Pith. sign in

REVIEW 5 major objections 6 minor 44 references

Efficient Recommendations via Graph Coarsening and Label Propagation

T0 review · 5 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Coarsening a telecom user graph into family-like communities and then running two-stage label propagation lifts NDCG@5 by 24% over full-graph propagation, and by over 50% when a GNN is used in the first stage.

desk verdict A practical telecom recommender pipeline that combines business-rule coarsening with two-stage label propagation; the ideas are sound and the reported gains are internally consistent, but the evaluation protocol has enough gaps (temporal cutoff for coarsening, missing thresholds, no artifacts) that I'd want fixes before I'd trust the numbers. read the letter →

arxiv 2607.22287 v1 pith:PTEHB7UI submitted 2026-07-24 cs.LG

Alessandro Sbandi , Federico Siciliano , Fabrizio Silvestri This is my paper
classification cs.LG
keywords graphcoarseninglabelpropagationrecommendersystemstelecommunicationsfamilydetectionneuralnetworksscalabilityNDCG@5
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

The paper attempts to show that a massive user-interaction graph can be made both faster and more accurate for offer recommendation by coarsening it first. Users are merged into super-nodes according to telecom business heuristics that approximate families or households, cutting the graph by more than 70%. A first propagation stage—either a lightweight label propagation algorithm or a small graph neural network—diffuses information across the coarsened graph, and a second stage propagates within each community to restore per-user personalization. On a real-world telecom dataset with over 13 million users, the LPA-only variant improves NDCG@5 by 24% over full-graph label propagation, and the GNN variant raises the gain above 50%. The message is that business-meaningful compression can simultaneously lower cost and improve ranking quality.

What carries the argument

The key object is the coarsening map C: V → V', a surjective assignment of users to super-nodes built from domain heuristics for family detection. The two-stage propagation is the mechanism that carries the argument: an inter-community diffusion step (LPA's iterative update Yhat = α D^{-1/2} A D^{-1/2} Y + (1-α) Y, or GraphSAGE on a heterogeneous community-item graph) produces community-level scores, and an intra-community LPA step on each subgraph, augmented by a representative node that injects the coarse score, converts them back to user-level recommendations. This division lets the expensive step run offline on a smaller graph and the cheap refinement run online and in parallel.

What would settle it

On the same or a similar telecom dataset, replace the business-rule coarsening with random node groupings of identical community sizes; if random coarsening matches the NDCG@5 improvement, the heuristics are not the cause. Also, run the full pipeline with and without the second-stage intra-community LPA; if removing it does not degrade NDCG@5, the two-stage claim collapses.

Watch

Extended reading notes

Core claim

The central discovery is that coarsening the user graph by business rules, rather than by generic structural methods, is what makes two-stage propagation outperform full-graph propagation. The paper's pipeline maps each user to one super-node via heuristics (call frequency, surname, care contacts, balance top-ups, number of lines), aggregates label vectors, and diffuses over the coarsened graph using LPA or a heterogeneous GraphSAGE model. It then runs a second LPA inside each community subgraph augmented with a representative node carrying the community-level scores, re-introducing individual-level personalization. On the telecom dataset this produces a 24% NDCG@5 gain with LPA in both stag

Load-bearing premise

The load-bearing premise is that the business heuristics—same surname, frequent calls, caring contacts, balance top-ups, number of lines—identify families whose aggregation preserves the structure that drives offer adoption; the paper gives no thresholds and no ground-truth validation of these groupings, so if the groupings are off, the gains may not reproduce.

Editorial extensions

If this is right

  • Telecom operators can use the LPA-only pipeline in low-latency settings because the first stage is offline and the per-community second stage runs in parallel, giving sub-second inference.
  • GNN-based recommendation becomes viable on graphs too large for direct training: the full graph caused out-of-memory, while the coarsened graph supported a lightweight GraphSAGE.
  • Deeper propagation helps only the business-coarsened graph: moving from one to five LPA steps raises NDCG@5 by over 17% on average for the proposed method, while baselines stay flat.
  • The coarsening drops more than 70% of nodes while preserving the original graph's sparse structure, which the authors link to stable, high-quality diffusion.
  • The method is tailored to shared-offer recommendations, where recommending the same item to multiple household members is redundant—coarsening removes exactly that redundancy.

