Pith. sign in

REVIEW 4 major objections 5 minor 57 references

Unnoticeable Community Deception via Multi-objective Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Community deception is best measured as a drop in adjusted Rand index, not modularity, and can be solved as a two-objective degree-preserving rewiring problem.

desk verdict Worth a serious look: the modularity critique is real and the Pareto-budget idea is a genuine improvement, but the evaluation leans on a self-referential ground truth that inflates the headline claims. read the letter →

arxiv 2509.01438 v1 pith:WHD7526G submitted 2025-09-01 cs.SI cs.AIcs.NE

Junyuan Fang , Huimin Liu , Yueqi Peng , Jiajing Wu , Zibin Zheng , Chi K. Tse This is my paper
classification cs.SIcs.AIcs.NE
keywords communitydetectiondeceptionhidingmulti-objectiveoptimizationNSGA-IIdegree-preservingrewiringadjustedRandindexadversarialgraphperturbation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that current community-deception attacks are judged by the wrong yardstick—how much modularity drops—and shows empirically that successful deception can even raise modularity. It proposes measuring deception as the decrease in adjusted Rand index (DARI), the drop in agreement between the detector's partition of the clean graph and its partition of the attacked graph. The attack is then cast as a two-objective optimization: maximize DARI while minimizing the number of rewired links, solved with NSGA-II using a rewiring operator that keeps every node's degree unchanged. The authors claim this yields a Pareto front of attacks from a single run, so an attacker can pick a solution for any available budget without rerunning the algorithm. On three benchmark networks and three detection algorithms, the two bias-guided variants produce Pareto fronts that match or dominate the modularity-based and autoencoder baselines.

What carries the argument

Degree-preserving rewiring: each mutation and crossover step deletes two existing links and inserts two new links so that every node's degree stays identical to its original value. This operator carries the 'unnoticeable' part of the attack and, combined with DICE-like community-biased selection, steers the NSGA-II search. The two objectives—DARI and DAT—define the Pareto landscape that the method explores.

What would settle it

Run the Label Propagation detector on the same clean graph many times: if its partition already fluctuates across restarts, then DARI gains on attacked graphs could come from detector randomness rather than attack effectiveness. More directly, generate synthetic graphs with planted community labels and measure UCD's DARI against those planted labels instead of against the detector's clean-graph output; if those scores collapse, the claim that DARI measures true community deception is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that community deception should be scored by how much the attacked detector's partition moves away from its own partition on the clean graph—the decrease of adjusted Rand index, DARI—rather than by the drop in modularity. Its synthetic experiments show that merging small communities or splitting a large one can raise modularity while still damaging community structure, so modularity-based objectives misreport attack success. The paper then frames deception as a two-objective maximization: DARI versus the inverse of attack budget (DAT), solved with NSGA-II. The key operator is degree-preserving rewiring: each perturbation removes two links and adds two links so th

Load-bearing premise

All deception scores rest on taking the detector's partition of the unmodified graph as the true community labels, so a large DARI only measures that the same detector changed its mind after rewiring—not that the graph's real communities became hideable.

Editorial extensions

If this is right

  • A single run of the multi-objective attack produces a Pareto front, so an attacker can choose a solution for any link budget without rerunning the optimizer.
  • All attacks generated by the three UCD variants preserve every node's degree exactly, so degree-distribution-based checks cannot detect the perturbation.
  • Because modularity can rise while deception succeeds, future community-hiding evaluations should report partition-distance metrics such as DARI, with modularity kept only as a secondary view.
  • The biased variants UCD(MIN) and UCD(MAX) consistently improve hypervolume over the unbiased UCD, indicating that degree- and community-informed node selection strengthens evolutionary deception attacks.

