: the 10-second evaluation horizon is short relative to production incident timescales; testing on longer horizons would reveal whether the temporal windows retain their advantage as interaction patterns drift.

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.

Desk Editor's Note

read the letter

Short version: this is an applied GNN paper with a sensible motivation and a flawed evaluation. The core claim—that a GAT with temporal windows and degree-weighted negative sampling predicts future microservice call edges with F1 0.92—is not supported by the evidence as presented.

What the paper does well: the problem is real and under-studied. The authors use a large public trace (Alibaba 2022), construct time-windowed call graphs, compare against LSTM and simple GNN variants, and they are unusually candid in Section 4.5 about the single-dataset and metric limitations. The code link is provided, which is a plus.

The soft spots are mostly concentrated in the evaluation section, and they are not minor. First, the test set is generated with the same Advanced Negative Sampling used in training (Algorithm 3, Section 3.6). That gives a roughly 50/50 positive/negative test distribution instead of the real, highly imbalanced distribution of all unobserved caller–callee pairs. Accuracy and precision numbers on that synthetic balanced sample do not tell an operator how many false alarms the model would produce in production. This alone undermines the headline result.

Second, Table 2 reports AUC 0.89 for 'Our Approach', lower than Simple GNN's 0.94, while Section 4.3.3 and Figure 5 claim an ROC curve with 'AUC approaching 1'. AUC is threshold-independent and standard for link prediction; you cannot have both statements true. The discrepancy suggests either a typo or selective reporting.

Third, the temporal modeling claim is internally inconsistent. Section 3.3 says node features are an identity matrix, and Section 3.5.1 says timestamps are embedded in node features. Both cannot hold. Also, the train/test split is 0–7000 ms vs 7000–10000 ms from a 13-day trace, which is a 10-second slice; generalizing from that to 'proactive monitoring' in production is a stretch.

There is no independent validation split for hyperparameters (alpha, threshold, window size), so the reported numbers are partially tuned on the test period. Single-run metrics with no error bars make it impossible to say whether the observed advantages over baselines are real or noise.

I want to be fair: the paper is not incoherent, and the authors do flag some limitations themselves. The combination of GAT + temporal windows + degree-weighted negative sampling is reasonable to try. But as it stands, the evidence does not establish the claim.

Who gets value: researchers working on microservice observability who want a starting point and a clear list of pitfalls to avoid. Not a paper to cite for performance numbers.

Recommendation: yes, send to peer review—a serious referee might get the authors to fix the evaluation. But in its current form I would not accept or trust the reported metrics.

Referee Report

5 major / 5 minor

Summary. The paper proposes a Graph Attention Network (GAT) with temporal segmentation and degree-based 'advanced negative sampling' for link prediction in microservice call graphs. Using the Alibaba 2022 Cluster Trace, it compares the approach against NodeSim, an adjusted NodeSim, an LSTM, and simple GNN baselines. The reported results claim high accuracy (0.91), recall (0.96), and F1 (0.92), with an AUC of 0.89 in Table 2. The authors argue that GNNs can support proactive monitoring in microservice architectures, and they provide qualitative visualizations including confusion matrices, attention heatmaps, PR curves, and an ROC curve.

Significance. If the central performance claim were established, the paper would offer a practical demonstration of GAT-based link prediction for microservice monitoring, a comparatively underexplored application. The use of real-world trace data and the public code repository are strengths, as are the comparisons against several baseline families. However, the evidence as presented is not sufficient to support the headline claims: the evaluation protocol uses a degree-biased balanced negative sample that does not reflect the real link-prediction task, the reported AUC contradicts the text, and the claimed temporal modeling is not implemented in the node features. The paper would be significant if these issues were resolved with a clear evaluation methodology and consistent reporting, but in its current form its conclusions are not reliably supported.

major comments (5)

