Pith. sign in

REVIEW 5 major objections 6 minor 44 references

Large Scalable Cross-Domain Graph Neural Networks for Personalized Notification at LinkedIn

T0 review · 5 major / 6 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 A single cross-domain graph of 100 billion edges trains a GNN that lifts LinkedIn notification CTR by 0.62% in production.

desk verdict An industrial cross-domain GNN at record scale with genuine deployment evidence, but the online A/B reporting is too thin to verify the headline lifts and the paper has several unforced errors. read the letter →

arxiv 2506.12700 v1 pith:AVMGHGGR submitted 2025-06-15 cs.LG

Shihai He , Julie Choi , Tianqi Li , Zhiwei Ding , Peng Du , Priya Bannur , Franco Liang , Fedor Borisyuk
show 3 more authors
Padmini Jaikumar Xiaobing Xue Viral Gupta
This is my paper · ORCID
classification cs.LG
keywords cross-domaingraphneuralnetworknotificationrecommendationheterogeneoustemporalmodelingmulti-tasklearningproductionrecommendersystemclick-throughratepredictionlarge-scale
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 claims that merging the professional network's notification, feed, email, and other interaction surfaces into one heterogeneous graph—8.6 billion nodes and over 100 billion edges—lets a graph neural network learn member and content representations that outperform models trained on any single domain. It further claims that adding temporal edge encoding and multi-task learning improves click and deeper-engagement predictions. Deployed in the notification ranking path, the cross-domain embeddings raised in-app click-through rate by 0.62%, push CTR by 0.30%, and weekly active users by 0.10%, all reported as statistically significant. The value of the claim is that one unified graph and one set of embeddings can replace per-domain modeling across a large industrial platform.

What carries the argument

The load-bearing object is the unified heterogeneous graph built from three edge categories: engagement edges (member clicked content), affinity edges (member engaged with another member's content), and intrinsic edges (member has title; notification mentions post). On top of it sits a two-tower architecture: a member encoder that combines heterogeneous GraphSAGE-style message passing with a transformer over the member's chronologically recent notification interactions, and an item encoder that runs message passing around the target notification. Edge timestamps are embedded with Time2Vec, a learnable time embedding with linear and sinusoidal components, and multi-task learning uses Multi-gate Mixture-of-Experts with shared and task-specific experts, each expert being an entity encoder, with gating networks choosing a weighted combination per task. The graph construction and temporal partitioning (35-day graph, 14-day training, 7-day validation) are what make the cross-domain signal usable without leakage.

What would settle it

Obtain the experiment logs and check pre-experiment covariate balance on member activity, CTR, and notification volume between the 6% treatment and 8% control groups; if the groups differ materially, or if the +0.62% CTR and +0.10% WAU lifts disappear after adjusting for those covariates, the central deployment claim is refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a Cross-domain GNN trained on a unified graph of member, content, and company nodes—with engagement, affinity, and intrinsic edge types—beats a domain-specific GNN baseline by +8.089% offline AUC on notification prediction. Temporal modeling contributes +0.265% AUC on notification click prediction over a non-temporal Sage model, and multi-task learning adds +0.063% AUC for clicks and +0.327% AUC for professional interactions. In the online A/B test, the GNN-powered member embeddings integrated into the second-pass ranker produced a +0.62% lift in in-app CTR, +0.30% in push CTR, +0.10% in weekly active users, and +0.07% in sessions. The authors attribute the gains to message passing across domains propagating preferences that single-domain graphs cannot see.

Load-bearing premise

The online result stands on the assumption that the 6% treatment and 8% control traffic groups are exchangeable; the paper lists balanced dimensions but gives no randomization protocol, covariate balance table, or confidence intervals, so unmeasured differences in member activity could produce the small observed lifts.

Editorial extensions

If this is right

  • If the offline AUC lift is real, platform teams can consolidate per-domain models into one graph, cutting duplicated feature engineering and training cost.
  • The temporal edge design means the model is evaluated on future interactions, so deploying it should generalize to new notifications rather than memorizing past sends.
  • Multi-task gains are larger for professional interactions than for clicks, suggesting the shared graph representation helps hardest for deeper engagement objectives.
  • Because member embeddings are refreshed daily and stored in a feature store, the same representation can be reused by email ranking, feed ranking, and other surfaces beyond notifications.

