Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Attention Enhanced Entity Recommendation for Intelligent Monitoring in Cloud Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper proposes DiRecGNN, a heterogeneous-graph recommender that learns which dimensions a cloud monitor should track, and reports that it outperforms existing methods by 43% in MRR on production data.

desk verdict Real industrial problem, good deployment story, but the HR@k table violates the definition of Hit Ratio and makes the headline gains untrustworthy. read the letter →

arxiv 2510.20640 v2 pith:YD5EYJCG submitted 2025-10-23 cs.LG

Fiza Husain , Anson Bastos , Anjaly Parayil , Ayush Choure , Chetan Bansal , Rujia Wang , Saravan Rajmohan This is my paper
classification cs.LG
keywords cloudmonitoringdimensionrecommendationheterogeneousgraphneuralnetworkattentionmechanismrandomwalksparselearningsystemproductiondeployment
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

Cloud services create monitors that watch metrics along chosen dimensions such as region, environment, and status code, but choosing the right subset is mostly manual and reactive. This paper turns that choice into a recommendation problem on a heterogeneous graph of monitors, metrics, and dimensions, and introduces DiRecGNN, which uses attention over both graph neighborhoods and random-walk paths, plus a loss that keeps attention sparse. On a production dataset, it reports HR@1 of 0.597, MRR of 0.714, and a 43.1% MRR improvement over the best baseline. Product engineers who evaluated the feature rated it 4.5 out of 5 for usefulness. The paper also reports the method scales linearly and generalizes to public benchmarks.

What carries the argument

The monitor entity graph is a heterogeneous graph with three node types (monitors, metrics, dimensions) and three relationship types (monitor-dimension, metric-dimension, monitor-metric). On this graph, DiRecGNN runs transformer-style graph convolutions: multi-head attention weights are computed between a node and its typed neighbors, and edge-aware message transformation aggregates neighbor value vectors. To capture global structure beyond the local neighborhood, the model samples random walks with restart from each monitor and applies a separate attention mechanism over the sequence of nodes in each path (Random Walk Attention, RWA), using position embeddings. The training objective combin

What would settle it

Ask domain experts to independently specify the ideal dimension set for a sample of production monitors, and check whether DiRecGNN's top-k recommendations match the expert choices more often than the historical assignments; alternatively, compare incident-detection and false-alert rates between monitors created with the recommendations and monitors created without them.

Watch

Extended reading notes

Core claim

The central claim is that the optimal subset of dimensions for a cloud monitor can be learned from historical interactions in a monitor entity graph, and that the right architecture for this task is a transformer-style heterogeneous GNN. The model treats monitors, metrics, and dimensions as typed nodes with three edge types, applies multi-head attention with edge-aware message passing to aggregate local neighborhoods, and additionally samples random walks from each target node and applies attention over the path sequences to capture long-range dependencies that local message passing misses. A key component is the attention-alignment loss, which penalizes divergence of attention heads from th

Load-bearing premise

The training labels are the existing monitor-to-dimension edges—the very assignments the paper describes as unstructured and reactive—so the model is evaluated on how well it reproduces current practice, not on whether those practices are correct.

Editorial extensions

If this is right

  • If correct, monitor configuration shifts from manual trial-and-error to a data-driven recommendation step, reducing both missed incidents from incomplete dimension coverage and noise from redundant dimensions.
  • The combination of BCE, TOP1-max ranking, and attention-alignment losses provides a concrete recipe for training sparse-graph recommenders, where the alignment loss alone accounts for large gains.
  • The finding that random-walk attention especially helps low-degree nodes suggests that explicit long-range path modeling is necessary for sparse heterogeneous graphs, not just deeper message passing.
  • The reported linear scaling and sub-linear sampling indicate the approach can be deployed on very large production graphs without architectural changes.
  • The competitive results on public benchmarks suggest the architecture is not domain-specific and transfers to other heterogeneous graph tasks such as node classification and link prediction.

