Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Breaking Structural Isolation: Scalable Graph Clustering via Community-Aware Sampling and Structural Entropy

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read SCISE breaks structural isolation in mini-batch graph clustering by expanding each batch with community context from constrained structural entropy, then learning contrastive embeddings on the denser structural subgraph.

desk verdict Solid scalable clustering pipeline that beats strong baselines on million-node graphs; the engineering is careful and the empirical package is thorough, even if the pieces are mostly known. read the letter →

arxiv 2607.05469 v1 pith:A33HTGV3 submitted 2026-07-06 cs.LG cs.AIcs.SI

Jingyun Zhang , Hao Peng , Jianxin Li , Angsheng Li , Philip S. Yu This is my paper
classification cs.LGcs.AIcs.SI
keywords unsupervisedgraphclusteringstructuralentropycommunity-awaresamplingmini-batchtrainingcontrastivelearninglarge-scalegraphsisolation
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

Unsupervised graph clustering on large networks is usually done with mini-batches, but ordinary random or neighborhood samplers split the same community across many batches. That severs the global topology the model needs, producing fragmented embeddings. SCISE first runs a constrained structural-entropy merge that forces the graph into a user-chosen number of cohesive communities instead of the tiny fragments that unconstrained entropy minimization creates. It then expands every training batch by pulling in other members of those communities (or a random-walk backbone for large ones). Inside the expanded batch it rebuilds edge weights from walk co-occurrence and trains a contrastive loss so the encoder sees higher-order community structure rather than isolated local edges. On six standard benchmarks, including graphs with millions of nodes, the resulting clusters beat or match ten recent methods while remaining memory-feasible where several baselines run out of memory.

What carries the argument

Structural Entropy Community Constraint (SECC): a greedy merge that repeatedly unites communities only while the number of communities exceeds a target N_comm, allowing temporary entropy increases when necessary so that the final partition contains exactly N_comm cohesive groups rather than thousands of micro-communities.

What would settle it

Replace the SECC partition with completely random community labels (or with a deliberately over-fragmented partition) and re-run the full pipeline on Ogbn-products or Reddit; if accuracy collapses to the level of the strongest non-community baseline, the claim that the constrained prior is doing essential work is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the combination of a cardinality-constrained structural-entropy community prior, community-aware batch expansion, and intra-batch structural-affinity contrastive learning is sufficient to restore the global community signal that ordinary mini-batch training destroys, and that this restoration yields state-of-the-art unsupervised clustering accuracy and scalability on graphs up to millions of nodes.

Load-bearing premise

That one fixed community partition computed once by SECC, using a user-chosen target number of communities, is a good enough structural scaffold for every subsequent mini-batch expansion.

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

3 major / 5 minor

Summary. The paper proposes SCISE, a scalable unsupervised graph clustering framework that addresses structural isolation in mini-batch GCL training. It combines three components: SECC (Algorithm 1), a constrained structural-entropy agglomeration that forces merges until a user-specified community count N_comm; CSampE, which expands each mini-batch with same-community nodes (full expansion for small communities, random-walk backbone for large ones); and StructCL, a multi-positive InfoNCE loss on a random-walk affinity matrix A* inside the expanded batch. After GCN encoding, K-means produces the final partitions. Experiments on six benchmarks (Photo through Ogbn-products) against ten baselines report best or second-best scores on 21 of 24 metrics (Table 2), with ablations (Table 3), hyper-parameter sweeps, noise/sparsity stress tests, runtime/memory tables, and t-SNE visualizations.

Significance. If the empirical claims hold, SCISE supplies a practical recipe for unsupervised clustering on million-node graphs that remain memory-feasible where several strong baselines OOM. The combination of a one-shot structural-entropy prior with community-aware batch expansion is a concrete engineering contribution that demonstrably restores long-range topology (Table 11) and yields large gains on large graphs (e.g., +17% F1 on Ogbn-arxiv). The paper ships extensive controls (ablations of every module and joint removals, noise ratios up to 100%, edge deletion up to 80%, runtime/memory breakdowns) and public code, which strengthens reproducibility and makes the work immediately usable for large-scale graph mining.