Reading between the lines

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

  • The representative-node design is a generic 'global prior, local refinement' pattern that could transfer to any coarsened graph, such as zip-code or shared-account clusters in e-commerce, where the business rules differ.
  • The reported gains depend on specific business-rule thresholds that are not stated; a natural next step is to make those thresholds tunable or learned, and to test whether the advantage persists across threshold settings.
  • Because the second stage is local and parallelizable, the framework suggests a streaming protocol for dynamic graphs: update only the affected communities online and refresh the offline inter-community step periodically.
  • One untested hypothesis implied by the paper is that the intra-community refinement stage is the main source of the improvement; verifying it by ablating that stage would isolate where the gain comes from.
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 proposes a two-stage diffusion framework for large-scale telecom recommendation. Users are first grouped into super-node communities by business heuristics (interaction frequency, surname, caring contacts, balance top-ups, number of lines), reducing graph size by roughly 70%. A Label Propagation Algorithm (LPA) or a lightweight GraphSAGE model propagates labels on the coarsened graph, and a second, parallelized LPA runs inside each community to produce user-level recommendations. On a TIM dataset with over 13 million users, the authors report NDCG@5 improvements of +24% for the LPA-only pipeline over full-graph LPA, and further substantial gains when a GNN is used in the first stage, together with sub-second inference and detailed runtime measurements. The paper also compares the coarsening strategy with Louvain and Unique baselines, evaluates structural fidelity, and ablates LPA depth and damping parameters.

Significance. If the reported results hold, the contribution is practically significant: it demonstrates a scalable, low-latency recommendation pipeline on a real industrial graph, separates offline coarse propagation from online fine-grained refinement, and provides a concrete use of business-driven coarsening. The runtime comparison and the ablation over L and alpha are useful, and the focus on a real 13M-node dataset is a strength. However, the empirical claims are currently undermined by protocol gaps: the temporal cutoff for graph construction and coarsening is not specified, hyperparameters are selected on the test set, coarsening rule thresholds are absent, and the evaluation protocol is under-specified. These issues affect the credibility of the headline numbers, but they are addressable within the scope of a revision.