[§3.4, §3.6, and Table 2] The evaluation computes accuracy, precision, recall, and F1 on a test set whose negatives are generated by the same Advanced Negative Sampling used in training (Algorithm 3), yielding a roughly balanced set of positive and negative pairs. This does not measure performance on the actual link-prediction task, where unobserved caller–callee pairs are the negative set and positives are sparse. As a result, the reported accuracy (0.91) and F1 (0.92) characterize a synthetic balanced distribution, not the operational task of forecasting future edges. Please evaluate on a realistic candidate-pair distribution, for example all unobserved pairs or an unbiased subsample, and report ranking metrics such as Hits@K or AUC on that distribution.
[Table 2 vs. §4.3.3 and Figure 5] Table 2 reports an AUC of 0.89 for 'Our Approach,' which is lower than the 0.94 AUC of Simple GNN, yet §4.3.3 and Figure 5 state that the ROC curve shows an 'AUC approaching 1.' Since AUC is threshold-independent and is the standard ranking metric for link prediction, this internal contradiction is unresolved and directly undercuts the claim that the proposed method outperforms the baselines. The authors should reconcile the numerical AUC with the figure and text, and report per-window AUC values with proper aggregation.
[§3.5.1, Algorithm 2, and §3.3] The paper claims in §3.5.1 that 'Temporal information is embedded in both the node and edge features by incorporating timestamps within the node features,' and in §3.5.2 that 'Temporal dependencies are implicitly captured through the inclusion of timestamp features.' However, Algorithm 2 sets node features to an identity matrix (g.x ← identity_matrix(n_nodes)), and §3.3 explicitly states that 'we use an identity matrix as the sole feature representation for nodes.' No timestamp features are present in the model input as described. The temporal modeling claim is therefore unsupported by the described implementation, and the manuscript should either implement temporal features or soften the claim accordingly.
[§4.3 and Table 2] All experiments appear to be single runs: Table 2 reports a single value per metric with no error bars, standard deviations, or repetitions. Furthermore, the Simple GNN and Simple Temporal GNN baselines are not described with enough detail (architecture, layer counts, hidden sizes, number of runs, or hyperparameter selection) to permit fair comparison. Without variance estimates and a precise baseline configuration, it is not possible to assess whether the reported differences are statistically meaningful. Please include repeated runs with error bars and specify all baseline settings.
[§4.1.2 and §3.4] The hyperparameters (α in Eq. (6), the link threshold τ, the time-window size, and the training/validation split) are chosen without an independent validation split. The training and test periods are fixed as 0–7000 ms and 7000–10000 ms, but the text indicates that α and τ were tuned on the data; if this tuning used the test period, the reported metrics are partially fitted to the test set. Please describe the validation procedure explicitly, or rerun the evaluation with a separate validation interval that is excluded from hyperparameter selection.

minor comments (5)

[Abstract and §4.3.2] The abstract and §4.3.2 state 'achieving an accuracy of 0.91 and an F1 score of 0.92' without mentioning that these are conditioned on the degree-biased negative-sample distribution; consider qualifying these numbers in both places to avoid overstatement.
[§3.2, Algorithm 1] Algorithm 1 uses the notation 'unique nodes' and 'map and encode' but does not define 'um' and 'dm' fields; these abbreviations should be spelled out or linked to the data description.
[§3.3, Eq. (5)] Equation (5) defines the simple negative sample set, but the notation for directed edges is inconsistent with the edge set definition in §3.3; please make the directed or undirected nature of E explicit throughout.
[§4.3.3, Figure 5] The ROC curve figure caption describes 'Time Window 21' but the text says the pattern is consistent across multiple windows; please indicate how many windows were evaluated and whether Figure 5 is representative or averages over windows.
[§4.5] The threats-to-validity section mentions that MRR and Hits@K are not used, that only one dataset is considered, and that the time range is limited to 10,000 ms; these are appropriate limitations, but they are not connected to the negative-sampling and validation issues raised above, which are more fundamental to the reported performance.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the GAT link-prediction pipeline is a standard empirical pipeline trained and evaluated on temporal splits, with no load-bearing self-citations and no prediction defined by its inputs.