major comments (3)
  1. Table 9 (100% random noise row) shows that completely random community labels leave SCISE competitive with or better than MAGI on several datasets. Combined with the modest drops when SECC is removed alone (Table 3, w/o SECC), this undercuts the claim that the quality of the SECC prior is load-bearing. The manuscript should either (a) quantify how much of the Table 2 gains survive under random or noisy priors on all six datasets, or (b) reframe SECC as a convenient but non-essential initializer whose main role is to supply a cheap community index for CSampE.
  2. Section 4.1 / Algorithm 1: the forced-merge rule (when no ΔH < 0 pairs remain, still merge until |C| = N_comm) is an ad-hoc axiom. The paper never proves or even bounds that the resulting partitions remain near-minimal entropy or that they preserve the hierarchical semantics structural entropy is supposed to capture. A short theoretical or empirical characterization of the entropy gap introduced by forced merges (beyond the heatmaps in Fig. 8) is needed to justify the operator as more than a constrained agglomerative heuristic.
  3. N_comm is a free hyper-parameter that must be chosen by the user (Table 13 lists values from 300 to 59 000). Although Fig. 5 and the appendix show flat sensitivity curves, the abstract and introduction still present SECC as automatically producing cohesive partitions. The paper should state clearly that N_comm is an ordinary hyper-parameter (or supply a default selection rule, e.g., the unconstrained SE minimum) so that the method is not oversold as parameter-free community discovery.
minor comments (5)
  1. Eq. (2) for ΔH is lengthy and contains nested volume/cut terms; a short derivation or reference to the exact structural-entropy difference formula would help readers verify correctness.
  2. Figure 1 caption and the surrounding text use both “structural isolation” and “structural barriers”; consistent terminology would improve clarity.
  3. Table 2 reports “Improv.(%)” only against the best baseline per metric; adding the absolute scores of the second-best method in the same row would make the margin transparent.
  4. Section 4.5 claims overall epoch complexity O(N·(w_l w_t + f d)); the constant factor introduced by |V*_batch| expansion is left implicit. A short remark on observed expansion ratios (already in Table 11) would make the bound more concrete.
  5. A few typographical issues: “studys of multi-relational” (p. 3), “structure-aware” vs. “Structural” capitalization inconsistency, and missing spaces around some equation references.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical SOTA claims rest on external ground-truth metrics and ablations; structural-entropy definitions and self-citations supply background operators, not the reported scores.

full rationale

SCISE is an engineering/ML systems paper whose central claims are quantitative outperformance (Table 2: best/second-best on 21/24 NMI/ARI/ACC/F1 metrics vs. ten baselines, including memory-feasible runs on Reddit/Ogbn-products where several competitors OOM) plus supporting ablations (Table 3), noise/sparsity robustness (Tables 9-10), and sensitivity (Figs. 4-7). The training objective (StructCL multi-positive InfoNCE on random-walk affinities inside CSampE-expanded batches) and the SECC preprocessing (constrained greedy merge of structural entropy, Algorithm 1) never receive ground-truth labels; final partitions are produced by ordinary K-means on the learned embeddings and scored against external labels. N_comm is an ordinary hyper-parameter (set far larger than true K; shown stable across wide ranges and even under 100% random community noise). Self-citations (Li & Pan structural entropy definition, prior SE applications by overlapping authors) appear only as background definitions or related-work baselines that the paper itself outperforms; none is a uniqueness theorem that forces the numerical results, nor is any fitted quantity renamed a prediction. The derivation chain therefore does not reduce by construction to its inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 3 invented entities

The central empirical claim rests on standard GNN/contrastive machinery plus three paper-specific design choices (cardinality-constrained SE merge, community-aware expansion threshold, random-walk affinity). No new physical entities are postulated; free parameters are the usual hyper-parameters of sampling and clustering.

