Pith. sign in

REVIEW 3 major objections 4 minor 54 references

SNAP-tFDP: Massively Scalable Graph Layouts via Sparse Negative Sampling

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that k random negative samples per edge reproduce the full degree-weighted force objective in expectation, making force-directed layout run in O(|E|) time and lay out 4-million-node graphs in seconds while sharpening cluste

desk verdict Practical, fast graph layout with a clean sampling idea, but the central derivation has a sign/antisymmetry error that must be fixed. read the letter →

arxiv 2608.01907 v1 pith:UWPES2VT submitted 2026-08-03 cs.GR

Xin Chen , Shuowei Hou , Yifan Wang , Mingliang Xue , Zezheng Feng , Oliver Deussen , Weidong Huang , Yunhai Wang This is my paper
classification cs.GR MSC 05C8568R10
keywords GraphLayoutNetworkVisualizationNegativeSamplingt-distributionforcesdegreeweightingclusterseparationGPUparallelizationO(|E|)algorithm
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 sets out to show that a graph-layout algorithm can be both faster and visually better by doing away with all-pairs repulsion. Its recipe is edge-centric negative sampling: for every edge, attract the two endpoints, then push each endpoint away from k randomly drawn other nodes. With a linearly normalized degree weight, (d_i+d_j)/2|V|, and the bounded short-range t-distribution force, the expected repulsion over an epoch equals the full degree-weighted repulsion, so the cheap stochastic loop optimizes almost the same objective as the expensive full model. If this is right, graph layouts no longer need spatial trees or multilevel structures, and graphs with millions of nodes can be drawn in seconds on a single GPU while preserving and sharpening community structure.

What carries the argument

Edge-centric negative sampling with immediate SGD-style updates. The algorithm keeps only an edge list and current positions; after shuffling edges, each directed edge triggers attraction on both endpoints and k uniform negative-sample repulsions on the source. The load-bearing identity is the expectation above, which converts the k samples into a degree-weighted pair-wise repulsion: node i accumulates repulsion from every other node j with weight proportional to d_i+d_j. This is what lets a constant-cost per-edge sample stand in for all-pairs repulsion. A second, supporting mechanism is bundle-based lock-free parallelization: grouping updates by source node and assigning each bundle to one

What would settle it

Two checks settle the central claim. First, accumulate the empirically realized repulsive forces during one full epoch on a graph and compare them with the closed-form expectation (k/(|V|-1))*sum_{j != i}(d_i+d_j)F_r(i,j); a large systematic gap would refute the unbiasedness claim. Second, rerun the method with omega_ij replaced by the quadratic-normalized weight (d_i+1)(d_j+1)/|V|^2 or by no weight at all, keeping all other settings fixed; if neighborhood preservation, silhouette index, or clustering quality does not consistently drop, the load-bearing degree-weight choice is falsified. The p

Watch

Extended reading notes

Core claim

The central claim is that the implicit degree weighting created by negative sampling is not a side effect but the main tool. For each directed edge (i,j), the method first applies the t-FDP attractive force and then samples k nodes uniformly from the other |V|-1 nodes, applying the t-FDP repulsive force between i and each sampled node. Linearity of expectation yields E[F_r^SNAP(i)] = k/(|V|-1) * sum_{j != i}(d_i+d_j)F_r(i,j), so after one epoch the sampled repulsive forces coincide, up to an overall factor, with the degree-weighted t-FDP objective whose pairwise weight is (d_i+d_j)/(2(|V|-1)). Thus the stochastic procedure is asymptotically unbiased with respect to a global objective that wo

Load-bearing premise

The load-bearing premise is the empirical choice of the linearly normalized degree weight omega_ij=(d_i+d_j)/2|V|: Section 3.2 selects it by visual comparison and defers the supporting ablation to the supplemental material, and if this weight is not actually better than alternatives on a range of graphs, the algorithm remains fast but its claimed cluster-separation and quality advantages weaken.

Editorial extensions

If this is right

  • SNAP-tFDP lays out the com-lj network (4 million nodes, 34 million edges) in 9.3 seconds on a GPU with 0.82 GB memory, so million-node layouts no longer require large clusters or CPU farms.
  • On com-friendster (65.6 million nodes, 1.8 billion edges), the same method finishes in about 1.2 hours on a 16-thread CPU with 29.48 GB RAM, bringing ultra-large graphs into single-workstation reach.
  • Because the sampled repulsion is an unbiased estimator of the degree-weighted t-FDP objective, the method's quality claims are directly tied to the choice of weight; the paper reports the highest average NP, SI, and CQ scores among the compared methods.
  • The lock-free bundle parallelization reaches a 16.0x speedup at 64 threads on the largest tested graph, meaning the linear-time algorithm can also use multicore hardware without synchronization overhead.
  • Memory is O(|V|+|E|) with no auxiliary spatial structure, reducing consumption by roughly 72 percent relative to the most memory-efficient baselines the paper compares against.

