Pith. sign in

REVIEW 3 major objections 6 minor 48 references

Robust Deep Signed Graph Clustering via Weak Balance Theory

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing Social Balance's 'enemy of my enemy is my friend' with Weak Balance's relaxed rule, plus edge-sign rewiring, makes deep K-way signed clustering more accurate and robust than every baseline it tests.

desk verdict The weak-balance pipeline is interesting, but VS-R as written cannot flip any edge on the sparse graphs tested, so the paper's central robustness claim is contradicted by its own equations. read the letter →

arxiv 2502.05472 v1 pith:2U3YWYXX submitted 2025-02-08 cs.SI

Peiyao Zhao , Xin Li , Zeyu Zhang , Mingzhong Wang , Xueying Zhu , Lejian Liao This is my paper · ORCID
classification cs.SI
keywords signedgraphclusteringweakbalancetheorysocialneuralnetworksrewiringnoiserobustnessK-waystochasticblockmodel
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

Signed graphs mark every link as positive or negative—friend or enemy, trust or distrust—which makes them expressive models for social and biological networks, but also fragile targets for clustering. The paper argues that the classical Social Balance rule 'an enemy of my enemy is my friend' implicitly presumes only two clusters and, in graphs with K>2 communities, pulls boundary nodes toward the wrong cluster. It proposes DSGC, an unsupervised deep clustering framework built on Weak Balance Theory, which replaces that rule with the weaker 'an enemy of my enemy might be my enemy.' Before learning, DSGC rewires the graph: Violation Sign-Refine flips the signs of edges it judges noisy, and Density-based Augmentation adds positive edges inside clusters and negative edges across clusters. Across the paper's 20 SSBM configurations and two real-world networks, DSGC reports higher accuracy and robustness than every spectral and deep baseline it compares against, with ablations tracing most of the gain to the rewiring steps.

What carries the argument

The load-bearing machinery is the rewiring pipeline defined by Eqs. (2)–(7). Definition 1 redefines a positive walk as one whose edges are all positive, and a negative walk as one with exactly one negative edge and the rest positive; Lemma 1 then expresses the difference of their counts as a matrix identity involving powers of the positive adjacency and one negative adjacency. The non-noise score $\Gamma_{ij}(L') = \sum_{l=1}^{L'} \alpha_l (\mu^+_l(i,j) - \mu^-_l(i,j))$ aggregates these counts up to length $L'$ with weights $\alpha_l$ that decay with $l$, and Eq. (5) thresholds $\Gamma$ at $\delta_+$ and $\delta_-$ to decide whether each edge should become $+1$, stay as is, or become $-1$. Density-based Augmentation follows, adding edges along long positive and negative walks via the matrix powers in Eq. (6). On the encoding side, the signed convolutions of Eqs. (9)–(10) are the second piece of machinery: the positive branch aggregates along positive walks, and the negative branch uses $-\bar{A}^-$ so that negative edges repel embeddings, widening cluster boundaries. The soft-assignment matrix $\Pi$ and the differentiable clustering loss of Eq. (14) with the volume regularizer of Eq. (15) close the loop, making the whole system trainable end-to-end.

What would settle it

Generate SSBM(1000, 5, 0.01, $\eta$) graphs at $\eta = 0.15$, $0.25$, and $0.35$ and measure DSGC's ACC against SPONGE: if the gap shrinks or reverses as $\eta$ approaches $0.4$, the robustness claim is confined to the low-noise regime the paper tested ($\eta \le 0.08$). A second check is to cluster a graph with a known K=2 structure inside a K=5 layout and see whether VS-R, which assumes violations are noise, erases legitimate boundaries.

