Pith. sign in

REVIEW 3 major objections 5 minor 95 references

Revisiting Self-Supervised Heterogeneous Graph Learning from Spectral Clustering Perspective

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read This paper claims that a rank-constrained spectral clustering loss makes self-supervised heterogeneous graph representations split into exactly the number of classes, and that dual node- and cluster-level consistency constraints improve…

desk verdict Solid, reproducible empirical method in search of a proof; the headline theorems are overclaimed, but the experiments deserve referee time. read the letter →

arxiv 2412.00742 v1 pith:2RQQZBDM submitted 2024-12-01 cs.AI

classification cs.AI
keywords self-supervisedheterogeneousgraphlearningspectralclusteringrankconstraintaffinitymatrixdualconsistencyconstraintsnodeclassification
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

Self-supervised heterogeneous graph learning (SHGL) methods implicitly cluster nodes, but they pass messages over graphs that contain connections between different classes, and they ignore cluster-level information. The paper tries to establish that both defects can be fixed by treating SHGL explicitly as spectral clustering: learn an affinity matrix whose graph has exactly $c$ connected components, one per class, and add node-level and cluster-level consistency losses. It proves that optimizing the resulting spectral loss drives the learned representations into $c$ partitions, matching the number of classes, and that the dual consistency constraints give a lower model-complexity bound and a higher generalization bound than node-level-only SHGL. If correct, this gives a principled reason to split representations into exactly the class count and supports stronger downstream classification and clustering.

What carries the argument

The load-bearing object is the learned affinity matrix $S$, updated in closed form as $s_{ij}=(-d_{ij}/(2\alpha)+\lambda)_+$ under the constraint that each row sums to one. The rank constraint on its Laplacian, $\mathrm{rank}(L_S)=n-c$, is designed to force exactly $c$ connected components; by Ky Fan's theorem this constraint can be rewritten as a trace spectral-clustering objective. Because eigendecomposition is expensive, the method replaces the eigenvectors with an orthogonalized cluster assignment matrix $Y$ produced by a projection head and a QR orthogonalization layer, and trains $Y$ with the spectral loss $L_{sp}$. The dual consistency losses $L_{nc}$ and $L_{cc}$ then align node representations across views and align each node with its cluster centroid, capturing both invariant and cluster-level information.

What would settle it

Run the method on any dataset with known labels using the paper's parameter settings, compute the affinity matrix $S$ from Eq. (10), and count the connected components of the graph it defines, or equivalently count the zero eigenvalues of its Laplacian. If the count differs from $c$ on a standard run, the central premise of Theorem 2.5 is empirically false.

Watch

Extended reading notes

Core claim

The central claim is Theorem 2.5: minimizing the spectral loss $L_{sp}$ is equivalent to performing spectral clustering on the affinity matrix $S$ with exactly $c$ connected components, and then to running the RatioCut graph-cut algorithm that splits node representations into $c$ partitions rather than into $d$ representation dimensions. The rank constraint on the Laplacian of $S$ is what enforces the $c$-component structure, so message passing through $Z = SH$ is supposed to aggregate only within-class nodes. Theorem 2.6 adds that the node-level and cluster-level consistency constraints together achieve a lower bound on model complexity and a higher bound on generalization ability than prior SHGL methods with only a node-level constraint. The experiments report consistent gains over prior methods on four heterogeneous and two homogeneous graph datasets for both node classification and node clustering.

Load-bearing premise

Everything rests on the learned affinity matrix really ending up with exactly one separate group of nodes per class, so that message passing never crosses class boundaries; the paper asserts this follows from minimizing the spectral loss but does not derive it.

Editorial extensions

If this is right

  • If the affinity matrix has exactly $c$ connected components, message passing through $Z = SH$ aggregates only nodes in the same class, so the learned representations are not contaminated by cross-class noise.
  • With representations partitioned into $c$ clusters instead of $d$ representation dimensions, the graph-cut objective matches the true label count, giving a cleaner optimization target than previous SHGL.
  • The dual consistency constraints provably lower the model-complexity measure and raise the generalization bound, so downstream classifiers trained on the representations should transfer better.
  • Replacing eigendecomposition with the orthogonal projection head keeps per-epoch cost linear in the number of nodes, so the method scales to large graphs.
  • The same recipe transfers to homogeneous graphs by swapping the heterogeneous encoder for a GCN, so the benefit is not tied to heterogeneous node types.

Reading between the lines

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

  • Editorial extension: the exact-$c$ guarantee is only as strong as the rank constraint actually being active; one can test on any dataset whether the learned $S$ has exactly $c$ connected components, and if it does not, the theory's precondition is violated even if the loss still decreases.
  • Editorial extension: the framework assumes the number of classes $c$ is known in advance; in open-set settings where the class count is unknown or drifts, the rank constraint would need a model-selection rule for $c$, which the paper does not provide.
  • Editorial extension: because the method builds the affinity matrix from node-feature distances, its noise-free claim is weaker when features are missing or uninformative; the paper itself flags feature dependence as a limitation, and a natural stress test is performance on feature-poor heterogeneous graphs.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes SCHOOL, a self-supervised heterogeneous graph learning method that revisits SHGL from a spectral clustering viewpoint. It learns a k-nearest-neighbor affinity matrix with a rank constraint, uses a spectral loss to fit an orthogonal cluster assignment matrix, and adds node- and cluster-level consistency constraints. The authors claim two central theoretical results: (i) optimizing the spectral loss partitions representations into exactly c clusters, and (ii) the dual consistency constraints yield lower model complexity and higher generalization bounds than previous SHGL methods. Experiments on four heterogeneous and two homogeneous graph datasets report consistent improvements over a wide range of baselines, with code released.

