Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Negative degree corrections let an untrained spectral GNN partition graphs 16x to 23x faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

InfraredGP shows that a negative degree correction in a random-input spectral GNN produces clusterable embeddings, yielding fast, competitive graph partitioning without training.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Real speedup, interesting negative-correction trick, but the 'negative correction alone' claim needs a cleaner ablation because τ, L, d are tuned per dataset. the 4 major comments →

arxiv 2508.19737 v1 pith:E3REEREN submitted 2025-08-27 cs.LG cs.SI

InfraredGP: Efficient Graph Partitioning via Spectral Graph Neural Networks with Negative Corrections

Meng Qin , Weihua Li , Jinqiang Cui , Sen Pei This is my paper
classification cs.LG cs.SI
keywords graph partitioningcommunity detectionspectral graph neural networksnegative degree correctiongraph signal processinginfrared frequenciesBIRCH clusteringHPEC Graph Challenge
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

InfraredGP claims that subtracting a positive constant from each node's degree in the graph Laplacian pushes part of the graph spectrum below its usual [0,2] band, and that this 'infrared' low-frequency band is rich enough in community structure that a randomly initialized, untrained spectral GNN can turn white noise into embeddings a generic clustering routine (BIRCH) partitions correctly. On the IEEE HPEC Graph Challenge benchmark, the paper reports 16x to 23x faster partitioning than the strongest baselines, with F1 and ARI scores staying within about half a percent of the best competitor and sometimes slightly better. If the effect is real, it means high-quality community detection on million-node graphs needs neither training, eigendecomposition, nor a task-specific model-selection pipeline, and the same forward pass extends to streaming graph updates via BIRCH's partial-fit mode.

Core claim

Starting from the normalized Laplacian L = I − D^{-1/2}AD^{-1/2}, the paper replaces D with D_τ = D + τ I and studies τ < 0, clipping entries so D_τ stays positive. With τ < 0, eigenvalues of the corrected Laplacian can fall below zero—the 'infrared' part of the spectrum, in the paper's optical analogy—whereas the standard Laplacian always has eigenvalues in [0,2]. InfraredGP feeds Gaussian noise through L layers of a spectral GNN whose low-pass kernel φ(Λ) = (θ + α)I − αΛ (with α = 1, θ = 0.1) amplifies low frequencies, interleaved with tanh, z-score normalization, and a final sigmoid; because the kernel equals 0.1I + D_τ^{-1/2}AD_τ^{-1/2}, no eigendecomposition is needed, only sparse matri

What carries the argument

The load-bearing object is the negatively corrected degree matrix D_τ := D + τ I for τ < 0, with entries renormalized by Eq. (1) so the matrix stays positive, and the associated Laplacian L_τ := I − D_τ^{-1/2}AD_τ^{-1/2}. Gershgorin's circle theorem is used to show that τ shifts the eigenvalues; negative τ spreads some frequencies below zero, which the paper calls infrared information. The carrying identity is φ(Λ) = (θ + α)I − αΛ with (α, θ) = (1, 0.1), which for the low-pass choice becomes 0.1I + D_τ^{-1/2}AD_τ^{-1/2}, so each GNN layer is just a degree-normalized neighbor average plus a small self-loop, applied to random input and followed by tanh and z-score normalization. BIRCH is the c

Load-bearing premise

The central claim rests on the assumption that, after a negative degree correction, the low-frequency eigenvectors of the corrected Laplacian still align with the true community structure, so amplifying those frequencies turns random noise into clusterable embeddings; the paper demonstrates this on the HPEC generator and one small graph but does not prove it or characterize when it fails.