Watch

Extended reading notes

Core claim

The central claim is that K-way signed graph clustering should be built on Weak Balance Theory rather than Social Balance Theory, and that the difference is measurable: the 'friend-of-my-enemy' reasoning of strict balance narrows cluster boundaries, while the relaxed 'enemy of my enemy might be my enemy' keeps nodes connected by negative edges well separated. To make that work, DSGC operates in two stages. In the preprocessing stage it overwrites the adjacency matrix: Violation Sign-Refine (VS-R) computes a non-noise score $\Gamma_{ij}$ from redefined positive and negative walk counts and flips edge signs that fall beyond thresholds $\delta_+$ and $\delta_-$; Density-based Augmentation (DA) then takes powers of the refined adjacency matrices to add positive edges within clusters and negative edges across clusters. In the learning stage, a signed convolutional encoder produces separate positive and negative embeddings, with the negative branch carrying an explicit minus sign so that negatively linked nodes push apart, and a regularized soft-assignment loss drives the final K-way partition. The evidence for the claim is a 20-setting SSBM benchmark plus the S&P1500 and Rainfall networks, where DSGC beats every baseline on ACC, NMI, ARI, and F1; ablations show that DA contributes the most, VS-R second, and that dropping the 'EEF' rule matters less than keeping the negative repulsion term in the encoder.

Load-bearing premise

The method stands or falls on the assumption that noisy edges are sparse enough for a hand-tuned score over high-order walks, with fixed weights and two thresholds, to tell real edges from noisy ones before learning begins; if that rewiring mislabels edges, every later stage trains on a corrupted graph.

Editorial extensions

If this is right

  • Table 2 shows that applying VS-R alone improves the ACC and NMI of spectral baselines such as BNC, SPONGE, and SPONGE_sym, so the denoising step can be used as a general preprocessor for other signed clustering methods.
  • If the EEF rule is as harmful as the paper argues, signed GNNs built for link prediction should not be transplanted unchanged into clustering; clustering-oriented encoders should derive their aggregation from Weak Balance instead.
  • The differentiable reformulation of the signed clustering objective means the NP-hard partition problem can be optimized end-to-end without labels, which is what allows the whole pipeline to train jointly.
  • The reported gains across K = 4 to 10 and N up to 1200 indicate that, within the tested regime, the method scales in cluster count and graph size while retaining an advantage over spectral baselines.

Reading between the lines

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

  • The robustness claim is probably confined to a moderate-noise window: the $\Gamma$ score uses hand-set weights and thresholds, so when sign flips become frequent ($\eta$ beyond roughly 0.15–0.2) the high-order walk 'vote' itself becomes unreliable and DSGC's advantage over spectral methods should erode.
  • The augmentation via matrix powers is a form of graph diffusion; replacing the hand-picked powers $m_+$ and $m_-$ with learnable or standard diffusion kernels (e.g., personalized PageRank) could yield similar or better rewiring with fewer hyperparameters.
  • The ablation showing DA as the most influential component suggests that edge addition, not sign correction, drives most of the performance; a direct test would be a version of DSGC that only runs DA on the raw graph, skipping VS-R entirely on low-noise graphs.
  • The argument that 'EEF' narrows boundaries is qualitative; converting the SoEN metric into a formal notion of cluster-boundary width could make the claim quantitative and testable across different balance axioms.
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 / 6 minor

Summary. The paper proposes DSGC, an unsupervised deep signed graph clustering framework built on Weak Balance Theory. It contains two graph rewiring preprocessing modules: Violation Sign-Refine (VS-R), intended to correct noisy edge signs using high-order walk counts, and Density-based Augmentation (DA), which adds positive edges within clusters and negative edges across clusters. The framework then employs a two-channel signed graph encoder that aggregates positive and negative walks while deliberately omitting the 'enemy of my enemy is my friend' rule, and optimizes a differentiable soft-assignment clustering loss. The authors evaluate on 20 synthetic SSBM configurations spanning noise level, sparsity, graph size, and cluster count, plus two real-world graphs, and report consistent superiority over 17 baselines in ACC, NMI, ARI, and F1.