free parameters (5)
  • N_comm (target community count)
    User-chosen cardinality that terminates SECC; different values per dataset (Table 13). Sensitivity plots show flatness, yet the number is still selected by the authors.
  • p (merge speed / parallel merge ratio)
    Controls how many candidate pairs are merged per iteration; set per dataset and swept in Figure 4.
  • θ (community-size expansion threshold)
    Decides full-community vs. random-walk backbone expansion; dataset-specific (Table 13).
  • w_t, w_l (random-walk count and length)
    Control both CSampE backbone extraction and StructCL affinity matrix; chosen per dataset and ablated.
  • batch size B, learning rate, temperature τ
    Standard training hyper-parameters that affect final metrics; swept but still free.
assumptions (4)
  • domain assumption Structural entropy of a graph is well-defined by the encoding-tree formula of Li & Pan (2016) and is a meaningful objective for community discovery.
    Invoked throughout §3–4.1; SECC is a constrained variant of that objective.
  • ad hoc to paper Forcing merges even when ΔH ≥ 0 until exactly N_comm communities remain still yields useful cohesive partitions.
    Core of Algorithm 1 lines 10–12; no optimality proof is given, only empirical size histograms.
  • domain assumption Random-walk visit counts inside a mini-batch are a sufficient proxy for higher-order structural affinity for contrastive learning.
    Defines A* in Eq. (5) and the positive set of L_struct (Eq. 6).
  • domain assumption K-means on the final GCN embeddings recovers the ground-truth semantic clusters.
    Final step of Algorithm 2; standard but unproven for the learned space.
invented entities (3)
  • SECC operator (Structural Entropy Community Constraint)
    purpose: Produce a fixed-cardinality community prior that avoids over-fragmentation of unconstrained SE minimization.
    New algorithmic wrapper around classical structural entropy; independent evidence is only the empirical ablations and size histograms inside this paper.
  • CSampE (Community-Aware Sampling Expansion)
    purpose: Restore community context inside mini-batches to break structural isolation.
    Paper-specific sampling policy; no external theoretical guarantee.
  • StructCL (Structural Contrastive Learning)
    purpose: Train embeddings with multi-positive InfoNCE on random-walk affinities rather than raw edges.
    Variant of existing contrastive losses; novelty is the affinity construction, not a new mathematical object.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking Structural Isolation: Scalable Graph Clustering via Community-Aware Sampling and Structural Entropy." pith.science (2026). https://pith.science/paper/A33HTGV3

@misc{pith2026260705469,
  author       = {Pith},
  title        = {Pith review of: Breaking Structural Isolation: Scalable Graph Clustering via Community-Aware Sampling and Structural Entropy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A33HTGV3}},
  note         = {Machine review of arXiv:2607.05469}
}
read the original abstract

Unsupervised graph clustering is a fundamental technique for uncovering underlying semantic patterns in large-scale networks. Although Graph Contrastive Learning has demonstrated promising performance, existing methods often suffer from the "structural isolation" issue during mini-batch training, making it challenging to capture cohesive community structures that characterize the global topological distribution. To address these challenges, we propose SCISE, a Scalable unsupervised graph Clustering framework that preserves structural Integrity by synergizing community-aware sampling with constrained Structural Entropy. Specifically, we first introduce the Structural Entropy Community Constraint operator (SECC), which optimizes structural information within a constrained solution space to mitigate community fragmentation and enhance partition cohesion. Second, to prevent global information loss during batch training, we design a Community-Aware Sampling Expansion (CSampE) mechanism that incorporates the community context of target nodes into sampling batches, effectively breaking structural barriers and preserving topological integrity. Finally, we devise a Structural Contrastive Learning (StructCL) module that refines edge weights based on intra-batch structural similarity, guiding the encoder to learn representations in a higher-order structural space. Extensive experiments on six mainstream benchmark datasets demonstrate that SCISE significantly outperforms state-of-the-art algorithms, with ablation studies and robustness analyses further validating its effectiveness and reliability for real-world large-scale graphs.

Figures

Figures reproduced from arXiv: 2607.05469 by the authors.