Reading between the lines

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

  • The evaluation treats the existing monitor-to-dimension edges as ground truth. If those assignments themselves are imperfect—which the paper's introduction says they are—the reported accuracy measures fidelity to current practice rather than quality of the resulting monitoring. A direct test would compare recommendations against expert-defined ideal configurations.
  • The user study's 4.5/5 rating likely reflects the general value of structured recommendations, not proof that these particular recommendations improve monitoring. A randomized deployment with incident-detection and false-alert rates as outcomes would separate those.
  • Since random-walk paths are sampled uniformly, a natural extension is to weight or learn the path sampling toward informative, sparsity-aware paths, which the paper itself mentions as future work; this could further help low-degree monitors.
  • If the architecture transfers to other sparse heterogeneous recommendation settings, the attention-alignment loss could become a reusable module for keeping attention sparse and consistent across heads.
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.

Referee Report

4 major / 6 minor

Summary. The paper introduces DiRecGNN, a heterogeneous graph neural network with transformer-style attention over random-walk paths, attention-alignment loss, and ranking loss, for the task of recommending monitoring dimensions for cloud monitors. The method is evaluated on a production monitor-entity graph with 18k monitors, 4.6k metrics, and 8.3k dimensions, reporting large gains over several baselines, and is complemented by a small user study. The central claim is that DiRecGNN achieves HR@1=0.597, MRR=0.714, NDCG@k=0.555, and Recall@5=0.748, corresponding to a 55.8% HR@1 improvement and a 43.1% MRR improvement over the best baseline, with an average user rating of 4.5/5.

Significance. If the reported results are correct, this is a meaningful industrial application of heterogeneous graph representation learning, with a real deployment and a user study. The paper also contributes an attention-alignment loss that is designed for sparse graphs and a random-walk attention module for long-range dependencies. The strengths are the production-scale dataset, the deployment insights, and the attempt at a controlled user study. However, the numerical claims in the abstract and Section 5.2 are compromised by internal inconsistencies in Table 2 and by ambiguity in the definition of the reported metrics. The label-validity concern about using existing monitor-dimension edges as ground truth is also not fully addressed, and this limits the conclusiveness of the recommendation-quality claim.

major comments (4)
  1. [Table 2, Section 5.2] The Hit Ratio values in Table 2 decrease with k for every method (e.g., DiRecGNN: HR@1=0.597, HR@3=0.265, HR@5=0.173; HGT: 0.396, 0.185, 0.131). This violates the standard cumulative definition of Hit Ratio, where a correct rank-1 recommendation is necessarily also a correct rank-3 and rank-5 recommendation, so HR@k must be non-decreasing in k. The paper defines HR@k in Section 5.1.2 without specifying a non-cumulative variant. Either the values reported are not cumulative Hit Ratios (e.g., they are precision or exact-rank metrics) or the computation is incorrect. Since the headline '55.8% improvement in HR@1' rests on this column, this must be clarified and corrected before the central quantitative claim can be accepted.