What would settle it

Construct a stochastic block model graph where the lowest eigenvectors of L_τ are deliberately misaligned with the planted blocks—for example, with strong degree heterogeneity that makes the corrected low-frequency subspace track degree rather than community—then run InfraredGP with τ from the paper's recipe. If BIRCH still recovers the planted partition, the 'infrared amplification' explanation is not the operative mechanism; if it fails, the method's success is bound to that eigenvector alignment, which is exactly the premise the paper assumes.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Million-node graphs can be partitioned in tens of seconds on a single CPU with no training, no eigendecomposition, and no learned model-selection layer, where several SBM and RaftGP baselines time out or run out of memory.
  • The same fixed forward pass works for streaming graphs: partial BIRCH updates on the new nodes' embeddings are faster than recomputing a static partition at every step, and occasionally more accurate.
  • Community-relevant signal is not confined to the conventional [0,2] spectral band; the 'infrared' band below zero can carry it, so other spectral algorithms should treat the assumed frequency range as a design choice rather than a fixed fact.
  • Because the quality gap to the best tuned baseline stays under about 0.5% in F1 and ARI, InfraredGP is a viable fast initial partitioner whose output could be refined by a more expensive method if maximum accuracy is required.
  • At the largest tested scales (500K and 1M nodes), InfraredGP keeps F1 around 99.4%, indicating the efficiency gain does not degrade as the graph grows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • By the same mechanism, any spectral GNN with random initialization might become a training-free unsupervised embedding generator for other tasks—node clustering, anomaly detection, or graph visualization—as long as the negative-correction regime is entered; this is an extrapolation, not a claim the paper tests.
  • The per-dataset values of τ, L, and d in Table II hint that the effect is tuned rather than fully automatic; a natural follow-up is an adaptive rule for τ based on spectral gap or degree heterogeneity that would remove the manual knob.
  • If the 'infrared' subspace is genuinely community-aligned, then the mirror-image 'ultraviolet' eigenvalues above 2 might encode complementary structural information (e.g., hub or role structure), which could be tested with a high-pass version of the same filter.
  • The streaming version occasionally beats rerunning static GP from scratch, which suggests the fresh random noise in each forward pass may act as an implicit regularizer; averaging over multiple noise draws could reveal whether partitions are stable or noise-dependent.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes InfraredGP, a training-free spectral GNN method for K-agnostic graph partitioning. It replaces the standard degree matrix by D_tau = D - min(|tau|, deg_i - epsilon) for tau < 0, so that the effective Laplacian spectrum extends below zero (called 'infrared' frequencies). The filter phi(L_tau) = 0.1 I + D_tau^{-1/2} A D_tau^{-1/2} is applied for L layers to random Gaussian node features, with z-score normalization and tanh/sigmoid nonlinearities, and the resulting embeddings are clustered with BIRCH. On synthetic HPEC Graph Challenge graphs from N=5K to 1M, the method reports 16x-23x speedups over SBM, Louvain, Locale, and RaftGP baselines while keeping F1 within 0.5% of the best baseline. A streaming variant partially updates BIRCH using embeddings of newly added nodes. The central claim is that the negative correction alone makes an untrained, random-feature spectral GNN produce clusterable embeddings.

Significance. If the central claim held, this would be a notable empirical result: a very simple, training-free spectral GNN with a negative degree correction can match the quality of much heavier community-detection baselines at a fraction of the cost on the HPEC benchmark. The paper reports standard deviations, uses the official benchmark generator, and makes code publicly available. The static efficiency/quality results are credible and the method is simple enough to reproduce. However, the evidence does not currently establish that the negative correction is the operative cause: the method's hyperparameters vary per dataset, the spectral mechanism is not derived, and the streaming evaluation lacks external baselines. The 'based solely on the negative correction' claim is therefore not yet supported at the level the abstract asserts.