Reading between the lines

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

  • The reported +8.089% offline lift compares cross-domain training against domain-specific training; if the baseline used less data or fewer parameters, part of the gain may be data scale rather than graph structure. A controlled ablation holding architecture and compute fixed would separate the two.
  • The affinity edges may be the main channel for cold-start members with sparse click histories; a testable extension is to measure embedding quality for members with few direct engagement edges, where cross-domain propagation should matter most.
  • The online lifts are small enough that allocation noise is a real alternative explanation; the strongest follow-up is a rerandomized A/B with covariate adjustment on pre-experiment CTR and activity, reported with confidence intervals.
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

5 major / 6 minor

Summary. The paper describes a cross-domain graph neural network system for personalized notification at LinkedIn. It constructs a unified heterogeneous graph spanning notifications, feed, email, and other surfaces, and proposes a two-tower GNN architecture with heterogeneous message passing, temporal aggregation via Time2Vec, and MMoE-style multi-task learning. The authors report offline AUC lifts for cross-domain integration, temporal modeling, and multi-task learning, and an online A/B test with improvements in Sessions, WAU, CTR, and professional interaction metrics. The central claim is that cross-domain GNN embeddings significantly improve notification personalization in production, with headline results of +0.62% CTR and +0.10% WAU.

Significance. If fully supported, this is a valuable industrial-scale demonstration that cross-domain GNNs can be deployed for notification ranking at the scale of billions of nodes and edges. The paper provides a concrete graph-construction framework, a production pipeline with daily embedding refresh and low-latency serving, and a rare online A/B evaluation. The claimed effects are modest but potentially meaningful at LinkedIn's scale, and the architectural choices (temporal modeling, MTL) are sensible. However, the evidence as currently reported is incomplete: the online A/B design is not fully specified, offline baselines are under-described, and there are internal numerical inconsistencies. The central claim is plausible but not yet fully verifiable from the manuscript.

major comments (5)
  1. [Section 6.3, Table 5] The randomization unit in the online A/B test is incompatible with the reported user-level metrics. The text states that 6% of notification traffic was allocated to treatment and 8% to control, yet WAU and Sessions are member-level metrics while CTR is per-notification. The manuscript does not state whether randomization was at the notification level or the member level, does not define the evaluation population, and does not provide a covariate balance table or confidence intervals. If allocation is notification-level, members can appear in both treatment and control, making a clean user-level WAU/Sessions comparison impossible; if allocation is member-level, the paper must say so and report balance. As written, the claimed p<0.01 for the +0.10% WAU and +0.62% CTR lifts cannot be checked, and the headline results are not interpretable.
  2. [Section 3.4] The section states that the graph data are partitioned into 'four distinct periods' but then lists only three: Graph Construction Period (35 days), Training Data Period (14 days), and Validation Data Period (7 days). No test period is described here, while Section 6.1.2 refers to an 80/10/10 temporal split. This inconsistency makes the offline evaluation protocol and the leakage-prevention claims unclear; please reconcile the partition description with the actual evaluation setup.
  3. [Section 6.2, Tables 3 and 4] The offline AUC lifts are reported as point estimates with no error bars, no number of runs, no dataset sizes, and no description of the 'domain-specific' baseline architecture, hyperparameters, training data, or label definitions. The +8.089% lift in Table 3 is much larger than typical AUC differences in CTR prediction and could be driven by differences in data domains, label distributions, or model capacity rather than by cross-domain signal alone. Please specify the baseline in detail and report repeated-run variability or confidence intervals.
  4. [Section 3.1 and Tables 1-2] The node and edge counts in Tables 1 and 2 do not sum to the totals stated in the text. The listed node counts sum to approximately 7.32B, not the stated 8.6B, and the listed edge counts sum to approximately 86.2B, not 'over 100 billion edges' as claimed in the Introduction and Section 3. Please reconcile the totals or clarify which node/edge types are omitted from the tables.
  5. [Section 6.3.3] The statement that 'all improvements are statistically significant (p<0.01)' is unsupported because no confidence intervals, standard errors, test statistics, or multiple-comparison corrections are provided. Given the small percentage lifts, the reader cannot distinguish a genuine effect from sampling variation. Please report the exact test used and interval estimates for the metrics in Table 5.
minor comments (6)
  1. [Equation (7)] The final loss uses summation index 'i' while the weights are denoted alpha_k; the notation should be consistent (e.g., sum over k of alpha_k * L_k).
  2. [Reference [5]] The LiGNN reference lists placeholder authors 'J. Doe, M. Smith, and Y. Li'; the actual author list and publication details should be provided before publication.
  3. [Figure 2 caption] The caption contains a typo: 'Tnotifciation' should be 'Notification'.
  4. [Section 6.3.4] The plus signs in the segment-level results are inconsistent: 'strong 0.10% Sessions Gain' lacks a plus sign, while 'WAU: +0.23%, 1.16% CTR' uses a plus sign on one metric but not the other. Please standardize.
  5. [Section 3.1] There is a missing period in the sentence '...over a longer period For example, ...'; please insert the period after 'period'.
  6. [Section 7.2] The claim of 'approximately 32 GPU-days per training cycle' should specify the GPU model and batch configuration to be informative.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all central claims are empirical comparisons against independent baselines; self-citations are contextual, not load-bearing.