full rationale

The paper's derivation chain is a standard supervised GNN link-prediction pipeline: temporal window graph construction (Algorithms 1-2), degree-weighted advanced negative sampling (Algorithm 3 and Eq. 6), GAT scoring (Eqs. 7-12), binary cross-entropy training (Eq. 10), and a temporally separated train/test split. No equation defines the target metric in terms of the fitted parameters, and no prediction is equivalent to an input by construction. The reported metrics are empirical outcomes on held-out future time windows, albeit evaluated against negatively sampled pairs. The evaluation-validity concerns (test negatives generated by the same sampling distribution, threshold chosen to balance precision and recall without a documented validation split, and the Table 2 AUC of 0.89 conflicting with the text's 'AUC approaching 1') are reporting and methodology weaknesses, not circularity. There are no load-bearing self-citations: the cited negative-sampling and GAT works are external references, and the core model is not justified by an author-invoked uniqueness theorem or ansatz. Under the hard rules, no specific reduction of a claim to its own inputs can be exhibited, so no circular step is established.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on standard GAT machinery plus several dataset-specific choices (alpha, threshold, window size, node features). The heaviest unexamined assumption is trace completeness, which determines the validity of the negative labels.

free parameters (5)

alpha (negative sampling) = 0.1
Exponent in degree-weighted sampling probability (Eq. 6); the paper says alpha is tuned but does not report a search procedure or validation set.
link threshold tau = not reported
Threshold in Eq. 13 chosen to balance precision and recall; the value is not given.
time window size = not reported
The paper gives train/test spans (0-7000 ms, 7000-10000 ms) but not the window width used to segment graphs.
embedding dimension / hidden size = not reported
Described as 'determined through experimentation' without reporting the value.
learning rate / epochs = not reported
Attention heatmaps show 200 epochs, but learning rate and other optimizer settings are absent.

assumptions (4)

domain assumption Observed trace is complete: every real microservice interaction appears in the trace, so all unobserved pairs are true negatives.
Section 3.4 builds the negative set from pairs not in the observed edges; if calls are missed, negatives are contaminated.
domain assumption Fixed time windows contain stationary enough structure that a static GAT trained on early windows transfers to later windows.
Section 3.2 segments data into fixed windows; Section 3.5 trains a static GAT per window without modeling edge timing.
ad hoc to paper Identity node features are sufficient to represent microservice identity; attribute vectors (service name, latency, RPC type) carry no predictive signal.
Section 3.3 discards attributes 'to avoid unnecessary complexity' even though the trace collects them.
standard math GAT equations, BCE loss, and backpropagation are correct (standard math).
Equations 7 to 10 are standard, cited from the literature.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Utilizing Graph Neural Networks for Effective Link Prediction in Microservice Architectures." pith.science (2026). https://pith.science/paper/RRAAOJD5

@misc{pith2026250115019,
  author       = {Pith},
  title        = {Pith review of: Utilizing Graph Neural Networks for Effective Link Prediction in Microservice Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRAAOJD5}},
  note         = {Machine review of arXiv:2501.15019}
}

read the original abstract

Managing microservice architectures in distributed systems is complex and resource intensive due to the high frequency and dynamic nature of inter service interactions. Accurate prediction of these future interactions can enhance adaptive monitoring, enabling proactive maintenance and resolution of potential performance issues before they escalate. This study introduces a Graph Neural Network GNN based approach, specifically using a Graph Attention Network GAT, for link prediction in microservice Call Graphs. Unlike social networks, where interactions tend to occur sporadically and are often less frequent, microservice Call Graphs involve highly frequent and time sensitive interactions that are essential to operational performance. Our approach leverages temporal segmentation, advanced negative sampling, and GATs attention mechanisms to model these complex interactions accurately. Using real world data, we evaluate our model across performance metrics such as AUC, Precision, Recall, and F1 Score, demonstrating its high accuracy and robustness in predicting microservice interactions. Our findings support the potential of GNNs for proactive monitoring in distributed systems, paving the way for applications in adaptive resource management and performance optimization.

Figures

Figures reproduced from arXiv: 2501.15019 by the authors.

**Figure 1.** Methodology Diagram a clear view of how microservice interactions evolve across distinct periods. For example, a microservice handling user requests may show high interaction frequency with a database service during peak usage times and much lower activity during off-peak hours. This fixed-window approach enables the model to learn sequential, time-dependent relationships, which is essential in microservice architec… view at source ↗

**Figure 2.** Confusion matrix illustrating the performance of [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 4.** Precision-Recall (PR) curves for representative [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Attention weight heatmaps for nodes 0-100 at selected training epochs (Epochs 0, 49, 99, 149, and 199). Each subfigure [PITH_FULL_IMAGE:figures/full_fig_p010_3.png]

**Figure 5.** Figure 5: Receiver Operating Characteristic (ROC) curve for [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 21 canonical work pages