major comments (5)
  1. [§3.1.1, §2.1] The chronological split is defined only for purchase labels ('the last month is used as test set, and the previous months as training set'). No statement is made about the temporal cutoff for the construction of the user graph G or for the five coarsening heuristics (interaction frequency, surname, caring, balance, number of lines). If these are computed over the full January–September window, the coarsening observes September call/balance/care activity that would not be available at inference time. Since the full-graph LPA baseline does not use these auxiliary behavioral features, the reported +24% and +54% NDCG@5 gains could be partly or wholly an artifact of future information. The protocol must specify that G and all coarsening inputs are restricted to the training period (or to an earlier cutoff), and ideally report results under both settings.
  2. [§3.5, Tables 5 and 2] The ablation study varies L and alpha and evaluates NDCG@5 directly on the test set; the text then says 'Based on these results, we set the optimal hyperparameters to L=5 and alpha=0.5', and Table 2 reports the headline results using this configuration. With no validation split or nested selection procedure described, this is selection on the test set. The magnitude of the claimed improvements is therefore optimistic. Please add a proper validation split for hyperparameter selection and report test-set performance only for the final configuration, or explicitly describe the selection procedure if it is nested.
  3. [§2.1] The core coarsening step is defined by heuristics ('interaction frequency', 'same surname', 'frequently contacts customer support on behalf of someone else', 'topping up another user’s mobile balance', 'number of lines') and by constraints ('node degree and community size') without any numeric thresholds or algorithmic definitions. Table 4 reports a reduction from 13.8M to 4.27M nodes, but the exact grouping is not reproducible, and the reader cannot assess whether the communities indeed correspond to the claimed 'implicit family/household' structure. Please provide the precise rule thresholds, the complete coarsening procedure, or a pseudo-code specification.
  4. [Abstract vs. §3.2 vs. Contributions] The reported GNN gain is inconsistent. The Abstract says 'more than 50%', the contributions list says 'over 90% (GNN) over a full-graph LPA baseline', and Section 3.2 says 'by 54% over its LPA counterpart and by 35% over the GNN trained on Louvain coarsened graph.' These are different claims and cannot all be correct. The authors should state exactly which comparison underlies the headline number and cite the Table 2 entries that support it.
  5. [§3.1.4] The evaluation protocol does not specify how recommendations are generated from LPA/GNN scores for ranking metrics. Specifically, is the ranking over all 561 items, or over a candidate set? Are negative items sampled? This matters because user purchase sets are extremely sparse (training density 0.06%). Different negative-sampling or candidate-generation choices can change NDCG values substantially and affect the comparison between methods. Please describe the full ranking/scoring protocol, including how the second-stage LPA scores are converted into an ordered list.
minor comments (6)
  1. [§2.1, Eq. (1)] The label-aggregation formula uses 'Y_j' inside the sum, but should reference the label vector of user i (e.g., C_ij Y_i); the current notation is ill-defined.
  2. [§2.3 / Algorithm 1] The representative node r_c is introduced, but the LPA update for the augmented subgraph is not written. Please specify the initialization of r_c, the propagation equation, and how the final score vector is read out.
  3. [Table 3] The row for 'Users' appears malformed ('-6.287·10^-7 -2.839·10^-7 13.80'), and it is unclear why the full graph is included as a 'coarsening' baseline in a structural-preservation table.
  4. [Table 2] The table header has more columns than can be matched to the row values. Please use a clearer layout, label each metric explicitly (e.g., P@5, NDCG@5, R@5), and include standard deviations or an explicit variance statement.
  5. [§3.2] The 'Users' baseline in Table 2 cannot follow the same 'PA then LPA within each community' pipeline because there is no coarsening stage. Clarify what 'same methodology' means for the full-graph baseline.
  6. [General] No data or code availability statement is provided. While the industrial dataset may be proprietary, the authors should state whether the code and any anonymized/processed graph statistics can be released for reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

LPA hyperparameters are selected on the test set, so the headline +24% NDCG@5 is a fitted best-case value; the coarsening framework itself is not circular.

  1. fitted input called prediction [Sec. 3.2 (Table 2) and Sec. 3.5 (Ablation RQ4)]
    "All of these use the best configuration found (see Sec. 3.5 for LPA). ... Based on these results, we set the optimal hyperparameters toL= 5 and 𝛼= 0.5 , as this configuration gives the highest NDCG@5."

    Table 5 sweeps L∈{1,3,5} and alpha∈{0.1,0.5,0.9} and reports NDCG@5 on the same chronological test set used for final evaluation; Sec. 3.1.1 defines only 'the last month is used as test set' with no LPA validation split. The best cell (L=5, alpha=0.5) is then used in Table 2, whose caption states 'All of these use the best configuration found.' Hence the headline +24% NDCG@5 is the maximum of the hyperparameter grid on the test set: the evaluation metric was the selection objective, so the reported performance is a fitted best-case value rather than an independent prediction. The qualitative advantage is not fully forced because Ours also leads on many non-optimal cells.