full rationale

The paper's derivation chain is not circular. The central claims—cross-domain GNN improves CTR and WAU, temporal modeling adds AUC, and MTL adds AUC—are supported by measured comparisons against baselines: Table 3 compares cross-domain versus domain-specific graphs, Table 4 compares temporal versus non-temporal and MTL versus single-task models, and Table 5 reports a production A/B test against a control system without GNN embeddings. None of these metrics is defined in terms of the model's own outputs, and no fitted parameter is relabeled as a prediction. The temporal module is adapted from LiGNN [5], which is self-referential in the sense that the paper calls it 'our earlier successes,' but the paper does not use that citation to prove effectiveness; it reports controlled experiments (Table 4) and an online A/B test (Table 5). The MMoE, PLE, and Time2Vec components are cited to external prior work. The validity of the A/B lift cannot be fully verified from the text because no covariate balance table, randomization protocol, or confidence intervals are provided, but that is a statistical-reporting and experimental-design concern, not circularity. No step reduces to its own input by construction.

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

No new physical or conceptual entities are introduced; the system uses established ML components. The ledger captures the design choices and assumptions that the empirical claims depend on.

free parameters (4)
  • Task importance weights α_k = not reported
    Equation (7) weights each task loss; these are tuned to prioritize tasks but no values or tuning procedure are given.
  • Time2Vec frequency and phase parameters = learned
    Equation (3) defines learnable temporal encoding parameters; the final learned values are not reported.
  • GNN depth and hidden dimensions = not reported
    The paper gives no layer count, hidden size, or number of experts, yet these capacity choices affect the reported lifts.
  • Temporal partition lengths = 35, 14, and 7 days
    Section 3.4 chooses graph construction, training, and validation window sizes by hand; changing them could change offline results.
assumptions (3)
  • domain assumption The unified graph with engagement, affinity, and intrinsic edge types is a sufficient and unbiased representation for notification personalization.
    Section 3.1 introduces these edge categories without an ablation; if important signal is missing, message passing cannot recover it.
  • domain assumption The temporal data partition prevents label leakage and gives a realistic offline estimate.
    Sections 3.3 to 3.4 claim edges are sampled before label timestamps, but the section says 'four distinct periods' then lists only three, leaving the test period undefined.
  • domain assumption The 6% versus 8% traffic split in the A/B test is free of allocation bias.
    Section 6.3.1 lists controlled dimensions but provides no balance diagnostics; asymmetric traffic allocation can introduce confounds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Scalable Cross-Domain Graph Neural Networks for Personalized Notification at LinkedIn." pith.science (2026). https://pith.science/paper/AVMGHGGR

@misc{pith2026250612700,
  author       = {Pith},
  title        = {Pith review of: Large Scalable Cross-Domain Graph Neural Networks for Personalized Notification at LinkedIn},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AVMGHGGR}},
  note         = {Machine review of arXiv:2506.12700}
}
read the original abstract

Notification recommendation systems are critical to driving user engagement on professional platforms like LinkedIn. Designing such systems involves integrating heterogeneous signals across domains, capturing temporal dynamics, and optimizing for multiple, often competing, objectives. Graph Neural Networks (GNNs) provide a powerful framework for modeling complex interactions in such environments. In this paper, we present a cross-domain GNN-based system deployed at LinkedIn that unifies user, content, and activity signals into a single, large-scale graph. By training on this cross-domain structure, our model significantly outperforms single-domain baselines on key tasks, including click-through rate (CTR) prediction and professional engagement. We introduce architectural innovations including temporal modeling and multi-task learning, which further enhance performance. Deployed in LinkedIn's notification system, our approach led to a 0.10% lift in weekly active users and a 0.62% improvement in CTR. We detail our graph construction process, model design, training pipeline, and both offline and online evaluations. Our work demonstrates the scalability and effectiveness of cross-domain GNNs in real-world, high-impact applications.