Reading between the lines

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

  • DARI is computed against the detector's own clean-graph partition, not against true planted communities; on synthetic graphs with known ground truth, scoring DARI against the planted labels would be a stricter test and could change the method's apparent ranking.
  • Preserving degrees does not guarantee global unnoticeability: a defender could inspect clustering coefficients, motif counts, or spectral fingerprints, which the current objective does not constrain.
  • The same two-objective Pareto framing likely transfers to other graph-privacy tasks, such as membership hiding or link obfuscation, where the trade-off is again privacy benefit versus number of structural edits.
  • The degree-preserving rewiring operator is defined for undirected unweighted graphs; extending it to directed or weighted networks would test whether the approach remains effective outside the current setting.
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

4 major / 5 minor

Summary. The paper proposes a new community-deception method, UCD, formulated as a bi-objective optimization problem in which the two objectives are DARI (the decrease of adjusted Rand index) and DAT (the decrease of attack budget). The method uses NSGA-II with degree-preserving rewiring, mutation, and crossover operators, and two variants, UCD(MIN) and UCD(MAX), bias the mutation toward low/high-degree nodes and cross-community links. The paper first argues that the widely used modularity-decrease metric is unreliable, using synthetic experiments in Table I, and then evaluates the proposed methods against GAQ and GCH on Karate, Dolphins, and Netscience under LOU, FN, and LPA detectors.

Significance. If the evaluation were fully convincing, the paper would make a useful contribution: the Table I counterexamples provide a concrete, reproducible illustration that modularity drop is not a faithful proxy for deception, and the degree-preserving rewiring operator is a clean mechanism for maintaining the degree sequence of the attacked graph. The multi-objective formulation also gives a budget-flexible set of attacks rather than a single budget-tuned solution, which is a practical improvement over GAQ. The central weakness is that the DARI metric is defined against the output of the same detector that is being attacked, so the reported effectiveness conflates 'fooling the detector' with 'hiding the underlying communities.' The proposed fixes are straightforward, so the paper is potentially publishable after major revision.

major comments (4)
  1. [III-C, Eq. (1)] The paper defines ground truth for DARI as 'we first employ a specific detection method on the clean graph and consider the obtained clustering labels as the ground truth labels.' This makes DARI a measure of how much the detector changes its own output, not whether the true community structure has been hidden. On Karate, a known ground-truth partition exists but is not used; on Netscience and Dolphins no external labels are used. For LPA, which is stochastic, the 'ground truth' is one arbitrary clean-graph run. All reported effectiveness claims (Fig. 6, Table III) depend on this choice. I request additional experiments with planted ground truth (e.g., LFR or the synthetic generator from Section III-B), plus results against the known Zachary labels, and mean/std over multiple LPA and NSGA-II runs.
  2. [V-C and V-E.1] The comparison against baselines is partly self-referential. UCD directly maximizes DARI under the target detector, while GAQ maximizes modularity decrease and GCH uses autoencoder prediction probabilities. Evaluating all methods on DARI therefore gives the proposed method a built-in advantage by construction. The paper should include transfer experiments: attack a graph with a UCD configured for detector A, then evaluate DARI and ARI-vs-true-labels under detectors B and C, and also evaluate with a detector different from the one used to obtain the clean-graph labels. This would separate 'the method fools the targeted detector' from 'the method genuinely degrades community structure.'
  3. [V-E.2, Table III] Table III reports average hypervolume and diversity without standard deviations or statistical significance. Several reported differences are tiny (e.g., Karate+FN: 0.5046 for UCD vs 0.5047 for UCD(MIN), and Karate+LPA: 0.8215 vs 0.8212), yet the text claims 'superiority' of the variants. Fig. 8 also lacks error bars. Since both NSGA-II and LPA are stochastic, numbers can vary from run to run. Please report means and standard deviations over at least 10 independent runs and state whether any hypothesis test was used or avoid over-claiming when differences are within noise.
  4. [IV-A, Figs. 3-4] The degree-preserving rewiring and crossover operations are described via examples and asserted to maintain the degree distribution. The paper should provide a precise invariant proof, especially for the crossover case where the two individuals' neighbor sets can overlap, and clarify how self-loops and multi-edges are avoided in all branches. This is load-bearing because 'unnoticeable' in the title is supported by exactly this degree-preservation claim.
minor comments (5)
  1. [Eq. (1)] The sentence 'where 1 indicates the original ARI value of the clean graph before the attacks' is confusing. Since the ground-truth labels are the detector output on the clean graph, the original ARI is identically 1, so DARI = 1 - ARI is not really a 'decrease' but a deviation from perfect agreement. Please rephrase.
  2. [Table II] Table heading has a typo: 'Datesets' should be 'Datasets.'
  3. [References] Reference [21] for NSGA-II is cited as IEEE Trans. Comput. Soc. Syst.; the canonical venue is IEEE Trans. Evol. Comput. Also in reference [25], 'NeruIPS' should be 'NeurIPS.'
  4. [V-E.4 / Fig. 10] The modularity fluctuations during UCD optimization support the claim that modularity drop is not a monotone proxy, but the paper should be careful not to imply that maximizing DARI is 'better than' maximizing modularity drop; the two metrics measure different things. A sentence acknowledging the scope of the claim would help.
  5. [V-D] Hypervolume is computed with reference point (0,0) and both objectives in [0,1]. This is a reasonable choice, but its sensitivity to the reference point should be noted, especially because DAT is normalized by T = 20% of links, which varies across datasets.