full rationale

No self-citation chain, imported uniqueness theorem, or ansatz-via-citation is load-bearing; the method is an empirical pipeline (business-rule coarsening + inter-community LPA/GNN + intra-community LPA) with standard propagation equations. The only circular step is the LPA hyperparameter choice: L=5 and alpha=0.5 are picked as the highest NDCG@5 in Table 5, and Table 2 then reports results using that best configuration, with no validation split for the LPA branch. This inflates the specific +24% figure. The GNN branch used Bayesian optimization on a validation set (Sec. 3.1.3), and the coarsening itself is not derived from the labels, so the framework has independent content. The unquantified business heuristics and the missing temporal cutoff for the graph are reproducibility/leakage risks but not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central result rests on four tuned quantities (alpha, L, GNN hyperparameters, and coarsening thresholds) plus several domain assumptions about families, leakage-free splitting, and ranking evaluation. The invented entities are the synthetic representative node and the inferred family communities. None of these has independent evidence outside the paper; the empirical gains are the only support.

free parameters (4)
  • LPA damping alpha = 0.5 (chosen from {0.1, 0.5, 0.9})
    Appears in the LPA update in Sec 2.2. Selected in Sec 3.5 because it 'gives the highest NDCG@5'; if based on the test set, this is selection-on-test.
  • LPA propagation layers L = 5 (chosen from {1, 3, 5})
    Selected in Sec 3.5; drives the claim that deeper propagation yields over 17% average NDCG@5 gain. Same test-set selection concern as alpha.
  • GNN hyperparameters (learning rate, hidden dim, layers, negative samples, batch size) = not reported
    Tuned by Bayesian optimization on the validation set (Sec 3.1.3), but exact values are omitted, preventing replication of the GNN variant.
  • Business-rule thresholds (call frequency, surname matching, care contacts, balance top-ups, degree and community-size co = unspecified
    Sec 2.1 defines the coarsening rules only verbally. These choices determine the communities and therefore all downstream results; without them the method is not reproducible.
assumptions (4)
  • domain assumption LPA diffusion with L iterations and damping alpha behaves as an effective propagation for recommendation
    Sec 2.2 and Sec 3.5 rely on LPA as a surrogate for learned propagation; no convergence proof is given for this variant beyond cited empirical studies.
  • ad hoc to paper Surname, billing, care-contact, and balance-top-up patterns identify family/household groups relevant for offers
    Sec 2.1 lists these heuristics as the basis of coarsening; there is no independent validation against ground-truth family labels.
  • domain assumption The chronological train/test split prevents leakage from test users into the coarsened graph and propagated labels
    Sec 3.1.1 describes the split, but the paper does not state whether test-period edges or attributes are excluded from graph construction and coarsening.
  • domain assumption Top-K NDCG computed without explicit candidate filtering is a valid measure of agent recommendation quality
    Sec 3.1.4 defines metrics but never specifies how the 561 catalog items are ranked per user, whether negative sampling is used, or whether already-owned offers are masked despite the stated business constraint.
invented entities (2)
  • Community representative node r_c
    purpose: Injects coarse-grained community-level predictions into each user subgraph before the final LPA refinement (Sec 2.3, Algorithm 1 line 20).
    A synthetic graph node whose only support is the empirical NDCG improvement; there is no external falsifiable handle outside this paper.
  • Implicit family/household communities
    purpose: Super-nodes used to coarsen the graph; inferred from business rules rather than observed family labels.
    The paper does not validate these groups against real family data; the structural metrics in Table 3 are internal consistency checks, not external evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Recommendations via Graph Coarsening and Label Propagation." pith.science (2026). https://pith.science/paper/PTEHB7UI

@misc{pith2026260722287,
  author       = {Pith},
  title        = {Pith review of: Efficient Recommendations via Graph Coarsening and Label Propagation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PTEHB7UI}},
  note         = {Machine review of arXiv:2607.22287}
}
read the original abstract