Figures

Figures reproduced from arXiv: 2506.12700 by the authors.

Figure 1
Figure 1. Our generalizable framework to collect all available [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. High-level architecture of the member encoder (left) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. High-level architecture of the Cross-domain GNN, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Production pipeline for the cross-domain GNN: the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Professional interactions that can occur after mem [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 29 canonical work pages

  1. [5]

    J. Doe, M. Smith, and Y. Li. LiGNN: Graph Neural Networks at LinkedIn. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2024

    work page 2024

  2. [1]

    Graphert–transformers-based temporal dynamic graph embedding

    Moran Beladev, Gilad Katz, Lior Rokach, Uriel Singer, and Kira Radinsky. Graphert–transformers-based temporal dynamic graph embedding. InProceed- ings of the 32nd ACM International Conference on Information and Knowledge Management, pages 68–77, 2023

    work page 2023

  3. [2]

    Graph representation learning for multi-task settings: A meta-learning approach

    Davide Buffelli and Fabio Vandin. Graph representation learning for multi-task settings: A meta-learning approach. InInternational Conference on Learning Representations, 2021

    work page 2021

  4. [3]

    Multitask learning.Machine learning, 28:41–75, 1997

    Rich Caruana. Multitask learning.Machine learning, 28:41–75, 1997

    work page 1997

  5. [4]

    Personalized pagerank graph attention networks

    Julie Choi. Personalized pagerank graph attention networks. InICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3578–3582, 2022

    work page 2022

  6. [6]

    Graph neural networks for social recommendation

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin. Graph neural networks for social recommendation. InThe world wide web conference, pages 417–426, 2019

    2019

  7. [7]

    A comprehensive survey of dynamic graph neural networks: Models, frame- works, benchmarks, experiments and challenges.arXiv preprint arXiv:2405.00476, 2024

    ZhengZhao Feng, Rui Wang, TianXing Wang, Mingli Song, Sai Wu, and Shuibing He. A comprehensive survey of dynamic graph neural networks: Models, frame- works, benchmarks, experiments and challenges.arXiv preprint arXiv:2405.00476, 2024

    arXiv 2024

  8. [8]

    A new model for learning in graph domains

    Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. InProceedings. 2005 IEEE international joint conference on neural networks, 2005., volume 2, pages 729–734. IEEE, 2005

    2005

Show all 44 references
  1. [9]

    Conet: Collaborative cross networks for cross-domain recommendation.arXiv preprint arXiv:1804.06769v3, 2018

    Qiang Yang Guangneng Hu, Yu Zhang. Conet: Collaborative cross networks for cross-domain recommendation.arXiv preprint arXiv:1804.06769v3, 2018

    2018 arXiv

  2. [10]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  3. [11]

    Dynamic neural networks: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(11):7436–7456, 2021

    Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(11):7436–7456, 2021

    2021

  4. [12]

    Multi-task learning on graphs with node and graph level labels

    Chester Holtz, Onur Atan, Ryan Carey, and Tushit Jain. Multi-task learning on graphs with node and graph level labels. InNeurIPS Workshop on Graph Representation Learning, 2019

    2019

  5. [13]

    Deepfm: a factorization-machine based neural network for ctr prediction.arXiv preprint arXiv:1703.04247v1, 2017

    Yunming Ye Zhenguo Li Xiuqiang He Huifeng Guo, Ruiming Tang. Deepfm: a factorization-machine based neural network for ctr prediction.arXiv preprint arXiv:1703.04247v1, 2017

    2017 arXiv

  6. [14]

    Recommendation through mixture-of-design- experts

    Wang-Cheng Kang and et al. Recommendation through mixture-of-design- experts. InProceedings of the 13th ACM Conference on Recommender Systems (RecSys), 2019

    2019

  7. [15]

    Time2vec: Learning a vector representation of time.arXiv preprint arXiv:1907.05321, 2019

    Seyed Mehran Kazemi, Rishab Goel, Sepehr Eghbali, Janahan Ramanan, Jaspreet Sahota, Sanjay Thakur, Stella Wu, Cathal Smyth, Pascal Poupart, and Marcus Brubaker. Time2vec: Learning a vector representation of time.arXiv preprint arXiv:1907.05321, 2019

    1907 arXiv

  8. [16]

    Kipf and Max Welling

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

    2017

  9. [17]

    Adatt: Adaptive task-to-task fusion network for multitask learning in recommendations

    Danwei Li, Zhengyu Zhang, Siyang Yuan, Mingze Gao, Weilin Zhang, Chaofei Yang, Xi Liu, and Jiyan Yang. Adatt: Adaptive task-to-task fusion network for multitask learning in recommendations. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,...