Significance. If the theoretical claims were rigorously established, the paper would be a valuable contribution: it would unify SHGL methods with spectral clustering and provide a principled mechanism for filtering noisy graph connections. The empirical study is a notable strength: results are repeated five times with reported deviations, the comparison set is broad, and the code is publicly released. The paper also makes a falsifiable prediction—that the learned affinity matrix has exactly c connected components and the representations split into c partitions—which is partially supported by the visualizations. However, the two central theorems are not proven as stated: the exact-rank partition claim and the generalization-bound claim rest on unsupported inferences, so the theoretical contribution currently overreaches what the proofs establish.

major comments (3)
  1. [Section 2.2, Eq. (8); Appendix C.3, Theorem 2.5] The proof of Theorem 2.5 does not establish that the affinity matrix S has exactly c connected components. For fixed S, the first term in Lsp equals (2/n^2)Tr(Y^T L_S Y); its minimum over orthogonal Y is (2/n^2) times the sum of the c smallest eigenvalues of L_S. That minimum is zero only if each of those c eigenvalues is zero, and nothing in Eq. (13) forces λ_c(L_S)=0. The entropy term −γH(Y) actively discourages Y from being the eigenvector matrix F used in Eq. (8), so the Ky Fan surrogate in Eq. (7) does not implement the hard rank constraint rank(L_S)=n−c introduced in Eq. (6). Furthermore, the closed-form update in Eq. (10) with α chosen in Appendix C.6 only enforces k-sparsity per row, not zero eigenvalues. The visualizations in Section 3.2.3 themselves say the heatmaps show 'nearly c' components, which confirms that exactness is not established. Thus the load-bearing premise of Theorem 2.5—that S has exactly c connected components and the representations are divided into exactly c partitions—is asserted rather than proven.
  2. [Appendix C.4, Theorem 2.6] The proof of Theorem 2.6 is close to circular. The complexity measure C in Definition C.5 is defined through within-class scatter S_i, and the cluster-level loss Lcc in Eq. (18) directly minimizes the distance from each projected representation to its cluster centroid, i.e., it directly reduces the quantities σ0^2 and σ1^2 that appear in the numerator of C. The derivation then shows a lower bound on C and observes that Lcc reduces that lower bound, but a reduction in a lower bound does not imply a reduction of the actual quantity inf(C_SCHOOL) versus inf(C_SHGL). Moreover, the generalization bound G in Definition C.6 is invoked from the literature without a derivation that applies to this setting, and the paper does not actually compute or compare G for the two methods. Consequently, the inequality in Eq. (20) is not a theorem with the stated proof; it is a restatement of the objective in new notation plus an unverified appeal to a generalization bound.
  3. [Appendix C.1, Theorem 2.2] The proof of the central 'revisiting' theorem contains an unjustified algebraic step. In the derivation for meta-path-based methods, the cross term in Eq. (25) is transformed in Eq. (29) into '2∑_{i,j,k} G_{i,j}G_{i,k}⟨(hi−hj)·(hi−hk)⟩' without defining G or showing how this follows from the preceding expression. The notation in Eq. (26) also conflates hi as a column vector and hij as an entry, and the line 'denote hi ∈ Rn, hj ∈ Rd' is malformed. Since Theorem 2.2 is the theoretical foundation for the paper's motivation—that previous SHGL methods perform spectral clustering with regularization—this proof needs to be made rigorous before the subsequent claims can be accepted.
minor comments (5)
  1. [Section 3.2.3, Figure 2] The text states that the heatmaps exhibit 'nearly c' components and 'almost all' large values fall in the block diagonal structure, while Theorem 2.5 claims exactly c connected components; these statements should be reconciled because the visual evidence supports approximate, not exact, behavior.
  2. [Eq. (13)] The entropy term H(Y) is defined with P(y_i) = (1/n)∑_j y_i^j, but the preceding text defines y_i as the i-th column and j-th row of Y; the indexing should be clarified to avoid confusion.
  3. [Eq. (10) and Appendix C.6] The Lagrange multiplier λ is used in Eq. (10) before it is introduced in Appendix C.6; please define it when Eq. (9) is presented.
  4. [Appendix C.6, Eq. (69)] Setting α_i to the upper bound in inequality (68) is stated to yield exactly k nonzeros, but the strict versus non-strict inequality conditions in the KKT derivation should be made explicit.
  5. [NeurIPS Checklist, item 3] The checklist answer states that complete and correct proofs are provided in Appendix C, but the gaps in the proofs of Theorems 2.5 and 2.6 contradict that claim; the checklist should be revised accordingly.