Graph-based recommendations are widely adopted in real-world industrial applications. However, graphs in these systems often reach a massive scale, posing notable scalability and efficiency challenges. This requires techniques that can effectively balance predictive quality with computational cost. One promising approach is graph coarsening, an adaptive graph reduction technique that offers a way to systematically construct smaller, yet structurally representative, versions of the original large-scale graphs. In this work, we propose a flexible two-stage diffusion framework that combines graph coarsening with multi-step label propagation in the telecommunications domain. Domain-specific heuristics are applied to first aggregate nodes into meaningful communities, reducing graph size while preserving essential business-relevant relationships. An initial diffusion process done by a Label Propagation Algorithm (LPA) or a Graph Neural Network (GNN) propagates labels across the coarsened graph to produce coarse-grained predictions. Finally, a second LPA within subgraphs generates the final recommendations for individual users. On a real-world telecommunications dataset, when using LPA in both stages, our method achieves up to +24% NDCG@5 over the full-graph LPA baseline. Incorporating a lightweight GNN in the first stage further boosts NDCG@5 by more than 50%, but requires substantial training and inference time. Through extensive experiments and a detailed ablation, we quantify these trade-offs and demonstrate that our coarsening-driven approach delivers an optimal balance between scalability, latency, and recommendation quality.

Figures

Figures reproduced from arXiv: 2607.22287 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. On the left, the original full graph, where colors illustrate the business-driven rules used to group nodes into communities, while dashed lines highlight inter￾community interactions. In the middle is shown the coarsened graph and inter-community propagation. On the right, the intra-community refinement stage is shown. accuracy [19]. However, directly applying LPA to industrial-s… view at source ↗
Figure 2
Figure 2. NDCG@5 values for all coarsening methods, varying the number of layers 𝐿, with corre￾sponding 𝛼 values indicated. 3.4. Efficiency and Scalability (RQ3) Tab. 4 summarizes, for each method, the graph size (nodes and edges) at each propagation step and the corresponding runtimes for both coarsening and propagation steps. The full-graph baseline (Users), which involves performing a single propagation on an original grap… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 4 linked inside Pith

  1. [1]

    W. Fan, Y. Ma, Q. Li, Y. He, E. Zhao, J. Tang, D. Yin, Graph neural networks for social recommen- dation, in: The world wide web conference, 2019, pp. 417–426

    2019

  2. [2]

    Sharma, Y.-C

    K. Sharma, Y.-C. Lee, S. Nambi, A. Salian, S. Shah, S.-W. Kim, S. Kumar, A survey of graph neural networks for social recommender systems, ACM Computing Surveys 56 (2024) 1–34

    2024

  3. [3]

    H. Wang, J. Yang, J. Wang, Leverage large-scale biological networks to decipher the genetic basis of human diseases using machine learning, in: Artificial Neural Networks, Springer, 2020, pp. 229–248

    2020

  4. [4]

    J. Wang, S. Zhang, Y. Xiao, R. Song, A review on graph neural network methods in financial applications, arXiv preprint arXiv:2111.15367 (2021)

    arXiv 2021

  5. [5]

    Huang, W

    Z. Huang, W. Chung, H. Chen, A graph model for e-commerce recommender systems, Journal of the American Society for information science and technology 55 (2004) 259–274

    2004

  6. [6]

    D. Xu, C. Ruan, E. Korpeoglu, S. Kumar, K. Achan, Product knowledge graph embedding for e- commerce, in: Proceedings of the 13th international conference on web search and data mining, 2020, pp. 672–680

    2020

  7. [7]

    S. Wu, F. Sun, W. Zhang, X. Xie, B. Cui, Graph neural networks in recommender systems: a survey, ACM Computing Surveys 55 (2022) 1–37

    2022

  8. [8]

    C. Gao, Y. Zheng, N. Li, Y. Li, Y. Qin, J. Piao, Y. Quan, J. Chang, D. Jin, X. He, et al., A survey of graph neural networks for recommender systems: Challenges, methods, and directions, ACM Transactions on Recommender Systems 1 (2023) 1–51

    2023

Show all 44 references
  1. [9]

    Y. Pang, L. Wu, Q. Shen, Y. Zhang, Z. Wei, F. Xu, E. Chang, B. Long, J. Pei, Heterogeneous global graph neural networks for personalized session-based recommendation, in: Proceedings of the fifteenth ACM international conference on web search and data mining, 2022, pp. 775–783