  2. [Section 5.2, Table 2, Abstract] The claimed improvements are not with respect to the best baseline in the table. Table 2 shows HGT achieving HR@1=0.396 and MRR=0.510, both higher than SAGE(v1)'s 0.383 and 0.499. Yet the text in Section 5.2 says the 'best baseline' is SAGE(v1) and reports a 55.8% HR@1 improvement; the abstract reports a 43.1% MRR increase. Relative to HGT, the HR@1 improvement is 0.597/0.396-1 = 50.8% and the MRR improvement is 0.714/0.510-1 = 40.0%. The inconsistency also appears between the caption (55.3%) and the text (55.8%). These numbers must be reconciled, and all relative improvements should be computed relative to the actual best baseline.
  3. [Section 1, Section 5.2, Table 2] The training and evaluation labels are the existing 'monitor associated with dimension' edges. Section 1 criticizes the existing approach to selecting dimensions as 'unstructured and reactive,' and Figure 2b shows that most monitors use only a subset of the emitted dimensions. If the historical assignments contain incomplete coverage or redundant dimension choices, the model is trained to reproduce those assignments. The evaluation then measures fidelity to current practice, not the quality of the recommended monitoring configuration. The paper should explicitly discuss this limitation and, if possible, provide evidence (e.g., incident-based validation or a qualitative analysis of recommended vs. non-recommended dimensions) that the learned recommendations correspond to better monitoring outcomes.
  4. [Section 4.2, Equations (2)-(3)] The description of the random walk path module is confusing. The text says 'for a given monitor node m_i' but then defines paths with a target node d_i and 'destination node type (dimension in this case).' It is unclear whether the paths are sampled from a monitor, a dimension, or both, and how the path embeddings are matched to the query node in Eq. (3). This makes the architecture hard to reproduce. Please clarify the node type from which the random walks start, the exact path schema, and the alignment between path sequences and the query embedding.
minor comments (6)
  1. [Abstract and Section 5.2] The percentage increases in HR@1 are stated as 55.8% in the text and 55.3% in the Table 2 caption. Please make these consistent.
  2. [Table 2] The abbreviation 'N@k' is not defined in Section 5.1.2. Please define the value of k used for NDCG and Recall, or report the full metric names and their cutoff values.
  3. [Section 5.1.2] The paper states that learning rate is reduced when validation loss does not improve for 5 consecutive epochs and early stopping uses patience 10, but the relation between these two mechanisms is unclear. Please clarify whether the scheduler is checked on validation loss or training loss.
  4. [Section 6.1] The user study reports that 10 engineers agreed out of 30 sampled, but the text earlier says 'approximately 30 users' and then '10 agreed.' Please clearly report the response rate and whether the 30 sampled participants all agreed to be interviewed; the current wording is inconsistent.
  5. [References] The reference list contains unresolved placeholders such as 'Conference acronym ’XX' in the copyright block and the ACM reference format. These should be replaced with the actual venue and year.
  6. [Section 4.2] The definition 'L = 2 + 4(l-1)' for path length is not used consistently in the experiments (path lengths are {2, 6, 10}). Please align the notation with the actual path lengths and the number of nodes in the path sequence.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the central claim is an empirical held-out edge-prediction result, and neither the equations nor the cited prior work reduce to the paper's own fitted outputs.