major comments (4)
  1. [Table II and Section IV-D] The central causal claim is that the negative correction alone accounts for clusterable embeddings. However, Table II changes (tau, L, d) per dataset: tau = -6, -3, -80, -80, -80, -80; L = 10, 9, 40, 60, 70, 60; d = 64, 64, 32, 32, 32, 32, and no model-selection procedure is reported. Fig. 4 fixes N=100K and one (L, d) combination; it does not show that the same negative tau works at other L,d, nor that tau=0 fails at other L,d. BIRCH's own hyperparameters are likewise unspecified. If these settings were selected per dataset with access to ground truth, the 'without any training' claim is misleading. Please report results under a fixed or principled hyperparameter rule across all N, and provide joint (tau, L, d) sensitivity analysis.
  2. [Section III.A / Fig. 1 / Theorem 1] The paper's load-bearing premise is that the low-frequency eigenspace of L_tau for tau<0 remains community-relevant, so that amplifying eigenvalues below zero yields clusterable embeddings. Theorem 1 only constrains eigenvalue locations; it says nothing about the eigenvectors or their alignment with community indicators. The claim that infrared information 'encodes more informative properties about community structures' is asserted rather than derived, and the evidence is limited to one Karate example and the HPEC generator. Please add a theoretical analysis (e.g., perturbation or regularization relation to degree-corrected spectral clustering) or, failing that, evaluate the mechanism on real graphs and across a wider range of SBM parameters to demonstrate that the effect is intrinsic to the negative correction rather than an artifact of the benchmark generator.
  3. [Section IV-C / Figs. 2-3] The streaming experiments compare Algorithm 2 only against Algorithm 1 run from scratch. No MC-SBM, Par-SBM, Louvain, Locale, RaftGP, or other streaming baselines are included. Consequently, the abstract's claim that InfraredGP achieves 'much better efficiency and competitive quality over various baselines' for both static and streaming GP is unsupported for the streaming setting. Please add external baselines for the snowball streaming model, or clearly restrict the claimed comparison to the static setting.
  4. [Algorithm 2 / Section III-C] In each streaming step, a fresh Gaussian matrix Theta is drawn (line 3), yet the previous embeddings Zhat_{t-1} are retained (line 10). The new nodes' embeddings are therefore computed from a different random realization than the old nodes' embeddings, and the combined matrix mixes inconsistent coordinate systems. The paper neither justifies this approximation nor ablates it against fixed-seed or full-recomputation variants. Since the streaming extension is a stated contribution, this issue should be addressed explicitly.
minor comments (5)
  1. [Section II] The notation 'Vtilde_t := S_r=1^t V_t' appears to use a corrupted union symbol; it should be a union over r=1..t. Please fix.
  2. [Algorithm 2] The symbol Zhat_t is used both for the newly added nodes' embeddings (line 9) and for the full cumulative embedding matrix (line 10). Rename one of these to avoid ambiguity.
  3. [Equation (4)] The text says 'standard derivation' in the definition of ZNorm; this should be 'standard deviation.'
  4. [Abstract / Section III.A] The phrase 'based solely on the negative correction mechanism' is too strong given that Section III.A admits z-score normalization and tanh/sigmoid nonlinearities are necessary, and the whole pipeline also requires a chosen number of layers, embedding dimension, and BIRCH. Please qualify the claim.
  5. [Fig. 4] The y-axis appears to show F1 values only up to 0.8, while Table VI reports F1 around 99.4% (0.994) for the same N=100K setting. If F1 is plotted as a percentage, the axis labels should be 20-100; if plotted as a fraction, the axis must extend to 1.0. Please harmonize the figure with the table.

Circularity Check

0 steps flagged

No significant circularity: the core claim is empirically evaluated against external benchmarks, not derived from its own inputs.

full rationale

The paper's central claim is that a negative degree correction (tau < 0) applied to the graph Laplacian produces eigenvalues outside [0,2] and that amplifying these 'infrared' components yields clusterable embeddings for untrained spectral GNNs. This is not circular: the negative correction is defined by Eq. (1), the spectral shift is a mathematical consequence of the modified Laplacian, and the community-relevance of the resulting embeddings is an empirical hypothesis tested on the external HPEC Graph Challenge benchmark against ground-truth communities. The method uses random noise inputs and standard components (tanh, z-score, sigmoid, BIRCH), none of which encode the target partition. The comparison with RaftGP (tau = 0) and the tau ablation in Fig. 4 provide independent empirical support that the negative sign, rather than the architecture alone, is responsible for the improved clusterability. Self-citations to the authors' prior work [1] and to RaftGP [12] are used for motivation, context, or as a baseline; they are not invoked as proof of the new negative-correction mechanism, and no uniqueness theorem or ansatz is imported from those works. Per-dataset hyperparameter choices in Table II are a potential confound for the causal claim, but they are not a fitted quantity renamed as a prediction: the paper does not claim to predict F1 from tau, and no equation reduces the output to the input. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The method depends on six hand-chosen or per-dataset fitted parameters, most importantly tau, L, and d. The core conceptual objects, infrared information and the negative-correction mechanism, are introduced ad hoc and validated only empirically. No new formal axioms beyond standard spectral graph theory are introduced, but the community-preservation property of the corrected spectrum is assumed without proof.