Circularity Check

1 steps flagged · score 4.0 of 10

DARI-based evaluation is partly self-referential: 'ground truth' is the attacked detector's own clean-graph output, so the reported deception performance is the objective being optimized.

  1. self definitional [Section III-C (Eq. 1), used in Section IV-A Step 4 and Section V-E]
    "Specifically, we first employ a specific detection method on the clean graph (i.e., before the attacks) and consider the obtained clustering labels as the ground truth labels. Our goal is to design the attack strategy to mislead nodes to be predicted to their corresponding ground truth labels. In particular, to further transform our task into a maximization problem, we define the measure of attack performance as the decrease of ARI, namely DARI, which is given as follows. DARI = 1 − ARI, (1)"

    DARI is defined with the attacked detector's own clean-graph output as the reference 'ground truth'. In Step 4 (Fitness Calculation), the method records DARI as the deception performance for each candidate graph, and the experiments (Fig. 6, Table III) evaluate success by the same DARI. Thus 'deception performance' reduces by construction to 'how much the same detector changed its partition after rewiring'. On real datasets with no external community labels, reporting high DARI is reporting the objective value that NSGA-II was explicitly maximizing, not independent evidence that the true underlying community structure was hidden. The comparison against GAQ/GCH on DARI is not forced, but the central interpretation of high DARI as successful community deception is built into Eq. (1).

full rationale

This paper has no self-citation chain and no fitted parameter that is later renamed as a prediction; the two MOO objectives (DARI and DAT) are defined independently, and the Pareto-front comparison against GAQ/GCH contains real comparative content. However, the evaluation is partly self-referential. The paper defines its deception metric using the output of the targeted detector on the clean graph as 'ground truth' (Section III-C), then optimizes that exact metric (Section IV-A Step 4), and then reports that metric as evidence of effectiveness (Section V-E). On Karate, Dolphins, and Netscience, which have no true community labels used in the evaluation, DARI measures the detector's partition shift, not whether the actual communities were hidden. This is especially problematic for LPA, which is stochastic: the 'ground truth' is one arbitrary run, and different clean-graph runs may yield different reference partitions. The paper does not report variance across runs or validate DARI against any external ground truth on these datasets. The synthetic experiment in Table I gives some independent support for preferring ARI-type measures over modularity, and the degree-preserving rewiring is an independent methodological contribution, so the paper is not wholly circular. But the central empirical claim of 'effective community deception' is substantially supported only by the objective that the method itself maximizes, relative to a self-defined ground truth. This is a partial, definitional circularity rather than a derivation-chain collapse, hence a score of 4.