Figure 1
Figure 1. Comparison of sampling strategies. (a) Random: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of SCISE. and parameter-free clustering. Furthermore, applying structural en￾tropy to unsupervised graph clustering [54] opens new avenues for synergizing structural information theory with neural networks. 2.2 Structural Information Theory Structural Information Theory has been validated in many applica￾tions. Introducing structural entropy in neural networks captures the underlying topological co… view at source ↗
Figure 3
Figure 3. Runtime comparison between SCISE and MAGI. Evaluation Metrics. We evaluate the accuracy and consistency of graph clustering with four metrics. NMI (Normalized Mutual Infor￾mation) evaluates how well the predicted clusters match the true clusters in terms of information shared. ARI (Adjusted Rand Index) assesses the similarity between the predicted and true cluster as￾signments, adjusting for random chance. ACC (Accu… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Sensitivity analysis of merge speed 𝑝 on Photo, Computers, Ogbn-arxiv, and Reddit datasets. 150 170 200 210 250 300 350 400 Ncomm ARI NMI F1 ACC 60 70 80 64.74 74.59 76.28 79.88 64.31 74.37 76.09 79.66 64.49 74.44 76.19 79.73 (a) Photo. 350 400 430 450 500 550 600 800 …
Figure 5
Figure 5. Figure 5: Sensitivity analysis of community number [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Sensitivity analysis of initial batch size [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Evolution of structural entropy and community distribution on Ogbn-arxiv. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Visualization on Pubmed. reduced boundary ambiguity. This observation suggests that the proposed community-aware structural expansion and contrastive learning strategy effectively capture global structural semantics, re￾sulting in more discriminative representations fo…
Figure 10
Figure 10. Figure 10: Analysis of hyper 𝑁𝑐𝑜𝑚𝑚 on Pubmed and Ogbn￾products datasets. 0.001 0.01 0.05 0.1 0.5 1 p 30 32 34 36 NMI SCISE (35.77) Best Baseline (33.30) (a) Pubmed. 0.001 0.01 0.05 0.1 0.5 1 p 55.1 55.3 55.5 55.7 NMI SCISE (55.61) Best Baseline (55.10) (b) Ogbn-products [PITH_F…
Figure 11
Figure 11. Figure 11: Analysis of hyper 𝑝 on Pubmed and Ogbn-products datasets. 1024 2048 4096 8192 Batch Size 65 72 74 Pubmed 1024 2048 4096 Batch Size 35 43 45 Ogbn-arxiv 1024 2048 4096 8192 Batch Size 80 88 92 Reddit 1024 2048 4096 Batch Size 35 40 44 Ogbn-products [PITH_FULL_IMAGE:fig…
Figure 12
Figure 12. Figure 12: Analysis of hyper [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Analysis of hyper 𝜃 with ACC. • MAGI [20] is a graph clustering framework that utilizes modu￾larity maximization as a contrastive pretext task to define positive and negative samples, capturing structural information without manual data augmentations [PITH_FULL_IMAGE…
Figure 14
Figure 14. Figure 14: Evolution of structural entropy and community distribution on Photo. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Evolution of structural entropy and community distribution on Ogbn-products. [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 2 linked inside Pith

  1. [1]

    Amit Aflalo, Shai Bagon, Tamar Kashti, and Yonina Eldar. 2023. Deepcut: Unsu- pervised segmentation using graph neural networks clustering. InProceedings of the IEEE/CVF International Conference on Computer Vision. 32–41

    2023

  2. [2]

    Leman Akoglu, Hanghang Tong, and Danai Koutra. 2015. Graph based anomaly detection and description: a survey.Data mining and knowledge discovery29, 3 (2015), 626–688

    2015

  3. [3]

    Deyu Bo, Xiao Wang, Chuan Shi, Meiqi Zhu, Emiao Lu, and Peng Cui. 2020. Structural deep clustering network. InProceedings of the web conference 2020. 1400–1410

    2020

  4. [4]

    Yuwei Cao, Hao Peng, Angsheng Li, Chenyu You, Zhifeng Hao, and Philip S Yu

  5. [5]

    InProceedings of the Conference on Uncertainty in Artificial Intelligence

    Multi-Relational Structural Entropy. InProceedings of the Conference on Uncertainty in Artificial Intelligence. 1–15

  6. [6]

    Yuwei Cao, Hao Peng, Zhengtao Yu, and Philip S Yu. 2024. Hierarchical and Incremental Structural Entropy Minimization for Unsupervised Social Event Detection. InProceedings of the AAAI conference on artificial intelligence. 1–13

    2024

  7. [7]