Circularity Check

2 steps flagged · score 6.0 of 10

Two of the paper's headline 'theoretical demonstrations' reduce to construction: Theorem 2.5 restates the imposed rank constraint as a prediction of c partitions, and Theorem 2.6 defines the complexity measure from the same intra-class scatter that the cluster-level loss directly minimizes.

  1. self definitional [Section 2.2, Eqs. (6)-(8) and (10); Theorem 2.5; Appendix C.3]
    "Based on Lemma 2.4, we can solve the above issue by adding the rank constraint on the affinity matrix, i.e., enforcing the smallest c eigenvalues of LS to be 0: rank(LS) = n − c ⇒ min Pc i=1 τi (LS). ... Therefore, the proposed method obtains the affinity matrix with exactly c connected components to mitigate noisy connections in an effective and efficient way. ... [Appendix C.3] Therefore, when the minimum of Lsp is attained, the constraints in the above function can be satisfied, i.e., rank(LS) = n−c holds. As a result, we can obtain the affinity matrix S with exactly c connected components."

    Eq. (6) defines the design goal as enforcing rank(LS)=n−c, i.e., exactly c connected components, and Eq. (8) adds that constraint to the optimization. Theorem 2.5's advertised conclusion—'S with c connected components' and 'divide the learned representations into c partitions'—is exactly this imposed rank condition restated. The proof does not show that minimizing Lsp (a Ky Fan surrogate with fixed β plus an entropy term) drives the c smallest eigenvalues of LS to zero; it merely asserts that the constraint 'can be satisfied' at the minimum. Since Appendix C.6 selects α only to make each row k-sparse, and the paper's own heatmaps show 'nearly c' components, the c-partition property is an input constraint or unproven assumption, not a derived prediction.

  2. self definitional [Section 2.3 Eq. (18); Theorem 2.6; Appendix C.4, Definitions C.5-C.6 and proof]
    "Eq. (18) enables the projected representation ˜qi and the cluster representation ˆqyi to align each other. As a result, representations capture the clustering information based on cluster indicators and reduce intra-cluster differences to improve the performance of downstream tasks. ... [Definition C.5] Si is a measure of scatter within representations of class i ... [Proof] Note that the cluster-level consistency constraint minimizes the first term in the S2 0 and S2 1 (i.e., σ2 0 and σ2 1). ..."

    The complexity measure C in Definition C.5 is the Davies-Bouldin index: intra-class scatter Si divided by inter-class separation Mi,j. The cluster-level loss Lcc in Eq. (18) is built to align every projected representation ˜qi with its cluster center ˆq_{y_i}, and the paper's own description says it 'reduce[s] intra-cluster differences.' The proof then observes that Lcc minimizes σ0 and σ1, which are exactly the scatter terms appearing in C, and concludes inf(C_SCHOOL)<inf(C_SHGL). Thus the generalization-bound theorem is the training objective re-expressed: a loss designed to minimize intra-class scatter lowers a complexity measure defined as intra-class scatter, and the 'higher generalization bound' is inherited from the formula G ≤ sqrt(C/n).

full rationale

The paper is not wholly circular: the algebraic reduction of earlier SHGL objectives to spectral-clustering and RatioCut forms (Theorems 2.2 and 2.3) has independent content under the stated orthogonality assumption, and the experimental comparisons against external benchmarks are self-contained and falsifiable. No load-bearing self-citation chain is present; the authors' prior HERO work is cited for comparison, and the rank-constrained clustering idea is attributed to earlier external literature. However, the two headline 'theoretically demonstrated' claims do reduce to construction. Theorem 2.5's c-partition conclusion is the rank constraint introduced in Eq. (6)/(8) asserted rather than derived from Lsp, and the proof's leap from minimizing a sum of c smallest eigenvalues to rank(LS)=n−c is not justified. Theorem 2.6's complexity/generalization advantage is definitional: the loss Lcc minimizes the same intra-class scatter that defines the complexity measure C, so the conclusion is the objective restated in new notation. The checklist's claim of complete proofs is therefore not supported by Appendix C.3. Because the abstract and introduction present exactly these two results as the paper's theoretical contributions, the circularity score is moderate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method relies on a handful of tuned hyperparameters (alpha, k, beta, gamma, eta, mu, delta) and on several unproven or heuristic assumptions in the theoretical sections, most notably that the surrogate spectral loss enforces the rank constraint and that the cluster-level loss reduces the variances used in the generalization bound. No new physical or conceptual entities are introduced.