Significance. The framework is clearly motivated and the experimental scope is broad: 20 SSBM configurations, two real-world graphs, a public code release, and ablations of each component. The weak-balance encoder and the soft violation-minimization loss are coherent and reasonable design choices. However, the central robustness claim rests on the VS-R denoising mechanism, and that mechanism, as specified in Eq. (5) with the thresholds reported in App. D, is internally inconsistent and cannot have produced the reported ablation and preprocessing gains. Until this is resolved, the stated contribution is not supported by the experiments as described.

major comments (3)
  1. [§4.1.1, Eqs. (3)–(5) and App. D] The VS-R rule cannot flip any edge on the evaluated SSBM graphs. For any existing edge, the l=1 term in Eq. (3) equals the observed sign A_ij ∈ {±1} exactly, so Γ_ij = ±1 + Σ_{l≥2} α_l (μ+_l − μ−_l). On the p=0.01–0.04 SSBM graphs used in Tables 1–2, the l≥2 corrections have expected magnitude below roughly 0.3 even at p=0.04, and closer to 0.02–0.05 at p=0.01–0.02 (e.g., (A+)^2_ij ≈ (N/K)p^2 ≈ 0.02 at N=1000, K=5, p=0.01). Consequently, for a negative edge Γ_ij remains above −1, and for a positive edge Γ_ij sits near +1 without crossing the δ+=1 threshold in a way that would change the sign. With δ+=1 and δ−=−1, neither the 'set to +1' branch nor the 'set to −1' branch of Eq. (5) can be triggered for an existing edge; hence  = A identically. This contradicts the large ACC/NMI improvements and violation-ratio reductions attributed to VS-R in Table 2 and the ablation curves in Fig. 4. The authors must either exclude the l=1 term from Γ, choose thresholds near 0, or specify a different correction rule, and then present experiments that actually exercise that rule.
  2. [Table 2 and §5.4] The reported violation-ratio reductions are impossible if VS-R is the identity. If no edge signs are changed, the violation ratio (defined as the ratio of violated to non-violated edges) is unchanged; however, the text states that VS-R 'consistently reduces the violation ratio' across the tested configurations. The discrepancy between the described algorithm and the reported Table 2 numbers must be resolved before the robustness claim can be evaluated.
  3. [App. D and Eq. (5)] The instruction that 'optimal performance is achieved when both δ+ and δ− are set to 1' directly violates the constraint δ− < 0 stated after Eq. (5), and it makes the 'retain' interval δ− ≤ Γ ≤ δ+ degenerate (it becomes Γ = 1 exactly). Fig. 13(a) shows ACC against a single δ axis up to 12, but the text does not specify whether δ+ and δ− are varied together, separately, or with what sign convention. This is not a cosmetic issue: the threshold values completely determine whether VS-R modifies any edge, so the reported hyperparameter analysis must be made consistent with the algorithm definition.
minor comments (6)
  1. [Eq. (14)] Specify whether L+ and A− in the clustering loss are built from the original adjacency matrix A or from the augmented A′′, since the encoder uses the latter; the loss–encoder consistency is otherwise unclear.
  2. [Abstract and §5.2] The paper says '15 baselines', but App. B enumerates 9 spectral, 2 deep signed (SiNE, SNEA), and 6 deep unsigned methods, totaling 17; the count should be aligned.
  3. [§5.2 and App. G] All results are averages over 5 runs without standard deviations or significance tests; given the claim of consistent superiority, at least standard deviations or error bars should be reported.
  4. [Notation] The symbol L is overloaded: it denotes both the signed Laplacian baseline (Table 1) and the clustering loss (Eq. (14)); rename one to avoid confusion.
  5. [§1] The phrase 'first Deep Signed Graph Clustering framework' should be softened, as SSSNET [16] is a deep signed clustering method, albeit semi-supervised.
  6. [Lemma 1] A one-line proof or citation of the walk-count identity in Eq. (2) would help readers verify the lemma.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DSGC components are constructive heuristics, the loss is the standard signed-cut objective, hyperparameters are validation-selected, and no load-bearing claim reduces to a fitted input or self-citation.