free parameters (6)
  • tau (negative degree correction) = -6, -3, -80, -100 depending on N
    Set per dataset in Table II; central mechanism of the method.
  • L (number of GNN layers) = 9 to 70 depending on N
    Set per dataset in Table II; affects propagation depth.
  • d (embedding dimension) = 16 to 64 depending on N
    Set per dataset in Table II; embedding width.
  • alpha = 1
    Low-pass filter weight chosen by hand; makes Eq. (3) simplify to a degree-normalized adjacency average.
  • theta = 0.1
    Identity-adjacency tradeoff chosen by hand to keep the filter non-degenerate.
  • epsilon = 0.001
    Clipping constant in Eq. (1) chosen by hand to keep D_tau entries positive when tau is negative.
axioms (5)
  • standard math Gershgorin Circle Theorem (Theorem 1) bounds eigenvalues of the corrected Laplacian and justifies that tau shifts the spectrum.
    Invoked in Section I to interpret the effect of tau, but the theorem is not used to prove community structure.
  • standard math Spectral decomposition of the symmetric matrix L_tau is valid, and graph Fourier transform properties hold.
    Used throughout Section III for Eq. (3) and the graph convolution formulation.
  • domain assumption The HPEC stochastic block model generator with heterogeneity 3 and edge-ratio 2.5 produces graphs whose ground-truth blocks match the clustering objective BIRCH recovers.
    All static and streaming experiments use this fixed generator; no real-world graphs are tested.
  • ad hoc to paper The negative correction D_tau = D - min(|tau|, deg-epsilon) with tau<0 produces a spectral basis whose low-frequency components encode communities.
    This is the central heuristic; supported only by one Karate example and the benchmark results, not by a derivation.
  • ad hoc to paper Random Gaussian inputs with z-score normalization and tanh/sigmoid nonlinearities suffice to produce clusterable embeddings.
    The paper states nonlinearities and normalization are necessary (Section III-A), but gives no formal justification.
invented entities (1)
  • Infrared graph information (eigenpairs with lambda < 0) no independent evidence
    purpose: Explains why negative correction helps community detection, by analogy to infrared light beyond the visible spectrum.
    The concept is defined solely by the paper's construction; there is no observable handle outside the method itself.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of InfraredGP: Efficient Graph Partitioning via Spectral Graph Neural Networks with Negative Corrections." pith.science (2026). https://pith.science/paper/E3REEREN

@misc{pith2026250819737,
  author       = {Pith},
  title        = {Pith review of: InfraredGP: Efficient Graph Partitioning via Spectral Graph Neural Networks with Negative Corrections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E3REEREN}},
  note         = {Machine review of arXiv:2508.19737}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Graph partitioning (GP), a.k.a. community detection, is a classic problem that divides nodes of a graph into densely-connected blocks. From a perspective of graph signal processing, we find that graph Laplacian with a negative correction can derive graph frequencies beyond the conventional range $[0, 2]$. To explore whether the low-frequency information beyond this range can encode more informative properties about community structures, we propose InfraredGP. It (\romannumeral1) adopts a spectral GNN as its backbone combined with low-pass filters and a negative correction mechanism, (\romannumeral2) only feeds random inputs to this backbone, (\romannumeral3) derives graph embeddings via one feed-forward propagation (FFP) without any training, and (\romannumeral4) obtains feasible GP results by feeding the derived embeddings to BIRCH. Surprisingly, our experiments demonstrate that based solely on the negative correction mechanism that amplifies low-frequency information beyond $[0, 2]$, InfraredGP can derive distinguishable embeddings for some standard clustering modules (e.g., BIRCH) and obtain high-quality results for GP without any training. Following the IEEE HPEC Graph Challenge benchmark, we evaluate InfraredGP for both static and streaming GP, where InfraredGP can achieve much better efficiency (e.g., 16x-23x faster) and competitive quality over various baselines. We have made our code public at https://github.com/KuroginQin/InfraredGP

Figures

Figures reproduced from arXiv: 2508.19737 by Jinqiang Cui, Meng Qin, Sen Pei, Weihua Li.