    Jianfei Chen, Jun Zhu, and Le Song. 2018. Stochastic Training of Graph Con- volutional Networks with Variance Reduction. InInternational Conference on Machine Learning. PMLR, 942–950

    2018

  8. [8]

    Mulin Chen, Bocheng Wang, and Xuelong Li. 2024. Deep contrastive graph learning with clustering-oriented guidance. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 11364–11372

    2024

Show all 63 references
  1. [9]

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

  2. [10]

    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

  3. [11]

    Fnu Devvrit, Aditya Sinha, Inderjit Dhillon, and Prateek Jain. 2022. S3GC: Scalable Self-Supervised Graph Clustering. InAdvances in Neural Information Processing Systems, Vol. 35. Curran Associates, Inc., 3248–3261

    2022

  4. [12]

    Santo Fortunato. 2010. Community detection in graphs.Physics reports486, 3-5 (2010), 75–174

    2010

  5. [13]

    Hongyang Gao, Zhengyang Wang, and Shuiwang Ji. 2018. Large-scale learn- able graph convolutional networks. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1416–1424

    2018

  6. [14]

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

    2017

  7. [15]

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. 2020. Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems 33 (2020), 22118–22133

    2020

  8. [16]

    Li Huang, Jingyun Zhang, Weikang Gong, Guangjie Zeng, Hao Peng, and Dong- sheng Chen. 2026. SECTOR: structural entropy-based learning of spatiotemporal organisation in spatial transcriptomics.Bioinformatics42, 6 (2026), btag367

    2026

  9. [17]

    Xiang Huang, Hao Peng, Li Sun, Hui Lin, Chunyang Liu, Jiang Cao, and Philip S Yu. 2025. Structural entropy guided probabilistic coding. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 17467–17475

    2025

  10. [18]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. https://openreview.ne...

    2017

  11. [19]

    Gayan K Kulatilleke, Marius Portmann, and Shekhar S Chandra. 2025. SCGC: Self-supervised contrastive graph clustering.Neurocomputing611 (2025), 128629

    2025

  12. [20]

    Angsheng Li and Yicheng Pan. 2016. Structural information and dynamical complexity of networks.IEEE TIT62, 6 (2016), 3290–3339

    2016

  13. [21]

    Mengyao Li, Zhibang Yang, Xu Zhou, Yixiang Fang, Kenli Li, and Keqin Li

  14. [22]

    Surveys57, 7 (2025), 1–36

    Clustering on attributed graphs: From single-view to multi-view.Comput. Surveys57, 7 (2025), 1–36

    2025

  15. [23]

    Yunfei Liu, Jintang Li, Yuehe Chen, Ruofan Wu, Ericbk Wang, Jing Zhou, Sheng Tian, Shuheng Shen, Xing Fu, Changhua Meng, et al . 2024. Revisiting modu- larity maximization for graph clustering: A contrastive learning perspective. In Proceedings of the 30th ACM SIGKDD Conferenc...

    2024

  16. [24]

    Yue Liu, Ke Liang, Jun Xia, Xihong Yang, Sihang Zhou, Meng Liu, Xinwang Liu, and Stan Z Li. 2023. Reinforcement graph clustering with unknown cluster number. InProceedings of the 31st ACM international conference on multimedia. 3528–3537

    2023

  17. [25]

    Yue Liu, Ke Liang, Jun Xia, Sihang Zhou, Xihong Yang, Xinwang Liu, and Stan Z Li. 2023. Dink-net: Neural clustering on large graphs. InInternational conference on machine learning. PMLR, 21794–21812

    2023

  18. [26]

    Yiwei Liu, Jiamou Liu, Zijian Zhang, Liehuang Zhu, and Angsheng Li. 2019. REM: From structural entropy to community structure deception.Proceedings of the Advances in Neural Information Processing Systems32 (2019), 1–11

    2019

  19. [27]

    Yue Liu, Jun Xia, Sihang Zhou, Xihong Yang, Ke Liang, Chenchen Fan, Yan Zhuang, Stan Z Li, Xinwang Liu, and Kunlun He. 2022. A survey of deep graph clustering: Taxonomy, challenge, application, and open resource.arXiv preprint arXiv:2211.12875(2022)