full rationale

The derivation chain in DSGC is self-contained and does not reduce any prediction to a fitted input. The non-noise score Γ (Eq. 3) and the refinement rule (Eq. 5) are constructive definitions built from walk counts; they are not defined in terms of the clustering assignments or the evaluation metrics. The clustering loss (Eq. 14) is the standard signed-cut violation objective made differentiable by replacing hard assignments with soft probabilities, so minimizing it is the stated objective rather than a post-hoc fit. Hyperparameters (δ±, m±, λ, L) are chosen by validation (App. D) and then applied across datasets; no parameter is fitted to the test configurations to manufacture a claimed prediction. The paper's self-citations ([43]–[45] by coauthor Z. Zhang) cite prior work on signed link prediction and graph augmentation; they are not invoked as a uniqueness theorem or as the justification for the central clustering claim. Weak Balance Theory is attributed to Davis [7], an external source, and the redefinition of positive/negative walks in Definition 1 is explicit rather than smuggled via citation. I therefore find no circular step. Separately, the reported optimal setting "both δ+ and δ− are set to 1" (App. D) conflicts with Eq. (5)'s requirement δ− < 0, and with δ± = 1 the refinement interval behavior would differ from the stated "otherwise retain" rule; this is a correctness/reproducibility inconsistency, not a circularity, and it does not affect the circularity score.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several hand-chosen hyperparameters (δ±, m±, λ, L, L', ε±) and on domain assumptions about Weak Balance and walk-based denoising. No new physical entities or external constants are introduced.

free parameters (9)
  • δ+ = 1 (from App D sensitivity analysis)
    Threshold for positive confidence in Violation Sign-Refine, Eq. (5). Chosen by grid search on SSBM(1000,5,0.01,0.02).
  • δ- = -1 (from App D sensitivity analysis)
    Threshold for negative confidence, Eq. (5), chosen same way.
  • m+ = 3 (from App D sensitivity analysis)
    Walk length for positive augmentation in Eq. (6), chosen on same SSBM.
  • m- = 2 (from App D sensitivity analysis)
    Walk length for negative augmentation in Eq. (6), chosen on same SSBM.
  • λ = 0.03 (App F)
    Regularization weight in Eq. (14).
  • L = 2 (App F)
    Number of encoder layers (receptive field).
  • L' = 3 (App F)
    Max walk length in VS-R non-noise score.
  • hidden dimension d = 32 (App F)
    Embedding dimension for positive and negative channels.
  • ε+, ε- = not reported
    Self-loop balance hyperparameters in Section 4.2; values are not specified in the manuscript.
assumptions (5)
  • domain assumption Weak Balance Theory (Davis 1967): K-way clusterable signed graphs have positive edges only within clusters and negative edges only between clusters.
    The paper's entire partition objective and encoder design assume this model of signed community structure; stated in Section 3.2 and App A.
  • ad hoc to paper Definition 1: positive walks are all-positive walks; negative walks are walks with exactly one negative edge.
    This redefinition is introduced to fit K-way weak balance, ignoring walks with two or more negative edges. Used in Lemma 1 and Eqs. (2)-(3).
  • standard math Lemma 1 combinatorial identity without proof.
    The identity in Eq. (2) is asserted as a lemma but no proof is given in main text or appendix; it is elementary but unverified in the manuscript.
  • ad hoc to paper Heuristic weighting α_l = 1/l! and the assumption that shorter walks are more informative for sign correction.
    Introduced without derivation in Eq. (4); the choice of weights affects Γ and thus the rewiring outcome.
  • domain assumption Assumption that noisy edges are sparse and high-order neighbor information revises them correctly.
    Stated as motivation in Section 4.1.1; if violations are dense, the denoising may compound errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Deep Signed Graph Clustering via Weak Balance Theory." pith.science (2026). https://pith.science/paper/2U3YWYXX

@misc{pith2026250205472,
  author       = {Pith},
  title        = {Pith review of: Robust Deep Signed Graph Clustering via Weak Balance Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2U3YWYXX}},
  note         = {Machine review of arXiv:2502.05472}
}
read the original abstract