free parameters (6)
  • alpha = n/a
    Sparsity parameter in affinity update, set per dataset via Eq. (70); controls the number of nonzero affinities per row.
  • k = n/a
    Number of nearest neighbors for sparse affinity; hyperparameter determining sparsity and indirectly the number of graph components.
  • beta = n/a
    Weight of spectral term in Eq. (8); trades off eigenvector fitting against distance-based affinity.
  • gamma = n/a
    Entropy regularization weight in Lsp; avoids trivial all-one cluster assignments.
  • eta = n/a
    Uniformity weight in Lnc; encourages dimensions to be uniformly distributed.
  • mu, delta = n/a
    Weights on node-level and cluster-level losses in J; tuned per dataset, and Figure 5 shows performance varies strongly with these.
assumptions (5)
  • domain assumption Representations H are orthogonal (H^T H = I)
    Explicit assumption of Theorems 2.2 and 2.3; not satisfied by prior SHGL methods.
  • standard math Ky Fan's theorem and Laplacian eigenvalue multiplicity (Lemma 2.4)
    Standard results used to justify the rank constraint and the spectral loss.
  • ad hoc to paper Minimizing Lsp with orthogonal Y yields the eigenspace of the c smallest eigenvalues of L_S and implies rank(L_S)=n-c
    The first part is Ky Fan's theorem; the rank implication is asserted without proof in Appendix C.3.
  • ad hoc to paper Lcc reduces the within-class scatter variances sigma_0^2 and sigma_1^2 of heterogeneous representations
    Needed for Theorem 2.6 but never derived; Lcc operates on projected Q, not on the linear heterogeneous encoder outputs used in the proof.
  • domain assumption The Davies-Bouldin-based complexity measure C and the generalization bound G from [33,16] are valid for comparing SHGL models
    Heuristic measures, not rigorous generalization bounds; the paper uses them as theorems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Self-Supervised Heterogeneous Graph Learning from Spectral Clustering Perspective." pith.science (2026). https://pith.science/paper/2RQQZBDM

@misc{pith2026241200742,
  author       = {Pith},
  title        = {Pith review of: Revisiting Self-Supervised Heterogeneous Graph Learning from Spectral Clustering Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2RQQZBDM}},
  note         = {Machine review of arXiv:2412.00742}
}
read the original abstract

Self-supervised heterogeneous graph learning (SHGL) has shown promising potential in diverse scenarios. However, while existing SHGL methods share a similar essential with clustering approaches, they encounter two significant limitations: (i) noise in graph structures is often introduced during the message-passing process to weaken node representations, and (ii) cluster-level information may be inadequately captured and leveraged, diminishing the performance in downstream tasks. In this paper, we address these limitations by theoretically revisiting SHGL from the spectral clustering perspective and introducing a novel framework enhanced by rank and dual consistency constraints. Specifically, our framework incorporates a rank-constrained spectral clustering method that refines the affinity matrix to exclude noise effectively. Additionally, we integrate node-level and cluster-level consistency constraints that concurrently capture invariant and clustering information to facilitate learning in downstream tasks. We theoretically demonstrate that the learned representations are divided into distinct partitions based on the number of classes and exhibit enhanced generalization ability across tasks. Experimental results affirm the superiority of our method, showcasing remarkable improvements in several downstream tasks compared to existing methods.

Figures

Figures reproduced from arXiv: 2412.00742 by the authors.

Figure 1
Figure 1. The flowchart of SCHOOL, which first employs the Multi-Layer Perception [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the affinity matrix S and t-SNE on DBLP and Aminer datasets. previous SHGL methods is equivalent to performing spectral clustering with additional regularization under the orthogonalization assumption. Then we proposed an efficient spectral clustering method with the rank constraint to learn an adaptive affinity matrix and mitigate noisy connections in previous methods. Moreover, we designed node-le… view at source ↗
Figure 3
Figure 3. Visualization plotted by t-SNE and the corresponding silhouette scores (SIL) of node [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Classification performance (i.e., Macro-F1) of the proposed method under different clusters. [PITH_FULL_IMAGE:figures/full_fig_p028_4.png]
Figure 5
Figure 5. Figure 5: The classification performance of the proposed method at different parameter settings [PITH_FULL_IMAGE:figures/full_fig_p029_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

95 extracted references · 72 canonical work pages

  1. [1]

    Sparse subspace clustering with entropy-norm

    Liang Bai and Jiye Liang. Sparse subspace clustering with entropy-norm. In ICML, pages 561–568, 2020

  2. [2]

    Piotr Bielak, Tomasz Kajdanowicz, and Nitesh V . Chawla. Graph barlow twins: A self-supervised representation learning framework for graphs. Knowledge-Based Systems, 256:109631, 2022

  3. [3]

    Convex optimization

    S Boyd, L Vandenberghe, and L Faybusovich. Convex optimization. IEEE Transactions on Automatic Control, 51(11):1859–1859, 2006

  4. [4]

    Chawla, and Ananthram Swami

    Yuxiao Dong, Nitesh V . Chawla, and Ananthram Swami. metapath2vec: Scalable representation learning for heterogeneous networks. In SIGKDD, pages 135–144, 2017

  5. [5]

    On a theorem of weyl concerning eigenvalues of linear transformations i

    Ky Fan. On a theorem of weyl concerning eigenvalues of linear transformations i. Proceedings of the National Academy of Sciences, 35(11):652–655, 1949

  6. [6]

    Algorithms for the qr decomposition

    Walter Gander. Algorithms for the qr decomposition. Res. Rep, 80(02):1251–1268, 1980

  7. [7]

    Provable guarantees for self- supervised deep learning with spectral contrastive loss