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

All numerical hyperparameters are hand-set and only T changes the objective normalization. The main free choice is the definition of ground truth as detector output, which is listed as a domain assumption. No invented entities are needed. The central method relies on the feasibility of degree-preserving swaps and on NSGA-II's approximation behavior, both listed as axioms.

free parameters (5)
  • T (maximum perturbation budget) = 20% of original links
    Appears in Eq. (3) as the normalization denominator of DAT; hand-set, and the whole Pareto trade-off is expressed relative to it. Changing T rescales all DAT values and therefore the shape of the Pareto front.
  • Population size (Omega) = 30
    NSGA-II population size, set without sensitivity analysis (Section V-D).
  • Crossover rate p_c = 0.5
    Hand-set (Section V-D).
  • Mutation rate p_m = 0.8
    Hand-set (Section V-D). Mutation rate 0.8 is unusually high, which may explain the reliance on mutation for search; no sensitivity analysis.
  • Max iterations = 500
    Hand-set termination condition (Section V-D); no study of convergence sensitivity.
assumptions (5)
  • domain assumption The output of a community detection algorithm on the clean graph is a valid ground-truth partition for measuring deception (Section III-C, Eq. 1).
    DARI compares attacked detection labels against clean detection labels, so the metric is only meaningful if the clean detector output approximates the true communities. The paper does not justify this for Karate, Dolphins, or Netscience, which lack ground-truth labels in the experiments.
  • domain assumption Degree distribution preservation is sufficient to make perturbations unnoticeable (Section IV-A, Section V-E3).
    The method only preserves each node's degree; other structural statistics (clustering, paths, assortativity) are unconstrained. The paper does not test whether defenders can detect the rewiring.
  • ad hoc to paper The rewiring and crossover operations can always be completed without creating multi-edges or self-loops (Section IV-A, Figs. 3-4).
    No existence proof is given for the required non-edge/edge choices; in dense or small graphs the search for a valid degree-preserving swap may be constrained.
  • standard math NSGA-II non-dominated sorting and crowding distance yield a good approximation of the true Pareto front (Section IV-A, Step 5).
    The paper relies on the well-known NSGA-II behavior, citing Deb et al. [21]; treated as background.
  • standard math ARI is an appropriate external validity metric (Section III-C).
    Standard adjusted Rand index properties assumed from Hubert and Arabie [18].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unnoticeable Community Deception via Multi-objective Optimization." pith.science (2026). https://pith.science/paper/WHD7526G

@misc{pith2026250901438,
  author       = {Pith},
  title        = {Pith review of: Unnoticeable Community Deception via Multi-objective Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WHD7526G}},
  note         = {Machine review of arXiv:2509.01438}
}
read the original abstract

Community detection in graphs is crucial for understanding the organization of nodes into densely connected clusters. While numerous strategies have been developed to identify these clusters, the success of community detection can lead to privacy and information security concerns, as individuals may not want their personal information exposed. To address this, community deception methods have been proposed to reduce the effectiveness of detection algorithms. Nevertheless, several limitations, such as the rationality of evaluation metrics and the unnoticeability of attacks, have been ignored in current deception methods. Therefore, in this work, we first investigate the limitations of the widely used deception metric, i.e., the decrease of modularity, through empirical studies. Then, we propose a new deception metric, and combine this new metric together with the attack budget to model the unnoticeable community deception task as a multi-objective optimization problem. To further improve the deception performance, we propose two variant methods by incorporating the degree-biased and community-biased candidate node selection mechanisms. Extensive experiments on three benchmark datasets demonstrate the superiority of the proposed community deception strategies.

Figures

Figures reproduced from arXiv: 2509.01438 by the authors.

Figure 1
Figure 1. Illustration of the generated graph with a strong community structure [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. By adopting the proposed graph generator, we then explore the relationship between the change of community distribution and modularity, as shown in Table I. We take the graph with 200 nodes (i.e., n=200), whose original community distribution is {100, 50, 25, 15, 10}, as an example. Six different operations have been conducted to mimic the attack operation at an extreme level, including the merge of the two largest … view at source ↗
Figure 2
Figure 2. Visualization of the correlation between the DAT and DARI metrics in three datasets using the LOU detection method based on 2,000 independent [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figures from the paper (9 more)
Figure 3
Figure 3. Figure 3: Illustration of the degree-preserving rewiring operation (and mutation [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 4
Figure 4. Figure 4: Illustration of the crossover process, where solid lines and dashed [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: A systematic framework of UCD-based methods, where the biased mutation represents the corresponding variant methods of the original UCD method [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the Pareto fronts of UCDs and baseline methods under three detection algorithms on three datasets. The red triangles, green squares, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Illustration of the hyper volume with (0, 0) as the reference point. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Comparison of the changes in hyper volume of UCDs with regard to the number of iterations under three detection algorithms on three datasets. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Visualization of the perturbed karate networks generated by corresponding community deception strategies. The green and red nodes denote the nodes [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: We can easily find that, with the increase of iteration [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 10
Figure 10. Figure 10: Comparison of the change of modularity of different solutions of Pareto fronts obtained from UCD (MAX) with regard to the number of iterations [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 57 canonical work pages