Signed graph clustering is a critical technique for discovering community structures in graphs that exhibit both positive and negative relationships. We have identified two significant challenges in this domain: i) existing signed spectral methods are highly vulnerable to noise, which is prevalent in real-world scenarios; ii) the guiding principle ``an enemy of my enemy is my friend'', rooted in \textit{Social Balance Theory}, often narrows or disrupts cluster boundaries in mainstream signed graph neural networks. Addressing these challenges, we propose the \underline{D}eep \underline{S}igned \underline{G}raph \underline{C}lustering framework (DSGC), which leverages \textit{Weak Balance Theory} to enhance preprocessing and encoding for robust representation learning. First, DSGC introduces Violation Sign-Refine to denoise the signed network by correcting noisy edges with high-order neighbor information. Subsequently, Density-based Augmentation enhances semantic structures by adding positive edges within clusters and negative edges across clusters, following \textit{Weak Balance} principles. The framework then utilizes \textit{Weak Balance} principles to develop clustering-oriented signed neural networks to broaden cluster boundaries by emphasizing distinctions between negatively linked nodes. Finally, DSGC optimizes clustering assignments by minimizing a regularized clustering loss. Comprehensive experiments on synthetic and real-world datasets demonstrate DSGC consistently outperforms all baselines, establishing a new benchmark in signed graph clustering.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 3 linked inside Pith

  1. [1]

    Singh, and Sourav Me- dya

    Aritra Bhowmick, Mert Kosan, Zexi Huang, Ambuj K. Singh, and Sourav Me- dya. 2024. DGCLUSTER: A Neural Framework for Attributed Graph Clustering via Modularity Maximization. In Thirty-Eighth AAAI Conference on Artificial Intelligence. 11069–11077

    2024

  2. [2]

    Deyu Bo, Xiao Wang, Chuan Shi, Meiqi Zhu, Emiao Lu, and Peng Cui. 2020. Structural Deep Clustering Network. In The Web Conference 2020. 1400–1410

    2020

  3. [3]

    Dorwin Cartwright and Frank Harary. 1956. Structural balance: a generalization of Heider’s theory. Psychological review 63, 5 (1956), 277

    1956

  4. [4]

    Yiqi Chen, Tieyun Qian, Huan Liu, and Ke Sun. 2018. "Bridge": Enhanced Signed Directed Network Embedding. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM . 773–782

    2018

  5. [5]

    Kai-Yang Chiang, Joyce Jiyoung Whang, and Inderjit S. Dhillon. 2012. Scalable clustering of signed networks using balance normalized cut. In 21st ACM Inter- national Conference on Information and Knowledge Management, CIKM . 615–624

    2012

  6. [6]

    Mihai Cucuringu, Peter Davies, Aldo Glielmo, and Hemant Tyagi. 2019. SPONGE: A generalized eigenproblem for clustering signed networks. In The 22nd In- ternational Conference on Artificial Intelligence and Statistics, AISTATS , Vol. 89. 1088–1098

    2019

  7. [7]

    James A Davis. 1967. Clustering and structural balance in graphs.Human relations 20, 2 (1967), 181–187

    1967

  8. [8]

    Aggarwal, and Jiliang Tang

    Tyler Derr, Charu C. Aggarwal, and Jiliang Tang. 2018. Signed Network Modeling Based on Structural Balance Theory. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM . 557–566

    2018

Show all 48 references
  1. [9]

    Tyler Derr, Yao Ma, and Jiliang Tang. 2018. Signed Graph Convolutional Networks. In IEEE International Conference on Data Mining, ICDM . 929–934

    2018

  2. [10]

    Fernando Diaz-Diaz and Ernesto Estrada. 2024. Signed graphs in data sciences via communicability geometry. arXiv preprint arXiv:2403.07493 (2024)

    2024 arXiv

  3. [11]

    André Fujita, Patricia Severino, Kaname Kojima, João Ricardo Sato, Alexan- dre Galvão Patriota, and Satoru Miyano. 2012. Functional clustering of time series gene expression data by Granger causality. BMC Syst. Biol. 6 (2012), 137

    2012

  4. [12]

    Gates and Yong-Yeol Ahn

    Alexander J. Gates and Yong-Yeol Ahn. 2017. The Impact of Random Models on Clustering Similarity. J. Mach. Learn. Res. 18 (2017), 87:1–87:28

    2017

  5. [13]

    Amir Ghasemian, Homa Hosseinmardi, and Aaron Clauset. 2020. Evaluating Overfit and Underfit in Models of Network Community Structure. IEEE Trans. Knowl. Data Eng. 32, 9 (2020), 1722–1735