    Jeff Z HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma. Provable guarantees for self- supervised deep learning with spectral contrastive loss. In NeurIPS, pages 5000–5011, 2021. 10

  8. [8]

    Contrastive multi-view representation learning on graphs

    Kaveh Hassani and Amir Hosein Khasahmadi. Contrastive multi-view representation learning on graphs. In ICML, pages 4116–4126, 2020

Show all 95 references
  1. [9]

    Analyzing heterogeneous networks with missing attributes by unsupervised contrastive learning

    Dongxiao He, Chundong Liang, Cuiying Huo, Zhiyong Feng, Di Jin, Liang Yang, and Weixiong Zhang. Analyzing heterogeneous networks with missing attributes by unsupervised contrastive learning. IEEE Transactions on Neural Networks and Learning Systems, 2022

  2. [10]

    Contrastive learning meets homophily: two birds with one stone

    Dongxiao He, Jitao Zhao, Rui Guo, Zhiyong Feng, Di Jin, Yuxiao Huang, Zhen Wang, and Weixiong Zhang. Contrastive learning meets homophily: two birds with one stone. In ICML, pages 12775–12789, 2023

  3. [11]

    Adversarial learning on heterogeneous information networks

    Binbin Hu, Yuan Fang, and Chuan Shi. Adversarial learning on heterogeneous information networks. In SIGKDD, pages 120–129, 2019

  4. [12]

    Leveraging meta-path based context for top-n recommendation with a neural co-attention model

    Binbin Hu, Chuan Shi, Wayne Xin Zhao, and Philip S Yu. Leveraging meta-path based context for top-n recommendation with a neural co-attention model. In SIGKDD, pages 1531–1540, 2018

  5. [13]

    Heterogeneous graph transformer

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. Heterogeneous graph transformer. In WWW, pages 2704–2710, 2020

  6. [14]

    Deep semantic clustering by partition confi- dence maximisation

    Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confi- dence maximisation. In CVPR, pages 8849–8858, 2020

  7. [15]

    Multiplex graph representation learning via bi-level optimization

    Yudi Huang, Yujie Mo, Yujing Liu, Ci Nie, Guoqiu Wen, and Xiaofeng Zhu. Multiplex graph representation learning via bi-level optimization. In IJCAI, 2024

  8. [16]

    Methods and analysis of the first competition in predicting generalization of deep learning

    Yiding Jiang, Parth Natekar, Manik Sharma, Sumukh K Aithal, Dhruva Kashyap, Natarajan Subramanyam, Carlos Lassance, Daniel M Roy, Gintare Karolina Dziugaite, Suriya Gunasekar, et al. Methods and analysis of the first competition in predicting generalization of deep learning. I...

  9. [17]

    Dual intent enhanced graph neural network for session-based new item recommendation

    Di Jin, Luzhi Wang, Yizhen Zheng, Guojie Song, Fei Jiang, Xiang Li, Wei Lin, and Shirui Pan. Dual intent enhanced graph neural network for session-based new item recommendation. In WWW, pages 684–693, 2023

  10. [18]

    Hdmi: High-order deep multiplex infomax

    Baoyu Jing, Chanyoung Park, and Hanghang Tong. Hdmi: High-order deep multiplex infomax. In WWW, pages 2414–2424, 2021

  11. [19]

    Diederik Kingma and Lei Jimmy Ba

    P. Diederik Kingma and Lei Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015

  12. [20]

    Thomas Kipf and Max Welling

    N. Thomas Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, pages 1–14, 2017

  13. [21]

    Deep spectral clustering learning

    Marc T Law, Raquel Urtasun, and Richard S Zemel. Deep spectral clustering learning. InICML, pages 1985–1994, 2017

  14. [22]

    A survey of knowledge graph reasoning on graph types: Static, dynamic, and multimodal

    Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, Xinwang Liu, and Fuchun Sun. A survey of knowledge graph reasoning on graph types: Static, dynamic, and multimodal. arXiv preprint arXiv:2212.05767, 2022

  15. [23]

    Self-supervised temporal graph learning with temporal and structural intensity alignment

    Meng Liu, Ke Liang, Yawei Zhao, Wenxuan Tu, Sihang Zhou, Xinbiao Gan, Xinwang Liu, and Kunlun He. Self-supervised temporal graph learning with temporal and structural intensity alignment. IEEE Transactions on Neural Networks and Learning Systems, 2024

  16. [24]

    Deep temporal graph clustering

    Meng Liu, Yue Liu, Ke Liang, Wenxuan Tu, Siwei Wang, Sihang Zhou, and Xinwang Liu. Deep temporal graph clustering. In ICLR, 2024

  17. [25]

    Revisiting graph contrastive learning from the perspective of graph spectrum

    Nian Liu, Xiao Wang, Deyu Bo, Chuan Shi, and Jian Pei. Revisiting graph contrastive learning from the perspective of graph spectrum. In NeurIPS, volume 35, pages 2972–2983, 2022

  18. [26]

    Dink-net: Neural clustering on large graphs

    Yue Liu, Ke Liang, Jun Xia, Sihang Zhou, Xihong Yang, Xinwang Liu, and Stan Z Li. Dink-net: Neural clustering on large graphs. In ICML, 2023