Reading between the lines

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

  • The expectation argument treats only uniform negative sampling; the same derivation would go through for any sampling distribution with known inclusion probabilities, suggesting a family of degree-weighted or structure-aware repulsion estimators beyond the paper's uniform choice.
  • On highly skewed degree distributions, the linear (d_i+d_j) weight may concentrate too much repulsion on hubs; replacing d_i with a concave transform such as log(1+d_i) is a direct variant the paper does not test, and would be a cheap ablation.
  • The identity shows SNAP-tFDP is effectively optimizing a contrastive objective, so the layout could inherit theoretical guarantees or design principles from contrastive learning, such as temperature schedules or hard-negative mining, in future work.
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 / 4 minor

Summary. The paper presents SNAP-tFDP, a stochastic negative-sampling force-directed layout method for large undirected graphs. It combines t-distribution forces with a linearly normalized degree-dependent repulsion weight omega_ij=(d_i+d_j)/2|V|. The main algorithmic claim is that an edge-centric negative-sampling scheme (k random negatives per directed edge, with immediate SGD-style updates) has expected repulsive force equal to k times the full degree-weighted repulsive force, so that the O(|E|)-time, O(|V|+|E|)-space algorithm optimizes essentially the same objective as the full O(|V|^2) t-FDP model. The paper also proposes a lock-free bundle-based parallelization, reports CPU/GPU implementations, and evaluates layout quality (NP, SI, CQ), runtime, and memory on 12 datasets up to 4M nodes/34M edges, plus a Friendster case study.

Significance. If the equivalence and evaluation hold, this is a practically valuable contribution: it provides a simple, memory-lean, GPU-friendly alternative to tree/grid-based FDP approximations, with large-scale cluster-oriented layout quality. The theoretical derivation is the kind of clean result that would make the paper more than an engineering report. The reproducibility elements (open-source library, 12 datasets, serial/parallel/GPU variants, memory table) are strengths. However, the central derivation currently contains a sign inconsistency, and the degree-weight choice on which all quality benefits depend is not fully evidenced in the main text.

major comments (3)
  1. [Section 3.3] The derivation of Eq. 6 is internally inconsistent. In the displayed definition of F_r^SNAP(i), the second term is written with a plus sign, + X_{jp}Y_{jp,i}F_r(j,i). But Algorithm 1 performs y_s <- y_s - eta F_r(i,s) for a negative sample s; when node i is sampled while processing directed edge (j,p), its update is therefore -F_r(j,i), not +F_r(j,i). The text then invokes 'the symmetry of the repulsive force (F_r(i,j)=F_r(j,i))'. The t-force in Table 1 is antisymmetric (F_r(i,j)=-F_r(j,i)), since e_{ij}=-e_{ji}. With the printed signs the displayed algebra does not produce the coefficient d_i+d_j. Replacing the plus sign by a minus sign and using antisymmetry does recover Eq. 6, so the final result is salvageable, but the proof as printed is incorrect and must be corrected.
  2. [Section 3.2 / Eq. (5)] The linearly normalized degree weight omega_ij=(d_i+d_j)/2|V| is load-bearing: SNAP-tFDP exactly samples a degree-weighted objective, so all cluster-separation advantages over unweighted t-FDP hinge on this choice. The main text justifies it by force-curve inspection (Figure 3) and two example layouts (Figure 2f). The systematic ablation and comparison of alternative normalization schemes are explicitly deferred to the supplemental material. Because this is the central design decision, the main text or the review materials should report the quantitative comparison across the 12 datasets; otherwise the claim that this weighting is 'key' is not established.
  3. [Section 4.1 / Figure 7] The evaluation protocol filters SNAP community datasets to the top 5,000 communities for all visual quality metrics. Since the method is specifically designed to improve separation of clearly defined clusters, this filtering could favor the proposed method and should be justified or shown not to affect conclusions. In addition, Figure 7 reports only average scores for a 5-run stochastic procedure without error bars or per-run spread, making it difficult to assess whether SNAP-tFDP's advantages are significant. Please add variability information and, if possible, a robustness check for the filtering choice.
minor comments (4)
  1. [Introduction / Section 4.4] The phrase '150% speedup' is ambiguous. Given the later statement that PMDS is '150% slower', the intended speedup factor should be stated explicitly (e.g., 2.5x or as a percentage slowdown).
  2. [Figure 7] The caption says 'nine methods' though the heatmap includes more columns; this inconsistent labeling should be fixed.
  3. [Table 3] Table 3 appears to have column-alignment issues, particularly for the GPU memory columns; the MRR row should clearly indicate which baselines are included.
  4. [Algorithm 1] Algorithm 1 updates node positions immediately inside the edge loop. The paper notes the directed-edge duplication in Section 3.3, but a brief comment in the algorithm caption would help avoid ambiguity about whether the input edge list E is directed or undirected.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the negative-sampling expectation identity is derived from Algorithm 1, and the degree weight is an empirical design choice rather than a fitted target.