Figure 1
Figure 1. Figure 1: An example about (a-c,f) graph spectrum (i.e., ED on graph Laplacian L with τ = 0) and (d,e,g,h) effects of degree correction (i.e., τ > 0 and < 0) on the Karate Club dataset with 34 nodes. Each color denotes a unique block member. our prior study [1] further demonstrated that ED on graph Laplacian Lτ := I − D −1/2 τ AD−1/2 τ (with Dτ := D + τ I), which involves a degree correction term τ > 0, may help de￾… view at source ↗
Figure 4
Figure 4. Figure 4: Parameter analysis of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation results of streaming GP with N = 1M. random noise inputs. Surprisingly, its quality degradation w.r.t. the best competitor can always be controlled within 0.5%. In some cases, InfraredGP can even achieve slightly better quality. Therefore, InfraredGP ensures a significantly better trade-off between the quality and efficiency of static GP. Although RaftGP adopts an architecture similar to In￾frar… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 33 canonical work pages

  1. [1]

    Efficient identity and position graph em- bedding via spectral-based random feature aggregation,

    M. Qin, J. Liu, and I. King, “Efficient identity and position graph em- bedding via spectral-based random feature aggregation,” in Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 2, 2025, pp. 2350–2361

    work page 2025

  2. [2]

    20 years of network community detection,

    S. Fortunato and M. E. Newman, “20 years of network community detection,” Nature Physics, vol. 18, no. 8, pp. 848–850, 2022

    work page 2022

  3. [3]

    Spectral clustering on protein- protein interaction networks via constructing affinity matrix using at- tributed graph embedding,

    K. Berahmand, E. Nasiri, Y . Li et al. , “Spectral clustering on protein- protein interaction networks via constructing affinity matrix using at- tributed graph embedding,” Computers in Biology and Medicine , vol. 138, p. 104933, 2021

    work page 2021

  4. [4]

    Optimal decomposition for large-scale infrastructure- based wireless networks,

    L. Dai and B. Bai, “Optimal decomposition for large-scale infrastructure- based wireless networks,” IEEE Transactions on Wireless Communica- tions, vol. 16, no. 8, pp. 4956–4969, 2017

    work page 2017

  5. [5]

    Graph partitioning models for parallel computing,

    B. Hendrickson and T. G. Kolda, “Graph partitioning models for parallel computing,” Parallel Computing, vol. 26, no. 12, pp. 1519–1534, 2000

    work page 2000

  6. [6]

    Towards a profiling view for unsupervised traffic classification by exploring the statistic features and link patterns,

    M. Qin, K. Lei, B. Bai, and G. Zhang, “Towards a profiling view for unsupervised traffic classification by exploring the statistic features and link patterns,” in Proceedings of the 2019 ACM SIGCOMM Workshop on Network Meets AI & ML , 2019, pp. 50–56

    work page 2019

  7. [7]

    Streaming graph challenge: Stochastic block partition,

    E. Kao, V . Gadepally, M. Hurley, M. Jones, J. Kepner, S. Mohindra, P. Monticciolo, A. Reuther, S. Samsi, W. Song et al., “Streaming graph challenge: Stochastic block partition,” in Proceedings of the 2017 IEEE High Performance Extreme Computing Conference (HPEC) . IEEE, 2017, pp. 1–12

    work page 2017

  8. [8]

    Fast stochastic block partitioning via sampling,

    F. Wanye, V . Gleyzer, and W.-c. Feng, “Fast stochastic block partitioning via sampling,” in Proceedings of 2019 IEEE High Performance Extreme Computing Conference (HPEC) . IEEE, 2019, pp. 1–7

    work page 2019

  9. [9]