  19. [27]

    Relation structure-aware heterogeneous information network embedding

    Yuanfu Lu, Chuan Shi, Linmei Hu, and Zhiyuan Liu. Relation structure-aware heterogeneous information network embedding. In AAAI, pages 4456–4463, 2019

  20. [28]

    Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks

    Qingsong Lv, Ming Ding, Qiang Liu, Yuxiang Chen, Wenzheng Feng, Siming He, Chang Zhou, Jianguo Jiang, Yuxiao Dong, and Jie Tang. Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks. In SIGKDD, pages 1150–1160, 2021. 11

  21. [29]

    Disentan- gled multiplex graph representation learning

    Yujie Mo, Yajie Lei, Jialie Shen, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Disentan- gled multiplex graph representation learning. In ICML, pages 24983–25005, 2023

  22. [30]

    Self-supervised heterogeneous graph learning: a homophily and heterogeneity view

    Yujie Mo, Feiping Nie, Zheng Zhang, Ping Hu, Heng Tao Shen, Xinchao Wang, and Xiaofeng Zhu. Self-supervised heterogeneous graph learning: a homophily and heterogeneity view. In ICLR, 2024

  23. [31]

    Multiplex graph representation learning via common and private information mining

    Yujie Mo, Zongqian Wu, Yuhuan Chen, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Multiplex graph representation learning via common and private information mining. In AAAI, pages 9217–9225, 2023

  24. [32]

    The laplacian spectrum of graphs

    Bojan Mohar, Y Alavi, G Chartrand, and OR Oellermann. The laplacian spectrum of graphs. Graph theory, combinatorics, and applications, 2(871-898):12, 1991

  25. [33]

    Representation based complexity measures for predicting generalization in deep learning

    Parth Natekar and Manik Sharma. Representation based complexity measures for predicting generalization in deep learning. arXiv preprint arXiv:2012.02775, 2020

  26. [34]

    Clustering and projected clustering with adaptive neighbors

    Feiping Nie, Xiaoqian Wang, and Heng Huang. Clustering and projected clustering with adaptive neighbors. In SIGKDD, pages 977–986, 2014

  27. [35]

    Self-weighted clustering with adaptive neighbors

    Feiping Nie, Danyang Wu, Rong Wang, and Xuelong Li. Self-weighted clustering with adaptive neighbors. IEEE Transactions on Neural Networks and Learning Systems, pages 3428–3441, 2020

  28. [36]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

  29. [37]

    Multi-view contrastive graph clustering

    Erlin Pan and Zhao Kang. Multi-view contrastive graph clustering. In NeurIPS, volume 34, pages 2148–2159, 2021

  30. [38]

    Unsupervised attributed multiplex network embedding

    Chanyoung Park, Donghyun Kim, Jiawei Han, and Hwanjo Yu. Unsupervised attributed multiplex network embedding. In AAAI, pages 5371–5378, 2020

  31. [39]

    Grlc: Graph representation learning with constraints

    Liang Peng, Yujie Mo, Jie Xu, Jialie Shen, Xiaoshuang Shi, Xiaoxiao Li, Heng Tao Shen, and Xiaofeng Zhu. Grlc: Graph representation learning with constraints. IEEE Transactions on Neural Networks and Learning Systems, 2023

  32. [40]

    Graph representation learning via graphical mutual information maximization

    Zhen Peng, Wenbing Huang, Minnan Luo, Qinghua Zheng, Yu Rong, Tingyang Xu, and Junzhou Huang. Graph representation learning via graphical mutual information maximization. InWWW, pages 259–270, 2020

  33. [41]

    Deepwalk: Online learning of social representations

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social representations. In SIGKDD, pages 701–710, 2014

  34. [42]

    Spectralnet: Spectral clustering using deep neural networks

    Uri Shaham, Kelly Stanton, Henry Li, Ronen Basri, Boaz Nadler, and Yuval Kluger. Spectralnet: Spectral clustering using deep neural networks. In ICLR, 2018

  35. [43]

    Pitfalls of graph neural network evaluation

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

  36. [44]

    Heterogeneous information network embedding for recommendation

    Chuan Shi, Binbin Hu, Wayne Xin Zhao, and S Yu Philip. Heterogeneous information network embedding for recommendation. IEEE Transactions on Knowledge and Data Engineering , 31(2):357–370, 2018

  37. [45]

    A survey of heterogeneous information network analysis

    Chuan Shi, Yitong Li, Jiawei Zhang, Yizhou Sun, and S Yu Philip. A survey of heterogeneous information network analysis. IEEE Transactions on Knowledge and Data Engineering, pages 17–37, 2016

  38. [46]

    Contrastive learning is spectral clustering on similarity graph

    Zhiquan Tan, Yifan Zhang, Jingqin Yang, and Yang Yuan. Contrastive learning is spectral clustering on similarity graph. In ICLR, 2024

  39. [47]