full rationale

The paper's derivation chain is an empirical ML pipeline: it builds a monitor-entity graph, defines a ranking model via message passing (Eq. 2), random-walk attention (Eq. 3), and a composite loss (Eq. 7), then evaluates on a held-out 10% edge split with fixed negatives (Section 5.1.2). The target quantities HR@k, MRR, NDCG, and Recall are computed on test edges that are not used to fit the model, so the reported improvements are not equal by construction to any training objective or fitted parameter. The only self-citations with author overlap are [2] (Bastos et al., used alongside external [1] to motivate why attention may be homophilic) and [22] (prior metric-recommendation work cited as related work); neither is invoked as a uniqueness theorem or as the source of the model's components, and removing them would not change the equations. The suspicious monotonic decrease of HR@k with k in Table 2 and the mismatch in the claimed 'best baseline' are internal-consistency/reporting concerns, not circularity: reporting a different statistic under the name HR@k does not make the prediction depend on its own input. No step in the derivation is self-definitional, no fitted value is renamed as a prediction, and no external benchmark result is imported from the authors' prior work. Hence the paper is self-contained with respect to the circularity question.

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

The model's performance rests mainly on hyperparameters and on the validity of using current monitor configurations as ground truth; no new entities are introduced.

free parameters (4)
  • lambda_al (attention-alignment strength) = not reported
    Weight on the alignment loss in Eq. 6; chosen by hand and never specified numerically; the model's sparsity behavior depends on it.
  • random walk samples W and path lengths L = W default not stated; L in {2,6,10}
    Number and length of sampled paths in the RWA module; ablation (Fig. 6) shows MRR varies with both.
  • negative sampling ratio = 2.0
    Ratio of negative to positive edges during dynamic negative sampling; affects ranking loss and final metrics.
  • dynamic loss balancing weights = not reported
    Weights of BCE/alignment/ranking losses are adjusted dynamically but the mechanism is not described; they affect every reported metric.
assumptions (4)
  • domain assumption Existing monitor-to-dimension association edges are a valid ground-truth target for learning optimal dimension selection
    Section 2.1/5.1: labels come from current monitors' chosen dimensions, even though the paper argues current selection is 'unstructured and reactive' (Section 1). If existing monitors are suboptimal, the model learns to reproduce that behavior.
  • domain assumption The monitor entity graph is static and complete for the training window
    Section 2.1: 'we assume a static graph.' Recommendations are generated from a snapshot; monitor changes over time are ignored.
  • domain assumption E5 text embeddings of metric/dimension names capture the semantic features needed for the task
    Section 5.1.1: node features are E5 embeddings; if the E5 model's semantics don't transfer to the cloud domain, the text-based inductive bias fails.
  • ad hoc to paper Self-attention over random-walk paths captures long-range dependencies beyond message passing
    Section 4.2: the RWA module is built on this premise; the ablation shows improvement but only on this proprietary dataset, so general validity is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attention Enhanced Entity Recommendation for Intelligent Monitoring in Cloud Systems." pith.science (2026). https://pith.science/paper/YD5EYJCG

@misc{pith2026251020640,
  author       = {Pith},
  title        = {Pith review of: Attention Enhanced Entity Recommendation for Intelligent Monitoring in Cloud Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YD5EYJCG}},
  note         = {Machine review of arXiv:2510.20640}
}
read the original abstract

In this paper, we present DiRecGNN, an attention-enhanced entity recommendation framework for monitoring cloud services at Microsoft. We provide insights on the usefulness of this feature as perceived by the cloud service owners and lessons learned from deployment. Specifically, we introduce the problem of recommending the optimal subset of attributes (dimensions) that should be tracked by an automated watchdog (monitor) for cloud services. To begin, we construct the monitor heterogeneous graph at production-scale. The interaction dynamics of these entities are often characterized by limited structural and engagement information, resulting in inferior performance of state-of-the-art approaches. Moreover, traditional methods fail to capture the dependencies between entities spanning a long range due to their homophilic nature. Therefore, we propose an attention-enhanced entity ranking model inspired by transformer architectures. Our model utilizes a multi-head attention mechanism to focus on heterogeneous neighbors and their attributes, and further attends to paths sampled using random walks to capture long-range dependencies. We also employ multi-faceted loss functions to optimize for relevant recommendations while respecting the inherent sparsity of the data. Empirical evaluations demonstrate significant improvements over existing methods, with our model achieving a 43.1% increase in MRR. Furthermore, product teams who consumed these features perceive the feature as useful and rated it 4.5 out of 5.

Figures

Figures reproduced from arXiv: 2510.20640 by the authors.

Figure 1
Figure 1. a) Monitor entity ranking problem: Nodes represent [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. a) Variation in jaccard similarity of set of dimen [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Characteristics of the Monitor Entity Graph: a) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The overall architecture of DiRecGNN framework. The framework uses an enhanced transformer-style graph convo [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: a) Impact of Ranking Loss on model performance, [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on the RWA module: a) The variation [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: a) The performance gains (↑ with sparsity) by using RWA module w.r.t node degree b) Plot of percent nodes sam￾pled vs time taken indicating linear scaling. this homophilic (neighborhood) information is already effectively captured by the message passing framework. Stud…
Figure 8
Figure 8. Figure 8: a) Attention map between monitor and dimension [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Example recommendation of dimensions for a mon [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 6 linked inside Pith

  1. [1]

    Muhammet Balcilar, Guillaume Renton, Pierre Héroux, Benoit Gaüzère, Sébastien Adam, and Paul Honeine. 2021. Analyzing the Expressive Power of Graph Neural Networks in a Spectral Perspective. InInternational Conference on Learning Representations. https://openreview.net/forum?id=-qh0M9XWxnv

    2021

  2. [2]

    Anson Bastos, Abhishek Nadgeri, Kuldeep Singh, Hiroki Kanezashi, Toyotaro Suzumura, and Isaiah Onando Mulang’. 2022. How Expressive are Transformers in Spectral Domain for Graphs?Transactions on Machine Learning Research (2022). https://openreview.net/forum?id=aRsLetumx1

    2022

  3. [3]