[1]

Djihad Arrar, Nadjet Kamel, and Abdelaziz Lakhfif. 2024. A comprehensive survey of link prediction methods. The journal of supercomputing 80, 3 (2024), 3902–3942

work page 2024
[2]

Lars Backstrom and Jure Leskovec. 2011. Supervised random walks: predicting and recommending links in social networks. In Proceedings of the fourth ACM international conference on Web search and data mining . 635–644

work page 2011
[3]

Barış Fındık. 2024. USING TOPOLOGICAL FEATURES OF MICROSERVICE CALL GRAPHS TO PREDICT THE RESPONSE TIME V ARIATION. Master’s thesis. Middle East Technical University

work page 2024
[4]

Weiwei Gu, Fei Gao, Xiaodan Lou, and Jiang Zhang. 2019. Link prediction via graph attention network. arXiv preprint arXiv:1910.04807 (2019)

arXiv 2019
[5]

Mohammad Al Hasan and Mohammed J Zaki. 2011. A survey of link prediction in social networks. Social network data analytics (2011), 243–275

work page 2011
[6]

Hangtao He, Linyu Su, and Kejiang Ye. 2023. GraphGRU: A graph neural network model for resource prediction in microservice cluster. In 2022 IEEE 28th Interna- tional Conference on Parallel and Distributed Systems (ICPADS) . IEEE, 499–506

work page 2023
[7]

Anand Padmanabha Iyer, Li Erran Li, Tathagata Das, and Ion Stoica. 2016. Time- evolving graph processing at scale. In Proceedings of the fourth international workshop on graph data management experiences and systems . 1–6

work page 2016
[8]

Chuanze Kang, Han Zhang, Zhuo Liu, Shenwei Huang, and Yanbin Yin. 2022. LR-GNN: A graph neural network based on link representation for predicting molecular associations. Briefings in Bioinformatics 23, 1 (2022), bbab513

2022

Show all 28 references

[9]

Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR)

2017
[10]

Bhushan Kotnis and Vivi Nastase. 2017. Analysis of the impact of negative sampling on link prediction in knowledge graphs.arXiv preprint arXiv:1708.06816 (2017)

2017 arXiv
[11]

Shanshan Li, He Zhang, Zijia Jia, Chenxing Zhong, Cheng Zhang, Zhihao Shan, Jinfeng Shen, and Muhammad Ali Babar. 2021. Understanding and addressing quality attributes of microservices architecture: A Systematic literature review. Information and software technology 131 (2021), 106449

2021
[12]

David Liben-Nowell and Jon Kleinberg. 2007. The link-prediction problem for social networks. Journal of the American Society for Information Science and Technology 58, 7 (2007), 1019–1031

2007
[13]

David B Lindenmayer and Gene E Likens. 2009. Adaptive monitoring: a new paradigm for long-term research and monitoring. Trends in ecology & evolution 24, 9 (2009), 482–486

2009
[14]

Linyuan Lü and Tao Zhou. 2011. Link prediction in complex networks: A survey. Physica A: statistical mechanics and its applications 390, 6 (2011), 1150–1170

2011
[15]

Shutian Luo, Huanle Xu, Chengzhi Lu, Kejiang Ye, Guoyao Xu, Liping Zhang, Jian He, and Chengzhong Xu. 2022. An in-depth study of microservice call graph and runtime performance. IEEE Transactions on Parallel and Distributed Systems 33, 12 (2022), 3901–3914

2022
[16]

Shang-Pin Ma, Chen-Yuan Fan, Yen Chuang, I-Hsiu Liu, and Ci-Wei Lan. 2019. Graph-based and scenario-driven microservice analysis, retrieval, and testing. Future Generation Computer Systems 100 (2019), 724–735

2019
[17]

Gianluca Ruberto. 2022. An experimental analysis of Link Prediction methods over Microservices Knowledge Graphs. (2022)

2022
[18]

Usha Ruby and Vamsidhar Yendapalli. 2020. Binary cross entropy with deep learning technique for image classification. Int. J. Adv. Trends Comput. Sci. Eng 9, 10 (2020)

2020
[19]

Prithwish Sarkar and Andrew W Moore. 2005. Dynamic social network analysis using latent space models. ACM SIGKDD Explorations Newsletter 7, 2 (2005), 31–40