    Unified one- step multi-view spectral clustering

    Chang Tang, Zhenglai Li, Jun Wang, Xinwang Liu, Wei Zhang, and En Zhu. Unified one- step multi-view spectral clustering. IEEE Transactions on Knowledge and Data Engineering, 35(6):6449–6460, 2022

  40. [48]

    Yijun Tian, Kaiwen Dong, Chunhui Zhang, Chuxu Zhang, and Nitesh V . Chawla. Heterogeneous graph masked autoencoders. In AAAI, pages 9997–10005, 2023

  41. [49]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(11), 2008

  42. [50]

    Graph attention networks

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In ICLR, pages 1–12, 2018. 12

  43. [51]

    Hamilton, Pietro Liò, Yoshua Bengio, and R

    Petar Velickovic, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R. Devon Hjelm. Deep graph infomax. In ICLR, pages 1–17, 2019

  44. [52]

    A tutorial on spectral clustering.Statistics and computing, 17:395–416, 2007

    Ulrike V on Luxburg. A tutorial on spectral clustering.Statistics and computing, 17:395–416, 2007

  45. [53]

    Spectral embedded adaptive neighbors clustering

    Qi Wang, Zequn Qin, Feiping Nie, and Xuelong Li. Spectral embedded adaptive neighbors clustering. IEEE Transactions on Neural Networks and Learning Systems, pages 1265–1271, 2018

  46. [54]

    Image segmentation with ratio cut

    Song Wang and Jeffrey Mark Siskind. Image segmentation with ratio cut. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(6):675–690, 2003

  47. [55]

    A survey on heterogeneous graph embedding: methods, techniques, applications and sources

    Xiao Wang, Deyu Bo, Chuan Shi, Shaohua Fan, Yanfang Ye, and S Yu Philip. A survey on heterogeneous graph embedding: methods, techniques, applications and sources. IEEE Transactions on Big Data, 9(2):415–436, 2022

  48. [56]

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S. Yu. Hetero- geneous graph attention network. In WWW, pages 2022–2032, 2019

  49. [57]

    Self-supervised heterogeneous graph neural network with co-contrastive learning

    Xiao Wang, Nian Liu, Hui Han, and Chuan Shi. Self-supervised heterogeneous graph neural network with co-contrastive learning. In SIGKDD, pages 1726–1736, 2021

  50. [58]

    Heteroge- neous graph contrastive multi-view learning

    Zehong Wang, Qi Li, Donghua Yu, Xiaolong Han, Xiao-Zhi Gao, and Shigen Shen. Heteroge- neous graph contrastive multi-view learning. In SDM, pages 136–144, 2023

  51. [59]

    Molformer: Motif-based transformer on 3d heterogeneous molecular graphs

    Fang Wu, Dragomir Radev, and Stan Z Li. Molformer: Motif-based transformer on 3d heterogeneous molecular graphs. In AAAI, volume 37, pages 5312–5320, 2023

  52. [60]

    Self-supervised learning on graphs: Contrastive, generative, or predictive

    Lirong Wu, Haitao Lin, Cheng Tan, Zhangyang Gao, and Stan Z Li. Self-supervised learning on graphs: Contrastive, generative, or predictive. IEEE Transactions on Knowledge and Data Engineering, pages 1–20, 2021

  53. [61]

    Decoupled self-supervised learning for graphs

    Teng Xiao, Zhengyu Chen, Zhimeng Guo, Zeyang Zhuang, and Suhang Wang. Decoupled self-supervised learning for graphs. In NeurIPS, 2022

  54. [62]

    Self-supervised graph neural networks via diverse and interactive message passing

    Liang Yang, Cheng Chen, Weixun Li, Bingxin Niu, Junhua Gu, Chuan Wang, Dongxiao He, Yuanfang Guo, and Xiaochun Cao. Self-supervised graph neural networks via diverse and interactive message passing. In AAAI, pages 4327–4336, 2022

  55. [63]

    Self-supervised graph neural networks via low-rank decomposition

    Liang Yang, Runjie Shi, Qiuliang Zhang, Zhen Wang, Xiaochun Cao, Chuan Wang, et al. Self-supervised graph neural networks via low-rank decomposition. In NeurIPS, 2024

  56. [64]

    Dealmvc: Dual contrastive calibration for multi-view clustering

    Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In ACM MM, pages 337–346, 2023

  57. [65]

    Cluster-guided contrastive graph clustering network

    Xihong Yang, Yue Liu, Sihang Zhou, Siwei Wang, Wenxuan Tu, Qun Zheng, Xinwang Liu, Liming Fang, and En Zhu. Cluster-guided contrastive graph clustering network. In AAAI, pages 10834–10842, 2023

  58. [66]

    Deep spectral clustering using dual autoencoder network

    Xu Yang, Cheng Deng, Feng Zheng, Junchi Yan, and Wei Liu. Deep spectral clustering using dual autoencoder network. In CVPR, pages 4066–4075, 2019

  59. [67]

    Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning

    Xingtong Yu, Yuan Fang, Zemin Liu, and Xinming Zhang. Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning. In AAAI, pages 16578–16586, 2024

  60. [68]