    2022 arXiv

  20. [28]

    Yixin Liu, Yu Zheng, Daokun Zhang, Hongxu Chen, Hao Peng, and Shirui Pan

  21. [29]

    InProceedings of the ACM Web Conference 2022

    Towards unsupervised deep graph structure learning. InProceedings of the ACM Web Conference 2022. 1392–1403

    2022

  22. [30]

    Hao Peng, Jingyun Zhang, Xiang Huang, Zhifeng Hao, Angsheng Li, Zhengtao Yu, and Philip S Yu. 2024. Unsupervised Social Bot Detection via Structural Information Theory.ACM Transactions on Information Systems42, 6 (2024), 42

    2024

  23. [31]

    Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. 2022. Recipe for a general, powerful, scalable graph transformer.Advances in Neural Information Processing Systems35 (2022), 14501–14515

    2022

  24. [32]

    Yazhou Ren, Jingyu Pu, Zhimeng Yang, Jie Xu, Guofeng Li, Xiaorong Pu, Philip S Yu, and Lifang He. 2024. Deep clustering: A comprehensive survey.IEEE trans- actions on neural networks and learning systems36, 4 (2024), 5858–5878

    2024

  25. [33]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective classification in network data.AI magazine29, 3 (2008), 93–93

    2008

  26. [34]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of graph neural network evaluation.arXiv preprint arXiv:1811.05868(2018)

    2018 arXiv

  27. [35]

    Xiao Shen, Dewang Sun, Shirui Pan, Xi Zhou, and Laurence T Yang. 2023. Neigh- bor contrastive learning on learnable graph augmentation. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 9782–9791

    2023

  28. [36]

    Zhihao Shi, Jie Wang, Fanghua Lu, Hanzhu Chen, Defu Lian, Zheng Wang, Jieping Ye, and Feng Wu. 2024. Label deconvolution for node representation learning on large-scale attributed graphs against learning bias.IEEE Transactions on Pattern Analysis and Machine Intelligence46, 12...

    2024

  29. [37]

    Hamed Shirzad, Ameya Velingker, Balaji Venkatachalam, Danica J Sutherland, and Ali Kemal Sinop. 2023. Exphormer: Sparse transformers for graphs. In International Conference on Machine Learning. PMLR, 31613–31632

    2023

  30. [38]

    Li Sun, Zhenhao Huang, Hao Peng, Yujie Wang, Chunyang Liu, and Philip S Yu. 2024. LSEnet: Lorentz Structural Entropy Neural Network for Deep Graph Clustering. InInternational Conference on Machine Learning. PMLR, 47078–47104

    2024

  31. [39]

    Anton Tsitsulin, John Palowitch, Bryan Perozzi, and Emmanuel Müller. 2023. Graph clustering with graph neural networks.Journal of Machine Learning Research24, 127 (2023), 1–21

    2023

  32. [40]

    Petar Veličković, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. 2018. Deep graph infomax. InInternational Conference on Learning Representations

    2018

  33. [41]

    Yifei Wang, Yupan Wang, Zeyu Zhang, Song Yang, Kaiqi Zhao, and Jiamou Liu

  34. [42]

    InProceedings of the AAAI Conference on Artificial Intelligence

    User: Unsupervised structural entropy-based robust graph neural network. InProceedings of the AAAI Conference on Artificial Intelligence. 10235–10243

  35. [43]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. InInternational conference on machine learning. Pmlr, 6861–6871

    2019

  36. [44]

    Junran Wu, Xueyuan Chen, Bowen Shi, Shangzhe Li, and Ke Xu. 2023. SEGA: Structural Entropy Guided Anchor View for Graph Contrastive Learning. In Proceedings of the ICML. PMLR, 1–20

    2023

  37. [45]

    Junran Wu, Xueyuan Chen, Ke Xu, and Shangzhe Li. 2022. Structural entropy guided graph hierarchical pooling. InProceedings of the International Conference on Machine Learning. PMLR, 24017–24030

    2022

  38. [46]