    An integrated ap- proach for accelerating stochastic block partitioning,

    F. Wanye, V . Gleyzer, E. Kao, and W.-c. Feng, “An integrated ap- proach for accelerating stochastic block partitioning,” in Proceedings of 2023 IEEE High Performance Extreme Computing Conference (HPEC) . IEEE, 2023, pp. 1–7

    work page 2023

  10. [10]

    Kalman filter driven estimation of com- munity structure in time varying graphs,

    L. J. Durbeck and P. Athanas, “Kalman filter driven estimation of com- munity structure in time varying graphs,” in Proceedings of 2022 IEEE High Performance Extreme Computing Conference (HPEC) . IEEE, 2022, pp. 1–7

    work page 2022

  11. [11]

    Preconditioned spectral clustering for stochastic block partition streaming graph challenge,

    D. Zhuzhunashvili and A. Knyazev, “Preconditioned spectral clustering for stochastic block partition streaming graph challenge,” in Proceed- ings of 2017 IEEE High Performance Extreme Computing Conference (HPEC). IEEE, 2017, pp. 1–6

    work page 2017

  12. [12]

    Raftgp: Random fast graph partitioning,

    Y . Gao, M. Qin, Y . Ding, L. Zeng, C. Zhang, W. Zhang, W. Han, R. Zhao, and B. Bai, “Raftgp: Random fast graph partitioning,” in Proceedings of the 2023 IEEE High Performance Extreme Computing Conference (HPEC). IEEE, 2023, pp. 1–7

    work page 2023

  13. [13]

    Towards faster graph partitioning via pre-training and inductive inference,

    M. Qin, C. Zhang, Y . Gao, Y . Ding, W. Jiang, W. Zhang, W. Han, and B. Bai, “Towards faster graph partitioning via pre-training and inductive inference,” in Proceedings of 2024 IEEE High Performance Extreme Computing Conference (HPEC) . IEEE, 2024, pp. 1–7

    work page 2024

  14. [14]

    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,” Journal of Anthropological Research , vol. 33, no. 4, pp. 452–473, 1977

    work page 1977

  15. [15]

    The emerging field of signal processing on graphs: Ex- tending high-dimensional data analysis to networks and other irregular domains,

    D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Van- dergheynst, “The emerging field of signal processing on graphs: Ex- tending high-dimensional data analysis to networks and other irregular domains,” IEEE Signal Processing Magazine , vol. 30, no. 3, pp. 83–98, 2013

    work page 2013

  16. [16]

    A tutorial on spectral clustering,

    U. V on Luxburg, “A tutorial on spectral clustering,” Statistics & com- puting, vol. 17, pp. 395–416, 2007

    work page 2007

  17. [17]

    Pacer: Network embedding from positional to structural,

    Y . Yan, Y . Hu, Q. Zhou, L. Liu, Z. Zeng, Y . Chen, M. Pan, H. Chen, M. Das, and H. Tong, “Pacer: Network embedding from positional to structural,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 2485–2496

    work page 2024

  18. [18]

    Irwe: Inductive random walk for joint inference of identity and position network embedding,

    M. Qin and D.-Y . Yeung, “Irwe: Inductive random walk for joint inference of identity and position network embedding,” Transactions on Machine Learning Research (TMLR) , 2024

    work page 2024

  19. [19]

    Gershgorin disks for multiple eigenvalues of non-negative matrices,

    I. B ´ar´any and J. Solymosi, “Gershgorin disks for multiple eigenvalues of non-negative matrices,” A Journey Through Discrete Mathematics: A Tribute to Ji ˇr´ı Matou ˇsek, pp. 123–133, 2017

    work page 2017

  20. [20]

    Birch: an efficient data clustering method for very large databases,

    T. Zhang, R. Ramakrishnan, and M. Livny, “Birch: an efficient data clustering method for very large databases,” ACM SIGMOD Record , vol. 25, no. 2, pp. 103–114, 1996

    work page 1996

  21. [21]

    Beyond low-frequency infor- mation in graph convolutional networks,

    D. Bo, X. Wang, C. Shi, and H. Shen, “Beyond low-frequency infor- mation in graph convolutional networks,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 35, no. 5, 2021, pp. 3950– 3957

    work page 2021

  22. [22]