    2022

  2. [10]

    Dhulipala, J

    L. Dhulipala, J. Łącki, J. Lee, V. Mirrokni, Terahac: Hierarchical agglomerative clustering of trillion- edge graphs, Proceedings of the ACM on Management of Data 1 (2023) 1–27

    2023

  3. [11]

    S. Sahu, A. Mhedhbi, S. Salihoglu, J. Lin, M. T. Özsu, The ubiquity of large graphs and surprising challenges of graph processing, Proceedings of the VLDB Endowment 11 (2017) 420–431

    2017

  4. [12]

    W. Ju, S. Yi, Y. Wang, Z. Xiao, Z. Mao, H. Li, Y. Gu, Y. Qin, N. Yin, S. Wang, et al., A survey of graph neural networks in real world: Imbalance, noise, privacy and ood challenges, arXiv preprint arXiv:2403.04468 (2024)

    2024

  5. [13]

    J. Dong, D. Zheng, L. F. Yang, G. Karypis, Global neighbor sampling for mixed cpu-gpu training on giant graphs, in: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 2021, pp. 289–299

    2021

  6. [14]

    Malewicz, M

    G. Malewicz, M. H. Austern, A. J. Bik, J. C. Dehnert, I. Horn, N. Leiser, G. Czajkowski, Pregel: a system for large-scale graph processing, in: Proceedings of the 2010 ACM SIGMOD International Conference on Management of data, 2010, pp. 135–146

    2010

  7. [15]

    Zhang, Z

    H. Zhang, Z. Yu, G. Dai, G. Huang, Y. Ding, Y. Xie, Y. Wang, Understanding gnn computational graph: A coordinated computation, io, and memory perspective, Proceedings of Machine Learning and Systems 4 (2022) 467–484

    2022

  8. [16]

    D. Chen, Y. Lin, W. Li, P. Li, J. Zhou, X. Sun, Measuring and relieving the over-smoothing problem for graph neural networks from the topological view, in: Proceedings of the AAAI conference on artificial intelligence, volume 34, 2020, pp. 3438–3445

    2020

  9. [17]

    Q. Li, Z. Han, X.-M. Wu, Deeper insights into graph convolutional networks for semi-supervised learning, in: Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  10. [18]

    Y. Xing, F. Meng, Y. Zhou, M. Zhu, M. Shi, G. Sun, A node influence based label propagation algorithm for community detection in networks, The Scientific World Journal 2014 (2014) 627581

    2014

  11. [19]

    Huang, H

    Q. Huang, H. He, A. Singh, S.-N. Lim, A. R. Benson, Combining label propagation and simple models out-performs graph neural networks, arXiv preprint arXiv:2010.13993 (2020)

    2010 arXiv

  12. [20]

    C. Cai, D. Wang, Y. Wang, Graph coarsening with neural networks, arXiv preprint arXiv:2102.01350 (2021)

    2021 arXiv

  13. [21]

    Huang, S

    Z. Huang, S. Zhang, C. Xi, T. Liu, M. Zhou, Scaling up graph neural networks via graph coarsening, in: Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp. 675–684

    2021

  14. [22]

    Hmimida, R

    M. Hmimida, R. Kanawati, A graph-coarsening approach for tag recommendation, in: Proceedings of the 25th International Conference Companion on World Wide Web, WWW ’16 Companion, Inter- national World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, CHE, ...