    Fedor Borisyuk, Shihai He, Yunbo Ouyang, Morteza Ramezani, Peng Du, Xiaochen Hou, Chengming Jiang, Nitin Pasumarthy, Priya Bannur, Birjodh Tiwana, et al

  4. [4]

    Ivan Cantador, Peter Brusilovsky, and Tsvi Kuflik. 2011. Second workshop on information heterogeneity and fusion in recommender systems (HetRec2011). InProceedings of the Fifth ACM Conference on Recommender Systems(Chicago, Illinois, USA)(RecSys ’11). Association for Computing Machinery, New York, NY, USA, 387–388. doi:10.1145/2043932.2044016

    arXiv 2011

  5. [5]

    Zhuangbin Chen, Yu Kang, Liqun Li, Xu Zhang, Hongyu Zhang, Hui Xu, Yangfan Zhou, Li Yang, Jeffrey Sun, Zhangwei Xu, et al. 2020. Towards intelligent incident management: why we need it and how we make it. InProceedings of the 28th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering. 1487–1497

    2020

  6. [6]

    Wei-Lin Chiang, Xuanqing Liu, Si Si, Yang Li, Samy Bengio, and Cho-Jui Hsieh

  7. [7]

    Xinyu Fu, Jiani Zhang, Ziqiao Meng, and Irwin King. 2020. Magnn: Metap- ath aggregated graph neural network for heterogeneous graph embedding. In Proceedings of the web conference 2020. 2331–2341

    2020

  8. [8]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems30 (2017)

    2017

Show all 44 references
  1. [9]

    Balázs Hidasi and Alexandros Karatzoglou. 2018. Recurrent neural networks with top-k gains for session-based recommendations. InProceedings of the 27th ACM international conference on information and knowledge management. 843–852

    2018

  2. [10]

    Huiting Hong, Hantao Guo, Yucheng Lin, Xiaoqing Yang, Zang Li, and Jieping Ye. 2020. An attention-based graph neural network for heterogeneous structural learning. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 4132–4139

    2020

  3. [11]

    Binbin Hu, Zhiqiang Zhang, Chuan Shi, Jun Zhou, Xiaolong Li, and Yuan Qi

  4. [12]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. 2020. Heterogeneous graph transformer. InProceedings of the web conference 2020. 2704–2710

    2020

  5. [13]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. 2020. Heterogeneous Graph Transformer. InProceedings of The Web Conference 2020(Taipei, Tai- wan)(WWW ’20). Association for Computing Machinery, New York, NY, USA, 2704–2710. doi:10.1145/3366423.3380027

    2020

  6. [14]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Cash-out user detection based on attributed heterogeneous information network with a hierarchical attention mechanism. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 33. 946–953

  7. [15]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907(2016)

    2016 arXiv

  8. [16]

    Qingsong Lv, Ming Ding, Qiang Liu, Yuxiang Chen, Wenzheng Feng, Siming He, Chang Zhou, Jianguo Jiang, Yuxiao Dong, and Jie Tang. 2021. Are we really making much progress? Revisiting, benchmarking, and refining heterogeneous graph neural networks. arXiv:2112.14936 [cs.LG] https...

    2021 arXiv

  9. [17]

    Wenbing Huang, Tong Zhang, Yu Rong, and Junzhou Huang. 2018. Adaptive sam- pling towards fast graph representation learning.Advances in neural information processing systems31 (2018)

    2018

  10. [18]

    Vinod Nair, Ameya Raul, Shwetabh Khanduja, Vikas Bahirwani, Qihong Shao, Sundararajan Sellamanickam, Sathiya Keerthi, Steve Herbert, and Sudheer Dhuli- palla. 2015. Learning a hierarchical monitoring system for detecting and di- agnosing service issues. InProceedings of the 21...