    2023

  10. [18]

    A survey of graph neural network based recommendation in social networks.Neurocomputing, 549:126441, 2023

    Xiao Li, Li Sun, Mengjie Ling, and Yan Peng. A survey of graph neural network based recommendation in social networks.Neurocomputing, 549:126441, 2023

    2023

  11. [19]

    Tgtod: A global temporal graph transformer for outlier detection at scale.arXiv preprint arXiv:2412.00984, 2024

    Kay Liu, Jiahao Ding, MohamadAli Torkamani, and Philip S Yu. Tgtod: A global temporal graph transformer for outlier detection at scale.arXiv preprint arXiv:2412.00984, 2024

    2024 arXiv

  12. [20]

    Linksage: Optimizing job matching using graph neural networks.arXiv preprint arXiv:2402.13430, 2024

    Ping Liu, Haichao Wei, Xiaochen Hou, Jianqiang Shen, Shihai He, Kay Qianqi Shen, Zhujun Chen, Fedor Borisyuk, Daniel Hewlett, Liang Wu, et al. Linksage: Optimizing job matching using graph neural networks.arXiv preprint arXiv:2402.13430, 2024

    2024 arXiv

  13. [21]

    Learning multiple tasks with multilinear relationship networks.Advances in neural information processing systems, 30, 2017

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Philip S Yu. Learning multiple tasks with multilinear relationship networks.Advances in neural information processing systems, 30, 2017

    2017

  14. [22]

    Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities.arXiv preprint arXiv:2302.01018, 2023

    Antonio Longa, Veronica Lachi, Gabriele Santin, Monica Bianchini, Bruno Lepri, Pietro Lio, Franco Scarselli, and Andrea Passerini. Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities.arXiv preprint arXiv:2302.01018, 2023

    2023 arXiv

  15. [23]

    Modeling task relationships in multi-task learning with multi-gate mixture-of-experts

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930–1939, 2018

    1930

  16. [24]

    Trans- former for graphs: An overview from architecture perspective.arXiv preprint arXiv:2202.08455, 2022

    Erxue Min, Runfa Chen, Yatao Bian, Tingyang Xu, Kangfei Zhao, Wenbing Huang, Peilin Zhao, Junzhou Huang, Sophia Ananiadou, and Yu Rong. Trans- former for graphs: An overview from architecture perspective.arXiv preprint arXiv:2202.08455, 2022

    2022 arXiv

  17. [25]

    Cross- stitch networks for multi-task learning

    Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross- stitch networks for multi-task learning. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3994–4003, 2016

    2016

  18. [26]

    Herograph: A heterogeneous graph framework for multi-target cross-domain recommendation.In ORSUM@ RecSys, 2020

    Yafeng Zhang Qing Zhang Qiang Cui, Tao Wei. Herograph: A heterogeneous graph framework for multi-target cross-domain recommendation.In ORSUM@ RecSys, 2020

    2020

  19. [27]

    Temporal graph networks for deep learning on dynamic graphs, 2020

    Emanuele Rossi and et al. Temporal graph networks for deep learning on dynamic graphs, 2020

    2020

  20. [28]

    Latent multi-task architecture learning

    Sebastian Ruder, Joachim Bingel, Isabelle Augenstein, and Anders Søgaard. Latent multi-task architecture learning. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 4822–4829, 2019

    2019

  21. [29]

    The graph neural network model.IEEE transactions on neural networks, 20(1):61–80, 2008

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model.IEEE transactions on neural networks, 20(1):61–80, 2008

    2008

  22. [30]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. InEuropean Semantic Web Conference (ESWC), 2018

    2018

  23. [31]

    A survey of graph neural networks for social recommender systems.ACM Computing Surveys, 56(10):1–34, 2024

    Kartik Sharma, Yeon-Chang Lee, Sivagami Nambi, Aditya Salian, Shlok Shah, Sang-Wook Kim, and Srijan Kumar. A survey of graph neural networks for social recommender systems.ACM Computing Surveys, 56(10):1–34, 2024