    2016

  15. [23]

    Hamilton, Z

    W. Hamilton, Z. Ying, J. Leskovec, Inductive representation learning on large graphs, Advances in neural information processing systems 30 (2017)

    2017

  16. [24]

    V. D. Blondel, J.-L. Guillaume, R. Lambiotte, E. Lefebvre, Fast unfolding of communities in large networks, Journal of statistical mechanics: theory and experiment 2008 (2008) P10008

    2008

  17. [25]

    Akiba, S

    T. Akiba, S. Sano, T. Yanase, T. Ohta, M. Koyama, Optuna: A next-generation hyperparameter optimization framework, in: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 2623–2631

    2019

  18. [26]

    U. N. Raghavan, R. Albert, S. Kumara, Near linear time algorithm to detect community structures in large-scale networks, Physical Review E—Statistical, Nonlinear, and Soft Matter Physics 76 (2007) 036106

    2007

  19. [27]

    T. Wang, S. Chen, X. Wang, J. Wang, Label propagation algorithm based on node importance, Physica A: Statistical Mechanics and its Applications 551 (2020) 124137

    2020

  20. [28]

    V. A. Traag, L. Šubelj, Large network community detection by fast label propagation, Scientific Reports 13 (2023) 2701

    2023

  21. [29]

    Clauset, M

    A. Clauset, M. E. Newman, C. Moore, Finding community structure in very large networks, Physical Review E—Statistical, Nonlinear, and Soft Matter Physics 70 (2004) 066111

    2004

  22. [30]

    F. Wu, B. A. Huberman, Finding communities in linear time: a physics approach, The European Physical Journal B 38 (2004) 331–338

    2004

  23. [31]

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, P. S. Yu, A comprehensive survey on graph neural networks, IEEE transactions on neural networks and learning systems 32 (2020) 4–24

    2020

  24. [32]

    Karypis, V

    G. Karypis, V. Kumar, Metis: A software package for partitioning unstructured graphs, partitioning meshes, and computing fill-reducing orderings of sparse matrices (1997)

    1997

  25. [33]

    V. A. Traag, L. Waltman, N. J. Van Eck, From louvain to leiden: guaranteeing well-connected communities, Scientific reports 9 (2019) 1–12

    2019

  26. [34]

    O. E. Livne, A. Brandt, Lean algebraic multigrid (lamg): Fast graph laplacian linear solver, SIAM Journal on Scientific Computing 34 (2012) B499–B522

    2012

  27. [35]

    Loukas, P

    A. Loukas, P. Vandergheynst, Spectrally approximating large graphs with smaller graphs, in: International conference on machine learning, PMLR, 2018, pp. 3237–3246

    2018

  28. [36]

    Z. Ying, J. You, C. Morris, X. Ren, W. Hamilton, J. Leskovec, Hierarchical graph representation learning with differentiable pooling, Advances in neural information processing systems 31 (2018)

    2018

  29. [37]

    H. Gao, S. Ji, Graph u-nets, in: international conference on machine learning, PMLR, 2019, pp. 2083–2092

    2019

  30. [38]

    J. Lee, I. Lee, J. Kang, Self-attention graph pooling, in: International conference on machine learning, pmlr, 2019, pp. 3734–3743

    2019

  31. [39]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, G. E. Dahl, Neural message passing for quantum chemistry, in: International conference on machine learning, PMLR, 2017, pp. 1263–1272

    2017

  32. [40]

    Veličković, G

    P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y. Bengio, Graph attention networks, arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  33. [41]

    R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, J. Leskovec, Graph convolutional neural networks for web-scale recommender systems, in: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 974–983

    2018

  34. [42]

    J. Wang, P. Huang, H. Zhao, Z. Zhang, B. Zhao, D. L. Lee, Billion-scale commodity embedding for e-commerce recommendation in alibaba, in: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 839–848

    2018

  35. [43]

    M. Wang, Y. Lin, G. Lin, K. Yang, X.-m. Wu, M2grl: A multi-task multi-view graph representation learning framework for web-scale recommender systems, in: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp. 2349–2358

    2020

  36. [44]

    Borisyuk, S

    F. Borisyuk, S. He, Y. Ouyang, M. Ramezani, P. Du, X. Hou, C. Jiang, N. Pasumarthy, P. Bannur, B. Tiwana, et al., Lignn: Graph neural networks at linkedin, in: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 4793–4803

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.