    2015

  11. [19]

    Rumelhart, Geoffrey E

    David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. 1986. Learning representations by back-propagating errors.Nature323, 6088 (01 Oct 1986), 533–536. doi:10.1038/323533a0

    1986 doi

  12. [20]

    Jesús Montes, Alberto Sánchez, Bunjamin Memishi, María S Pérez, and Gabriel Antoniu. 2013. GMonE: A complete approach to cloud monitoring.Future Generation Computer Systems29, 8 (2013), 2026–2040

    2013

  13. [21]

    Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjin Wang, and Yu Sun. 2020. Masked label prediction: Unified message passing model for semi- supervised classification.arXiv preprint arXiv:2009.03509(2020)

    2020 arXiv

  14. [22]

    Pooja Srinivas, Fiza Husain, Anjaly Parayil, Ayush Choure, Chetan Bansal, and Saravan Rajmohan. 2024. Intelligent Monitoring Framework for Cloud Services: A Data-Driven Approach. InProceedings of the 46th International Conference on Software Engineering: Software Engineering i...

    2024

  15. [23]

    Aravind Sankar, Yozen Liu, Jun Yu, and Neil Shah. 2021. Graph neural networks for friend ranking in large-scale social platforms. InProceedings of the Web Conference 2021. 2535–2546

    2021

  16. [24]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 [cs.CL] https://arxiv.org/abs/1706.03762

    2023 arXiv

  17. [25]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. 2017. Graph attention networks.stat1050, 20 (2017), 10–48550

    2017

  18. [26]

    Chellammal Surianarayanan and Pethuru Raj Chelliah. 2019. Essentials of Cloud Computing.Cham: Springer International Publishing(2019)

    2019

  19. [27]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533(2022)

    2022 arXiv

  20. [28]

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu

  21. [29]

    Hulong Wang, Zesheng Shen, Shuzhen Jiang, Guang Sun, and Ren-Jie Zhang

  22. [30]

    Carl Yang, Aditya Pal, Andrew Zhai, Nikil Pancha, Jiawei Han, Charles Rosenberg, and Jure Leskovec. 2020. MultiSage: Empowering GCN with contextualized multi- embeddings on web-scale multipartite networks. InProceedings of the 26th ACM SIGKDD international conference on knowle...

    2020

  23. [31]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 974–983

    2018

  24. [32]

    Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim

  25. [33]

    InThe world wide web conference

    Heterogeneous graph attention network. InThe world wide web conference. 2022–2032

    2022

  26. [34]

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. 2019. Heterogeneous Graph Attention Network. InThe World Wide Web Conference(San Francisco, CA, USA)(WWW ’19). Association for Computing Machinery, New York, NY, USA, 2022–2032. doi:10.1145/330855...

    2019

  27. [35]

    Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V. Chawla. 2019. Heterogeneous Graph Neural Network. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (Anchorage, AK, USA)(KDD ’19). Association for Computing M...

    2019

  28. [36]

    Muhan Zhang, Zhicheng Cui, Marion Neumann, and Yixin Chen. 2018. An end- to-end deep learning architecture for graph classification. InProceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  29. [37]

    Jianan Zhao, Xiao Wang, Chuan Shi, Binbin Hu, Guojie Song, and Yanfang Ye

  30. [38]

    Graph transformer networks.Advances in neural information processing systems32 (2019)

    2019

  31. [39]

    Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. 2019. Graphsaint: Graph sampling based inductive learning method. arXiv preprint arXiv:1907.04931(2019)

    2019 arXiv

  32. [40]

    Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V Chawla. 2019. Heterogeneous graph neural network. InProceedings of the 25th Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Fiza Hussain, Anson Bastos, Anjaly Parayil, Ayush Choure, Chetan Bansal, Ruj...

    2019

  33. [44]

    In Proceedings of the AAAI conference on artificial intelligence, Vol

    Heterogeneous graph structure learning for graph neural networks. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 4697–4705

  34. [2019]

    InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining

    Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 257–266

  35. [2021]

    Journal of Physics: Conference Series1757, 1 (jan 2021), 012168

    User-based Collaborative Filtering Algorithm Design and Implementation. Journal of Physics: Conference Series1757, 1 (jan 2021), 012168. doi:10.1088/1742- 6596/1757/1/012168

    2021 doi

  36. [2024]

    InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Lignn: Graph neural networks at linkedin. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4793–4803

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.