    2024

  24. [32]

    Deep learning based recommender system: A survey and new perspectives.arXiv preprint arXiv:1707.07435v7, 2019

    Aixin Sun Yi Tay Shuai Zhang, Lina Yao. Deep learning based recommender system: A survey and new perspectives.arXiv preprint arXiv:1707.07435v7, 2019

    2019 arXiv

  25. [33]

    Foundations and model- ing of dynamic networks using dynamic graph neural networks: A survey.iEEE Access, 9:79143–79168, 2021

    Joakim Skarding, Bogdan Gabrys, and Katarzyna Musial. Foundations and model- ing of dynamic networks using dynamic graph neural networks: A survey.iEEE Access, 9:79143–79168, 2021

    2021

  26. [34]

    Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recom- mendations

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recom- mendations. InProceedings of the 14th ACM conference on recommender systems, pages 269–278, 2020

    2020

  27. [35]

    Graph attention networks

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. InInternational Conference on Learning Representations (ICLR), 2018

    2018

  28. [36]

    Multi- domain recommendation with embedding disentangling and domain alignment

    Weiwen Liu Reynold Cheng Rui Zhang Bo Tang Wentao Ning, Xiao Yan. Multi- domain recommendation with embedding disentangling and domain alignment. arXiv preprint arXiv:2308.05508v2, 2023

    2023 arXiv

  29. [37]

    Minet: Mixed interest network for cross-domain click-through rate prediction.arXiv preprint arXiv:2008.02974v1, 2020

    Lei Zhao Jinmei Luo Yu Zhang Heng Zou Zhaojie Liu Yanlong Du Wentao Ouyang, Xiuwu Zhang. Minet: Mixed interest network for cross-domain click-through rate prediction.arXiv preprint arXiv:2008.02974v1, 2020

    2008 arXiv

  30. [38]

    A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems, 32(1):4–24, 2020

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems, 32(1):4–24, 2020

    2020

  31. [39]

    Heterogeneous graph-based framework with disentangled rep- resentations learning for multi-target cross domain recommendation.arXiv preprint arXiv:2407.00909v2, 2024

    Chongqi Ren Shenghui Xu Zhaoming Pan Zhimin Zhang Xiaopeng Liu, Juan Zhang. Heterogeneous graph-based framework with disentangled rep- resentations learning for multi-target cross domain recommendation.arXiv preprint arXiv:2407.00909v2, 2024

    2024 arXiv

  32. [40]

    A multi-task representation learning architecture for enhanced graph classification.Frontiers in Neuroscience, 13:1395, 2020

    Xiang Xie, Haoran Hou, Mengting Li, and Jian Tang. A multi-task representation learning architecture for enhanced graph classification.Frontiers in Neuroscience, 13:1395, 2020

    2020

  33. [41]

    Inductive representation learning on temporal graphs, 2020

    Da Xu, Dragomir Radev, Yu Meng, Jiaxin Huang, Ahmed Ejaz, and Yuzhou Wang. Inductive representation learning on temporal graphs, 2020

    2020

  34. [42]

    Hamil- ton, and Jure Leskovec

    Rex Ying, Ruining He, Kaifeng Chen, Pimnapat Eksombatchai, William L. Hamil- ton, and Jure Leskovec. Graph convolutional neural networks for web-scale recommender systems. InProceedings of the 24th ACM SIGKDD Conference on Knowledge Discovery & Data Mining (KDD), 2018. Large S...

    2018

  35. [43]

    Graphstorm an easy-to-use and scalable graph neural network framework: From beginners to heroes

    Jian Zhang, Da Zheng, Xiang Song, Theodore Vasiloudis, Israt Nisa, and Jim Lu. Graphstorm an easy-to-use and scalable graph neural network framework: From beginners to heroes. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5790–5791, 2023

    2023

  36. [44]

    Gigl: Large-scale graph neural networks at snapchat.arXiv preprint arXiv:2502.15054, 2025

    Tong Zhao, Yozen Liu, Matthew Kolodner, Kyle Montemayor, Elham Ghazizadeh, Ankit Batra, Zihao Fan, Xiaobin Gao, Xuan Guo, Jiwen Ren, et al. Gigl: Large-scale graph neural networks at snapchat.arXiv preprint arXiv:2502.15054, 2025

    2025 arXiv

Pith tools

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