    2020

  6. [14]

    Fan Chung Graham, Alexander Tsiatas, and Wensong Xu. 2013. Dirichlet PageR- ank and Ranking Algorithms Based on Trust and Distrust. Internet Math. 9, 1 (2013), 113–134

    2013

  7. [15]

    Frank Harary. 1953. On the notion of balance of a signed graph. Michigan Mathematical Journal 2, 2 (1953), 143–146

    1953

  8. [16]

    Yixuan He, Gesine Reinert, Songchao Wang, and Mihai Cucuringu. 2022. SSSNET: Semi-Supervised Signed Network Clustering. In Proceedings of the 2022 SIAM International Conference on Data Mining, SDM . 244–252

    2022

  9. [17]

    Junjie Huang, Huawei Shen, Liang Hou, and Xueqi Cheng. 2019. Signed Graph At- tention Networks. In Artificial Neural Networks and Machine Learning , Vol. 11731. Springer, 566–577

    2019

  10. [18]

    Junjie Huang, Huawei Shen, Liang Hou, and Xueqi Cheng. 2021. SDGNN: Learn- ing Node Representation for Signed Directed Networks. In Thirty-Fifth AAAI Conference on Artificial Intelligence. 196–203

    2021

  11. [19]

    Aditya Prakash, and Naren Ramakrishnan

    Mohammad Raihanul Islam, B. Aditya Prakash, and Naren Ramakrishnan. 2018. SIGNet: Scalable Embeddings for Signed Networks. In Advances in Knowledge Discovery and Data Mining , Vol. 10938. 157–169

    2018

  12. [20]

    Jinhong Jung, Jaemin Yoo, and U Kang. 2020. Signed Graph Diffusion Network. (2020). arXiv:2012.14191

    2020 arXiv

  13. [21]

    Andrew V. Knyazev. 2001. Toward the Optimal Preconditioned Eigensolver: Locally Optimal Block Preconditioned Conjugate Gradient Method. SIAM J. Sci. Comput. 23, 2 (2001), 517–541

    2001

  14. [22]

    Jérôme Kunegis, Stephan Schmidt, Andreas Lommatzsch, Jürgen Lerner, Ernesto William De Luca, and Sahin Albayrak. 2010. Spectral Analysis of Signed Graphs for Clustering, Prediction and Visualization. In Proceedings of the SIAM Interna- tional Conference on Data Mining, SDM . 559–570

    2010

  15. [23]

    Yeon-Chang Lee, Nayoun Seo, and Sang-Wook Kim. 2020. Are Negative Links Really Beneficial to Network Embedding?: In-Depth Analysis and Interesting Results. In The 29th ACM International Conference on Information and Knowledge Management. 2113–2116

    2020

  16. [24]

    Yu Li, Meng Qu, Jian Tang, and Yi Chang. 2023. Signed Laplacian Graph Neural Networks. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence . AAAI Press, 4444–4452

    2023

  17. [25]

    Yu Li, Yuan Tian, Jiawei Zhang, and Yi Chang. 2020. Learning Signed Network Em- bedding via Graph Attention. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI. 4772–4779

    2020

  18. [26]

    Haoxin Liu, Ziwei Zhang, Peng Cui, Yafeng Zhang, Qiang Cui, Jiashuo Liu, and Wenwu Zhu. 2021. Signed Graph Neural Network with Latent Groups. InThe 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1066–1075

    2021

  19. [27]

    Yunfei Liu, Jintang Li, Yuehe Chen, Ruofan Wu, Ericbk Wang, Jing Zhou, Sheng Tian, Shuheng Shen, Xing Fu, Changhua Meng, Weiqiang Wang, and Liang Chen

  20. [28]

    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, ICML , Vol. 202. 21794–21812

    2023

  21. [29]

    Yue Liu, Wenxuan Tu, Sihang Zhou, Xinwang Liu, Linxuan Song, Xihong Yang, and En Zhu. 2022. Deep Graph Clustering via Dual Correlation Reduction. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI . 7603–7611

    2022

  22. [30]