  1. [1]

    Information network or social network? The structure of the twitter follow graph,

    S. A. Myers, A. Sharma, P. Gupta, and J. Lin, “Information network or social network? The structure of the twitter follow graph,” in Proc. WWW, 2014, pp. 493–498

    work page 2014

  2. [2]

    On power-law relation- ships of the internet topology,

    M. Faloutsos, P. Faloutsos, and C. Faloutsos, “On power-law relation- ships of the internet topology,” ACM Comput. Commun. Rev. , vol. 29, no. 4, pp. 251–262, 1999

    work page 1999

  3. [3]

    Deeptrax: Embedding graphs of financial transactions,

    A. Khazane, J. Rider, M. Serpe, A. Gogoglou, K. Hines, C. B. Bruss, and R. Serpe, “Deeptrax: Embedding graphs of financial transactions,” in Proc. ICMLA. IEEE, 2019, pp. 126–133

    work page 2019

  4. [4]

    The power grid as a complex network: A survey,

    G. A. Pagani and M. Aiello, “The power grid as a complex network: A survey,” Phys. A Stat. Mech. Appl. , vol. 392, no. 11, pp. 2688–2700, 2013

    work page 2013

  5. [5]

    Node classification in social networks,

    S. Bhagat, G. Cormode, and S. Muthukrishnan, “Node classification in social networks,” Soc. Netw. Data Anal. , pp. 115–148, 2011

    work page 2011

  6. [6]

    Link prediction in complex networks: A survey,

    L. L ¨u and T. Zhou, “Link prediction in complex networks: A survey,” Phys. A Stat. Mech. Appl. , vol. 390, no. 6, pp. 1150–1170, 2011. FANG et al. : UNNOTICEABLE COMMUNITY DECEPTION VIA MULTI-OBJECTIVE OPTIMIZATION 13

    work page 2011

  7. [7]

    A survey on community detection methods in social networks,

    M. Khatoon and W. A. Banu, “A survey on community detection methods in social networks,” Int. J. Educ. Manag. Eng. , vol. 5, no. 1, p. 8, 2015

    work page 2015

  8. [8]

    Fast algorithm for detecting community structure in networks,

    M. E. Newman, “Fast algorithm for detecting community structure in networks,” Phys. Rev. E , vol. 69, no. 6, p. 066133, 2004

    work page 2004

Show all 57 references
  1. [9]

    Modularity and community structure in networks,

    ——, “Modularity and community structure in networks,” Proc. Natl. Acad. Sci. , vol. 103, no. 23, pp. 8577–8582, 2006

    2006

  2. [10]

    Fast unfolding of communities in large networks,

    V . D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfolding of communities in large networks,” J. Stat. Mech. Theor . Exp., vol. 2008, no. 10, p. P10008, 2008

    2008

  3. [11]

    The impact of unlinkability on adversarial community detection: Effects and countermeasures,

    S. Nagaraja, “The impact of unlinkability on adversarial community detection: Effects and countermeasures,” in Proc. PETS. Springer, 2010, pp. 253–272

    2010

  4. [12]

    Community deception or: How to stop fearing community detection algorithms,

    V . Fionda and G. Pirro, “Community deception or: How to stop fearing community detection algorithms,” IEEE Trans. Knowl. Data Eng. , vol. 30, no. 4, pp. 660–673, 2017

    2017

  5. [13]