    Qitian Wu, Wentao Zhao, Chenxiao Yang, Hengrui Zhang, Fan Nie, Haitian Jiang, Yatao Bian, and Junchi Yan. 2023. Sgformer: Simplifying and empowering transformers for large-graph representations.Advances in neural information processing systems36 (2023), 64753–64773

    2023

  39. [47]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. 2020. A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems32, 1 (2020), 4–24

    2020

  40. [48]

    Junyuan Xie, Ross Girshick, and Ali Farhadi. 2016. Unsupervised deep embedding for clustering analysis. InInternational conference on machine learning. PMLR, 478–487

    2016

  41. [49]

    Minhui Xie, Hao Peng, Pu Li, Guangjie Zeng, Shuhai Wang, Jia Wu, Peng Li, and Philip S Yu. 2025. Hierarchical superpixel segmentation via structural information theory. InProceedings of the 2025 SIAM International Conference on Data Mining (SDM). SIAM, 242–251

    2025

  42. [50]

    Xihong Yang, Yue Liu, Sihang Zhou, Siwei Wang, Wenxuan Tu, Qun Zheng, Xinwang Liu, Liming Fang, and En Zhu. 2023. Cluster-guided contrastive graph clustering network. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 10834–10842

    2023

  43. [51]

    Xihong Yang, Cheng Tan, Yue Liu, Ke Liang, Siwei Wang, Sihang Zhou, Jun Xia, Stan Z Li, Xinwang Liu, and En Zhu. 2023. Convert: Contrastive graph clustering with reliable augmentation. InProceedings of the 31st ACM international conference on multimedia. 319–327

    2023

  44. [52]

    Zhenyu Yang, Ge Zhang, Jia Wu, Jian Yang, Quan Z Sheng, Hao Peng, Angsheng Li, Shan Xue, and Jianlin Su. 2023. Minimum entropy principle guided graph neural networks. InProceedings of the Sixteenth ACM International Conference on Web Search and Data Mining. 114–122

    2023

  45. [53]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale 13 recommender systems. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 974–983

    2018

  46. [54]

    Guangjie Zeng, Hao Peng, Angsheng Li, Li Sun, Chunyang Liu, Shengze Li, Yicheng Pan, and Philip S Yu. 2026. Hyperbolic continuous structural entropy for hierarchical clustering. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 28094–28102

    2026

  47. [55]

    Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. 2019. GraphSAINT: Graph Sampling Based Inductive Learning Method. InInternational Conference on Learning Representations

    2019

  48. [56]

    Xianghua Zeng, Hao Peng, and Angsheng Li. 2025. Proactive Bot Detection Based on Structural Information Principles.IEEE Transactions on Pattern Analysis and Machine Intelligence(2025)

    2025

  49. [57]

    Xianghua Zeng, Hao Peng, Angsheng Li, Chunyang Liu, Lifang He, and Philip S. Yu. 2023. Hierarchical State Abstraction based on Structural Information Princi- ples. InProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, Edith Elk...

    2023

  50. [58]

    Xianghua Zeng, Hao Peng, Yicheng Pan, Angsheng Li, and Guanlin Wu. 2026. Structural Information-based Hierarchical Diffusion for Offline Reinforcement Learning.Advances in Neural Information Processing Systems38 (2026), 109539– 109567

    2026

  51. [59]

    Jingyun Zhang, Hao Peng, Li Sun, Guanlin Wu, Chunyang Liu, and Zhengtao Yu

  52. [60]

    InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Unsupervised graph clustering with deep structural entropy. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

  53. [61]

    Jingyun Zhang, Hao Peng, Mingdai Yang, and Philip S Yu. 2025. Enhanced Pre-training for Recommendation via Hypergraph Structural Entropy.ACM Transactions on Information Systems44, 2 (2025), 1–48

    2025

  54. [62]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2021. Graph contrastive learning with adaptive augmentation. InProceedings of the web conference 2021. 2069–2080

    2021

  55. [63]

    Dongcheng Zou, Hao Peng, and Chunyang Liu. 2024. A structural information guided hierarchical reconstruction for graph anomaly detection. InProceed- ings of the 33rd ACM International Conference on Information and Knowledge Management. 4318–4323. 14 7 APPENDIX A NOTATIONS The ...

    2024

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.