    Pedro Mercado, Jessica Bosch, and Martin Stoll. 2019. Node Classification for Signed Social Networks Using Diffuse Interface Methods. In Machine Learning and Knowledge Discovery in Databases , Vol. 11906. 524–540

    2019

  23. [31]

    Pedro Mercado, Francesco Tudisco, and Matthias Hein. 2016. Clustering Signed Networks with the Geometric Mean of Laplacians. In Advances in Neural Infor- mation Processing Systems 29 . 4421–4429

    2016

  24. [32]

    Pedro Mercado, Francesco Tudisco, and Matthias Hein. 2019. Spectral Clustering of Signed Graphs via Matrix Power Means. InProceedings of the 36th International Conference on Machine Learning, ICML , Vol. 97. 4526–4536

    2019

  25. [33]

    Shirui Pan, Ruiqi Hu, Guodong Long, Jing Jiang, Lina Yao, and Chengqi Zhang

  26. [34]

    Yiru Pan, Xingyu Ji, Jiaqi You, Lu Li, Zhenping Liu, Xianlong Zhang, Zeyu Zhang, and Maojun Wang. 2024. CSGDN: Contrastive Signed Graph Diffusion Network for Predicting Crop Gene-Trait Associations. arXiv:2410.07511 (2024)

    2024 arXiv

  27. [35]

    Moshen Shahriari and Mahdi Jalili. 2014. Ranking Nodes in Signed Social Net- works. Soc. Netw. Anal. Min. 4, 1 (2014), 172

    2014

  28. [36]

    Wenxuan Tu, Sihang Zhou, Xinwang Liu, Xifeng Guo, Zhiping Cai, En Zhu, and Jieren Cheng. 2021. Deep Fusion Clustering Network. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI . 9978–9987

    2021

  29. [37]

    Ruo-Chun Tzeng, Bruno Ordozgoiti, and Aristides Gionis. 2020. Discovering con- flicting groups in signed networks. In Advances in Neural Information Processing Systems 33

    2020

  30. [38]

    Chun Wang, Shirui Pan, Ruiqi Hu, Guodong Long, Jing Jiang, and Chengqi Zhang

  31. [39]

    Aggarwal, Jiliang Tang, and Huan Liu

    Suhang Wang, Charu C. Aggarwal, Jiliang Tang, and Huan Liu. 2017. Attributed Signed Network Embedding. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, CIKM . ACM, 137–146

    2017

  32. [40]

    Aggarwal, Yi Chang, and Huan Liu

    Suhang Wang, Jiliang Tang, Charu C. Aggarwal, Yi Chang, and Huan Liu. 2017. Signed Network Embedding in Social Media. In Proceedings of the 2017 SIAM International Conference on Data Mining, Houston . SIAM, 327–335

    2017

  33. [41]

    Pinghua Xu, Wenbin Hu, Jia Wu, and Bo Du. 2019. Link Prediction with Signed Latent Factors in Signed Social Networks. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 1046–1054

    2019

  34. [42]

    Cheung, and Jiming Liu

    Bo Yang, William K. Cheung, and Jiming Liu. 2007. Community Mining from Signed Social Networks. IEEE Trans. Knowl. Data Eng. 19, 10 (2007), 1333–1348

    2007

  35. [43]

    Zeyu Zhang, Lu Li, Shuyan Wan, Sijie Wang, Zhiyi Wang, Zhiyuan Lu, Dong Hao, and Wanli Li. 2024. DropEdge not Foolproof: Effective Augmentation Method for Signed Graph Neural Networks. NeurIPS (2024)

    2024

  36. [44]

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

  37. [2018]

    In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI

    Adversarially Regularized Graph Autoencoder for Graph Embedding. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI. 2609–2615

  38. [2019]

    In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI

    Attributed Graph Clustering: A Deep Attentional Embedding Approach. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI. 3670–3676

  39. [2023]

    In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Contrastive learning for signed bipartite graphs. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1629–1638

  40. [2024]

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

    Revisiting Modularity Maximization for Graph Clustering: A Contrastive Learning Perspective. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1968–1979

    1968

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.