    Hiding individuals and communities in a social network,

    M. Waniek, T. P. Michalak, M. J. Wooldridge, and T. Rahwan, “Hiding individuals and communities in a social network,” Nat. Hum. Behav. , vol. 2, no. 2, pp. 139–147, 2018

    2018

  6. [14]

    Ga-based q-attack on community detection,

    J. Chen, L. Chen, Y . Chen, M. Zhao, S. Yu, Q. Xuan, and X. Yang, “Ga-based q-attack on community detection,” IEEE Trans. Comput. Soc. Syst., vol. 6, no. 3, pp. 491–503, 2019

    2019

  7. [15]

    Community deception in directed influence networks,

    S. A. Madi and G. Pirr `o, “Community deception in directed influence networks,” Soc. Netw. Anal. Min. , vol. 13, no. 1, p. 122, 2023

    2023

  8. [16]

    Enhancing community deception based on graph autoencoder and genetic algo- rithm,

    X. Wang, J. Li, Y . Guan, J. Yuan, H. Tao, and S. Zhang, “Enhancing community deception based on graph autoencoder and genetic algo- rithm,” in Proc. ICCC. IEEE, 2023, pp. 742–746

    2023

  9. [17]

    Community hiding by link perturbation in social networks,

    X. Chen, Z. Jiang, H. Li, J. Ma, and S. Y . Philip, “Community hiding by link perturbation in social networks,” IEEE Trans. Comput. Soc. Syst. , vol. 8, no. 3, pp. 704–715, 2021

    2021

  10. [18]

    Comparing partitions,

    L. Hubert and P. Arabie, “Comparing partitions,” J. Classif., vol. 2, pp. 193–218, 1985

    1985

  11. [19]

    Multi-objective optimization,

    K. Deb, K. Sindhya, and J. Hakanen, “Multi-objective optimization,” in Decis. Sci. CRC Press, 2016, pp. 161–200

    2016

  12. [20]

    A review of multi-objective optimization: Methods and its applications,

    N. Gunantara, “A review of multi-objective optimization: Methods and its applications,” Cogent Eng., vol. 5, no. 1, p. 1502242, 2018

    2018

  13. [21]

    A fast and elitist multiobjective genetic algorithm: Nsga-ii,

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: Nsga-ii,” IEEE Trans. Comput. Soc. Syst., vol. 6, no. 2, pp. 182–197, 2002

    2002

  14. [22]

    Community structure in social and biological networks,

    M. Girvan and M. E. Newman, “Community structure in social and biological networks,” Proc. Natl. Acad. Sci. , vol. 99, no. 12, pp. 7821– 7826, 2002

    2002

  15. [23]

    On spectral clustering: Analysis and an algorithm,

    A. Ng, M. Jordan, and Y . Weiss, “On spectral clustering: Analysis and an algorithm,” Proc. NeurIPS, vol. 14, 2001

    2001

  16. [24]

    Stochastic blockmodels: First steps,

    P. W. Holland, K. B. Laskey, and S. Leinhardt, “Stochastic blockmodels: First steps,” Soc. Netw., vol. 5, no. 2, pp. 109–137, 1983

    1983

  17. [25]

    Mixed membership stochastic blockmodels,

    E. M. Airoldi, D. Blei, S. Fienberg, and E. Xing, “Mixed membership stochastic blockmodels,” Proc. NeruIPS, vol. 21, 2008

    2008

  18. [26]

    Modularity based community detection with deep learning

    L. Yang, X. Cao, D. He, C. Wang, X. Wang, and W. Zhang, “Modularity based community detection with deep learning.” in Proc. IJCAI, vol. 16, no. 2016, 2016, pp. 2252–2258

    2016

  19. [27]

    Learning community embedding with community detection and node embedding on graphs,

    S. Cavallari, V . W. Zheng, H. Cai, K. C.-C. Chang, and E. Cambria, “Learning community embedding with community detection and node embedding on graphs,” in Proc. SIGKDD, 2017, pp. 377–386

    2017

  20. [28]