2005
[20]

Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. 2001. Item-based collaborative filtering recommendation algorithms. In Proceedings of the 10th international conference on World Wide Web. 285–295

2001
[21]

Akrati Saxena, George Fletcher, and Mykola Pechenizkiy. 2022. NodeSim: node similarity based network embedding for diverse link prediction.EPJ Data Science 11, 1 (2022), 24

2022
[22]

Da Sun Handason Tam, Yang Liu, Huanle Xu, Siyue Xie, and Wing Cheong Lau. 2023. Pert-gnn: Latency prediction for microservice-based cloud-native applications via graph neural networks. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2155–2165

2023
[23]

Victor Velepucha and Pamela Flores. 2023. A survey on microservices architec- ture: Principles, patterns and migration challenges. IEEE Access (2023)

2023
[24]

Huan Wang, Ziwen Cui, Ruigang Liu, Lei Fang, and Ying Sha. 2023. A multi- type transferable method for missing link prediction in heterogeneous social networks. IEEE Transactions on Knowledge and Data Engineering 35, 11 (2023), 10981–10991. Ghazal Khodabandeh, Alireza Ezaz, Ma...

2023
[25]

Wen Wang, Wei Zhang, Shukai Liu, Qi Liu, Bo Zhang, Leyu Lin, and Hongyuan Zha. 2021. Incorporating link prediction into multi-relational item graph mod- eling for session-based recommendation. IEEE Transactions on Knowledge and Data Engineering 35, 3 (2021), 2683–2696

2021
[26]

Yuxin Wang, Xiannian Hu, Quan Gan, Xuanjing Huang, Xipeng Qiu, and David Wipf. 2024. Efficient Link Prediction via GNN Layers Induced by Negative Sampling. IEEE Transactions on Knowledge and Data Engineering (2024)

2024
[27]

Haixia Wu, Chunyao Song, Yao Ge, and Tingjian Ge. 2022. Link prediction on complex networks: an experimental survey. Data science and engineering 7, 3 (2022), 253–278

2022
[28]

Guotong Xue, Ming Zhong, Jianxin Li, Jia Chen, Chengshuai Zhai, and Ruochen Kong. 2022. Dynamic network embedding survey. Neurocomputing 472 (2022), 212–223

2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Djihad Arrar, Nadjet Kamel, and Abdelaziz Lakhfif. 2024. A comprehensive survey of link prediction methods. The journal of supercomputing 80, 3 (2024), 3902–3942

work page 2024

[2] [2]

Lars Backstrom and Jure Leskovec. 2011. Supervised random walks: predicting and recommending links in social networks. In Proceedings of the fourth ACM international conference on Web search and data mining . 635–644

work page 2011

[3] [3]

Barış Fındık. 2024. USING TOPOLOGICAL FEATURES OF MICROSERVICE CALL GRAPHS TO PREDICT THE RESPONSE TIME V ARIATION. Master’s thesis. Middle East Technical University

work page 2024

[4] [4]

Weiwei Gu, Fei Gao, Xiaodan Lou, and Jiang Zhang. 2019. Link prediction via graph attention network. arXiv preprint arXiv:1910.04807 (2019)

arXiv 2019

[5] [5]

Mohammad Al Hasan and Mohammed J Zaki. 2011. A survey of link prediction in social networks. Social network data analytics (2011), 243–275

work page 2011

[6] [6]

Hangtao He, Linyu Su, and Kejiang Ye. 2023. GraphGRU: A graph neural network model for resource prediction in microservice cluster. In 2022 IEEE 28th Interna- tional Conference on Parallel and Distributed Systems (ICPADS) . IEEE, 499–506

work page 2023

[7] [7]

Anand Padmanabha Iyer, Li Erran Li, Tathagata Das, and Ion Stoica. 2016. Time- evolving graph processing at scale. In Proceedings of the fourth international workshop on graph data management experiences and systems . 1–6

work page 2016

[8] [8]

Chuanze Kang, Han Zhang, Zhuo Liu, Shenwei Huang, and Yanbin Yin. 2022. LR-GNN: A graph neural network based on link representation for predicting molecular associations. Briefings in Bioinformatics 23, 1 (2022), bbab513

2022