    Molecular representation learning via heterogeneous motif graph neural networks

    Zhaoning Yu and Hongyang Gao. Molecular representation learning via heterogeneous motif graph neural networks. In ICML, pages 25581–25594, 2022

  61. [69]

    A simple meta-path-free framework for heterogeneous network embedding

    Rui Zhang, Arthur Zimek, and Peter Schneider-Kamp. A simple meta-path-free framework for heterogeneous network embedding. In CIKM, pages 2600–2609, 2022

  62. [70]

    Costa: Covariance- preserving feature augmentation for graph contrastive learning

    Yifei Zhang, Hao Zhu, Zixing Song, Piotr Koniusz, and Irwin King. Costa: Covariance- preserving feature augmentation for graph contrastive learning. In SIGKDD, pages 2524–2534, 2022

  63. [71]

    Heterogeneous graph structure learning for graph neural networks

    Jianan Zhao, Xiao Wang, Chuan Shi, Binbin Hu, Guojie Song, and Yanfang Ye. Heterogeneous graph structure learning for graph neural networks. In AAAI, pages 4697–4705, 2021

  64. [72]

    Collaborative knowledge distillation for heterogeneous information network embedding

    Sheng Zhou, Kang Yu, Defang Chen, Bolang Li, Yan Feng, and Chun Chen. Collaborative knowledge distillation for heterogeneous information network embedding. In WWW, pages 1631–1639, 2022. 13

  65. [73]

    Multi-view spectral clustering with optimal neighborhood laplacian matrix

    Sihang Zhou, Xinwang Liu, Jiyuan Liu, Xifeng Guo, Yawei Zhao, En Zhu, Yongping Zhai, Jianping Yin, and Wen Gao. Multi-view spectral clustering with optimal neighborhood laplacian matrix. In AAAI, pages 6965–6972, 2020

  66. [74]

    Structure-enhanced heterogeneous graph contrastive learning

    Yanqiao Zhu, Yichen Xu, Hejie Cui, Carl Yang, Qiang Liu, and Shu Wu. Structure-enhanced heterogeneous graph contrastive learning. In SDM, pages 82–90, 2022

  67. [75]

    Deep graph contrastive representation learning

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Deep graph contrastive representation learning. arXiv preprint arXiv:2006.04131, 2020

  68. [76]

    Graph contrastive learning with adaptive augmentation

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Graph contrastive learning with adaptive augmentation. In WWW, pages 2069–2080, 2021. 14 A Related Work This section briefly reviews topics related to this work, including self-supervised heterogeneous graph l...

  69. [77]

    [ 46] demonstrates that contrastive learning with the standard InfoNCE loss is equivalent to spectral clustering on the similarity graph

    demonstrates contrastive learning performs spectral clustering on the population augmentation graph by replacing the standard InfoNCE [ 36] with its proposed spectral contrastive loss. [ 46] demonstrates that contrastive learning with the standard InfoNCE loss is equivalent to...

  70. [78]

    (23), we can rewrite Eq

    (24) Based on Eq. (23), we can rewrite Eq. (24) as: min θ nX i ||z(1) i − z(2) i ||2 2 = min θ nX i ||hi + {hj, vj ∈ N(vi)(1)} −hi − {hk, vk ∈ N(vi)(2)}||2 2 = min θ nX i ||hi − {hk, vk ∈ N(vi)(2)} + {hj, vj ∈ N(vi)(1)} −hi||2 2 = min θ nX i ||hi − {hk, vk ∈ N(vi)(2)}||2 2 + |...

  71. [79]

    Therefore, based on Eq

    (28) where Tr(·) indicates the matrix trace. Therefore, based on Eq. (25) and Eq. (28), we can obtain min θ nX i ||z(1) i − z(2) i ||2 2 = min θ 2Tr(HT LH) + 2 nX i ⟨(hi − {hk, vk ∈ N(vi)(2)}) · (hj, vj ∈ N(vi)(1)} −hi})⟩ = min θ 2Tr(HT LH) + 2 nX i,j,k Gi,jGi,k⟨(hi − hj) · (h...

  72. [80]

    Hetero” and “Homo

    and the spectral clustering. Denote the self-expressive matrix in [ 30] as S, and denote the representations after projection by linear transformation as H. Moreover, denote DS as the degree matrix of S and denote LS = DS − S as the graph Laplacian. Given that the self-express...

  73. [81]

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes]

  74. [82]

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] We discuss the limitations of the work in Section 4

  75. [83]

    Theory Assumptions and Proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] We provide the assumptions and complete proof in Appendix C

  76. [84]

    Experimental Result Reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  77. [85]

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] We released codes and data at https://gith...

  78. [86]

    Experimental Setting/Details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] We specify all the training and test details...

  79. [87]

    Experiment Statistical Significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes]

  80. [88]

    Experiments Compute Resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] We list the details of experiments com...

  81. [89]

    Code Of Ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes]

  82. [90]

    Broader Impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] We discuss broder impacts in Section 4

  83. [91]

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? 30 Answer: [NA]

  84. [92]

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes]

  85. [93]

    New Assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes]

  86. [94]

    Crowdsourcing and Research with Human Subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  87. [95]

    Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.