    Commdgi: community detection oriented deep graph infomax,

    T. Zhang, Y . Xiong, J. Zhang, Y . Zhang, Y . Jiao, and Y . Zhu, “Commdgi: community detection oriented deep graph infomax,” in Proc. CIKM , 2020, pp. 1843–1852

    2020

  21. [29]

    Near linear time algorithm to detect community structures in large-scale networks,

    U. N. Raghavan, R. Albert, and S. Kumara, “Near linear time algorithm to detect community structures in large-scale networks,” Phys. Rev. E , vol. 76, no. 3, p. 036106, 2007

    2007

  22. [30]

    Finding community structure in networks using the eigenvectors of matrices,

    M. E. Newman, “Finding community structure in networks using the eigenvectors of matrices,” Phys. Rev. E, vol. 74, no. 3, p. 036104, 2006

    2006

  23. [31]

    Community detection based on minimum-cut graph partitioning,

    Y . Wang, H. Huang, C. Feng, and Z. Liu, “Community detection based on minimum-cut graph partitioning,” in Proc. WAIM. Springer, 2015, pp. 57–69

    2015

  24. [32]

    A graph-cut-based approach to commu- nity detection in networks,

    H. Shin, J. Park, and D. Kang, “A graph-cut-based approach to commu- nity detection in networks,” Appl. Sci. , vol. 12, no. 12, p. 6218, 2022

    2022

  25. [33]

    A game theory based approach for community detection in social networks,

    L. Zhou, K. L ¨u, C. Cheng, and H. Chen, “A game theory based approach for community detection in social networks,” in Proc. BNCOD. Springer, 2013, pp. 268–281

    2013

  26. [34]

    Dynamic community detection based on game theory in social networks,

    F. Jiang and J. Xu, “Dynamic community detection based on game theory in social networks,” in Proc. Big Data . IEEE, 2015, pp. 2368– 2373

    2015

  27. [35]

    Community detection in networks: a game-theoretic framework,

    Y . Chen, X. Cao, and K. J. R. Liu, “Community detection in networks: a game-theoretic framework,” EURASIP J. Adv. Signal Process. , vol. 2019, pp. 1–12, 2019

    2019

  28. [36]

    Community detection using time-dependent personalized pagerank,

    H. Avron and L. Horesh, “Community detection using time-dependent personalized pagerank,” in Proc. ICML. PMLR, 2015, pp. 1795–1803

    2015

  29. [37]

    Improving pagerank for local community detection,

    A. Hollocou, T. Bonald, and M. Lelarge, “Improving pagerank for local community detection,” arXiv, 2016

    2016

  30. [38]

    Extended clique percolation method to detect overlapping community structure,

    S. Maity and S. K. Rath, “Extended clique percolation method to detect overlapping community structure,” in Proc. ICACCI. IEEE, 2014, pp. 31–37

    2014

  31. [39]

    Clique percolation method: memory efficient almost exact communi- ties,

    A. Baudin, M. Danisch, S. Kirgizov, C. Magnien, and M. Ghanem, “Clique percolation method: memory efficient almost exact communi- ties,” in Proc. ADMA. Springer, 2022, pp. 113–127

    2022

  32. [40]

    Safeness-based community penetration,

    Z. Yu, Z. Jiang, J. Ma, and Y . Zheng, “Safeness-based community penetration,” IEEE Tran. Circ. Syst. II: Exp. Briefs , vol. 68, no. 7, pp. 2690–2694, 2021

    2021

  33. [41]

    Multiscale evolutionary perturbation attack on community detection,

    J. Chen, Y . Chen, L. Chen, M. Zhao, and Q. Xuan, “Multiscale evolutionary perturbation attack on community detection,” IEEE Trans. Comput. Soc. Syst. , vol. 8, no. 1, pp. 62–75, 2020

    2020

  34. [42]

    Lsha: A local structure-based community detection attack heuristic approach,

    H. Yang, L. Chen, F. Cheng, J. Qiu, and L. Zhang, “Lsha: A local structure-based community detection attack heuristic approach,” IEEE Trans. Comput. Soc. Syst. , 2023

    2023

  35. [43]