    Adagnn: Graph neural networks with adaptive frequency response filter,

    Y . Dong, K. Ding, B. Jalaian, S. Ji, and J. Li, “Adagnn: Graph neural networks with adaptive frequency response filter,” in Proceedings of the 30th ACM International Conference on Information & Knowledge Management (CIKM), 2021, pp. 392–401

    work page 2021

  23. [23]

    Towards a better tradeoff between quality and efficiency of community detection: an inductive embedding method across graphs,

    M. Qin, C. Zhang, B. Bai, G. Zhang, and D.-Y . Yeung, “Towards a better tradeoff between quality and efficiency of community detection: an inductive embedding method across graphs,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 17, no. 9, pp. 1–34, 2023

    work page 2023

  24. [24]

    Efficient monte carlo and greedy heuristic for the inference of stochastic block models,

    T. P. Peixoto, “Efficient monte carlo and greedy heuristic for the inference of stochastic block models,” Physical Review E, vol. 89, no. 1, p. 012804, 2014

    work page 2014

  25. [25]

    A scalable community detection algorithm for large graphs using stochastic block models,

    C. Peng, Z. Zhang, K.-C. Wong, X. Zhang, and D. Keyes, “A scalable community detection algorithm for large graphs using stochastic block models,” in Proceedings of the 24th International Joint Conference on Artificial Intelligence (IJCAI) , 2015

    work page 2015

  26. [26]

    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,” Journal of Statistical Mechanics: Theory & Experiment , vol. 2008, no. 10, p. P10008, 2008

    work page 2008

  27. [27]

    Community detection using fast low- cardinality semidefinite programming,

    P.-W. Wang and J. Z. Kolter, “Community detection using fast low- cardinality semidefinite programming,” Advances in Neural Information Processing Systems (NIPS) , vol. 33, pp. 3374–3385, 2020

    work page 2020

  28. [28]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proceedings of the 5th International Con- ference on Learning Representations (ICLR) , 2017, p. 1609.02907

    Pith/arXiv arXiv 2017

  29. [29]

    Adaptive community detection incorporating topology and content in social net- works,

    M. Qin, D. Jin, K. Lei, B. Gabrys, and K. Musial-Gabrys, “Adaptive community detection incorporating topology and content in social net- works,” Knowledge-Based Systems, vol. 161, pp. 342–356, 2018

    work page 2018

  30. [30]

    Community detection in node-attributed social networks: How structure-attributes correlation affects clustering quality,

    P. Chunaev, T. Gradov, and K. Bochenina, “Community detection in node-attributed social networks: How structure-attributes correlation affects clustering quality,” Procedia Computer Science , vol. 178, pp. 355–364, 2020

    work page 2020

  31. [31]

    Dual-channel hybrid community detection in attributed networks,

    M. Qin and K. Lei, “Dual-channel hybrid community detection in attributed networks,” Information Sciences, vol. 551, pp. 146–167, 2021

    work page 2021

  32. [32]

    The trade-off between topology and content in community detection: An adaptive encoder–decoder-based nmf approach,

    Z. Zhao, Z. Ke, Z. Gou, H. Guo, K. Jiang, and R. Zhang, “The trade-off between topology and content in community detection: An adaptive encoder–decoder-based nmf approach,” Expert Systems with Applications, vol. 209, p. 118230, 2022

    work page 2022

  33. [33]

    High-quality temporal link prediction for weighted dynamic graphs via inductive embedding aggregation,

    M. Qin, C. Zhang, B. Bai, G. Zhang, and D.-Y . Yeung, “High-quality temporal link prediction for weighted dynamic graphs via inductive embedding aggregation,” IEEE Transactions on Knowledge and Data Engineering (TKDE), vol. 35, no. 9, pp. 9378–9393, 2023

    work page 2023

  34. [34]

    Temporal link prediction: A unified frame- work, taxonomy, and review,

    M. Qin and D.-Y . Yeung, “Temporal link prediction: A unified frame- work, taxonomy, and review,”ACM Computing Surveys (CSUR), vol. 56, no. 4, pp. 1–40, 2023

    work page 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.