    Community hiding: Completely escape from community detection,

    Z. Chang, J. Liang, S. Ma, and D. Liu, “Community hiding: Completely escape from community detection,” Inf. Sci., vol. 672, p. 120665, 2024

    2024

  36. [44]

    Prohico: A probabilistic framework to hide communities in large networks,

    X. Liu, L. Fu, X. Wang, and J. E. Hopcroft, “Prohico: A probabilistic framework to hide communities in large networks,” in Proc. INFOCOM. IEEE, 2021, pp. 1–10

    2021

  37. [45]

    Hiding ourselves from community detection through genetic algorithms,

    D. Liu, Z. Chang, G. Yang, and E. Chen, “Hiding ourselves from community detection through genetic algorithms,” Inf. Sci. , vol. 614, pp. 123–137, 2022

    2022

  38. [46]

    Community hiding using a graph autoencoder,

    ——, “Community hiding using a graph autoencoder,” Knowl.-Based Syst., vol. 253, p. 109495, 2022

    2022

  39. [47]

    Variational graph auto-encoders,

    T. N. Kipf and M. Welling, “Variational graph auto-encoders,” in Proc. NeurIPS, 2016

    2016

  40. [48]

    Obfuscating community structure in complex network with evolutionary divide-and-conquer strategy,

    J. Zhao and K. H. Cheong, “Obfuscating community structure in complex network with evolutionary divide-and-conquer strategy,” IEEE Trans. Evol. Comput. , vol. 27, no. 6, pp. 1926–1940, 2023

    1926

  41. [49]

    A self-adaptive evolution- ary deception framework for community structure,

    J. Zhao, Z. Wang, J. Cao, and K. H. Cheong, “A self-adaptive evolution- ary deception framework for community structure,” IEEE Trans. Syst. Man Cybern. Syst. , vol. 53, no. 8, pp. 4954–4967, 2023

    2023

  42. [50]

    A unified framework of community hiding using symmetric nonnegative matrix factorization,

    D. Liu, R. Jia, X. Liu, and W. Zhang, “A unified framework of community hiding using symmetric nonnegative matrix factorization,” Inf. Sci. , vol. 663, p. 120235, 2024

    2024

  43. [51]

    Adversarial attack on community detection by hiding individuals,

    J. Li, H. Zhang, Z. Han, Y . Rong, H. Cheng, and J. Huang, “Adversarial attack on community detection by hiding individuals,” in Proc. WWW, 2020, pp. 917–927

    2020

  44. [52]

    How to protect ourselves from overlapping community detection in social networks,

    D. Liu, G. Yang, Y . Wang, H. Jin, and E. Chen, “How to protect ourselves from overlapping community detection in social networks,” IEEE Trans. Big Data. , vol. 8, no. 4, pp. 894–904, 2022

    2022

  45. [53]

    Overlapping community hid- ing method based on multi-level neighborhood information,

    G. Yang, Y . Wang, Z. Chang, and D. Liu, “Overlapping community hid- ing method based on multi-level neighborhood information,” Symmetry, vol. 14, no. 11, p. 2328, 2022

    2022

  46. [54]

    The micro-scale adversarial detection algorithm for overlapping communities based on genetic algorithms,

    S. Ji, Y . Liu, X. Guo, and M. Jia, “The micro-scale adversarial detection algorithm for overlapping communities based on genetic algorithms,” in Proc. ICIC. Springer, 2024, pp. 306–315

    2024

  47. [55]

    An information flow model for conflict and fission in small groups,

    W. W. Zachary, “An information flow model for conflict and fission in small groups,” J. Anthropol. Res. , vol. 33, no. 4, pp. 452–473, 1977

    1977

  48. [56]

    The bottlenose dolphin community of doubtful sound features a large proportion of long-lasting associations: can geographic isolation explain this unique trait?

    D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase, E. Slooten, and S. M. Dawson, “The bottlenose dolphin community of doubtful sound features a large proportion of long-lasting associations: can geographic isolation explain this unique trait?” Behav. Ecol. Sociobiol. , vol. 5...

    2003

  49. [57]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proc. ICLR, 2017

    2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.