full rationale

The central equivalence claim (Eq. 6 vs. Eq. 4/5) is not circular. Section 3.3 defines the edge-centric sampler in Algorithm 1, computes E[Y_ij,s]=k/(|V|-1), and decomposes the expected repulsive force on node i into two roles (as source drawing negatives, and as a negative sample for other nodes' edges). The resulting coefficient k(d_i+d_j)/(|V|-1) is obtained by linearity of expectation from the algorithm's update rules, not assumed from Eq. 5. Equation 5 is chosen earlier in Section 3.2 by comparing force-balance curves for alternative degree weights; the fact that the derived sampler coefficient approximately matches Eq. 5 is a consequence of that design, not a self-definition. The energy validation in Figure 4 checks that the stochastic process tracks the designed objective, which is an unbiasedness check rather than a circular prediction. Parameters k and T are tuned empirically, but the reported NP/SI/CQ metrics are measured outputs of the full layout, not quantities fitted by the same parameters. The t-FDP force model and parameters are inherited from reference [47], which shares authors with this paper; however, t-FDP is a published, independently evaluated baseline, and the present derivation does not depend on the validity of [47]'s parameter choices to compute the expectation identity. No uniqueness theorem or ansatz is smuggled in via self-citation. One non-circular caveat: the printed derivation in Section 3.3 invokes 'symmetry of the repulsive force (F_r(i,j)=F_r(j,i))' and writes a plus sign for the negative-sample term, which is inconsistent with the antisymmetric vector convention in Table 1. This is a correctness flaw in the displayed proof (the intended identity is recoverable with the correct convention), not a circularity, and it does not raise the circularity score.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its free parameters are the sampling count k, epoch count T, learning rate scheme, and the inherited t-FDP force constants. The key domain assumptions are the unbiasedness of uniform negative sampling, the empirical optimality of the linear degree weight, and the convergence of lock-free updates.

free parameters (4)
  • k (number of negative samples) = 3
    Selected as default based on trade-off between visual quality and runtime on the APH dataset (Figure 5); controls repulsive strength and runtime.
  • T (number of epochs) = 50
    Chosen based on silhouette-index convergence on the com-lj graph (Figure 6); quality saturates around 50 epochs.
  • learning rate eta
    No explicit value or schedule is given; described as standard practice with decay over epochs. Affects convergence and final layout quality.
  • t-FDP force parameters alpha, beta, gamma = alpha=0.1, beta=8, gamma=2
    Inherited from t-FDP [47]; not tuned here, but the central quality claims depend on them.
assumptions (4)
  • domain assumption Uniform negative sampling yields unbiased gradient estimates with respect to the degree-weighted objective
    Used in Section 3.3 expectation analysis; validated on only one small dataset (Figure 4), not across the full test suite.
  • ad hoc to paper Linearly normalized degree weight (d_i+d_j)/(2|V|) is the right balance between compactness and readability
    Chosen by empirical comparison in Section 3.2; ablation deferred to supplemental material; no derivation of optimality.
  • domain assumption HOGWILD!-style sparsity conditions imply convergence of the lock-free bundle parallelization
    Section 3.4 approximates Delta and rho and asserts sparsity, but convergence is shown empirically, not proven for skewed degree distributions.
  • domain assumption t-FDP parameters from [47] are appropriate defaults
    Inherited directly from prior work without re-validation on the 12 evaluation datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SNAP-tFDP: Massively Scalable Graph Layouts via Sparse Negative Sampling." pith.science (2026). https://pith.science/paper/UWPES2VT

@misc{pith2026260801907,
  author       = {Pith},
  title        = {Pith review of: SNAP-tFDP: Massively Scalable Graph Layouts via Sparse Negative Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWPES2VT}},
  note         = {Machine review of arXiv:2608.01907}
}
abstract

Force-Directed Placement (FDP) is a widely used approach for network visualization, yet scaling it to massive graphs while preserving clear community structures remains a major computational and visual challenge. Existing approximation methods often rely on auxiliary data structures (e.g., spatial trees), which introduce substantial memory overhead; furthermore, traditional power-function-based forces frequently fail to separate dense clusters effectively. In this paper, we present a negative sampling-based algorithm that achieves O(|E|) time complexity with a low memory footprint, without requiring complex multi-level representations. In a first step, we introduce a linearly normalized degree-weighting scheme, which, combined with short-range bounded $t$-distribution forces, effectively untangles dense structures and enhances visual cluster separation. To optimize for this formulation efficiently, we introduce an edge-centric negative sampling strategy that naturally reconstructs the global degree-weighted objective. Furthermore, we design a lock-free, bundle-based parallelization scheme that leverages the sparsity of stochastic updates to achieve significant speedups while mitigating access conflicts. Comprehensive evaluations on 12 large-scale graphs demonstrate that the proposed method outperforms state-of-the-art algorithms in neighborhood preservation and cluster separation. Compared to existing baselines, our method reduces memory consumption by 72% on average and leverages simple GPU parallelism to generate a high-quality layout for a graph with 4 million nodes and 34 million edges in below 10 seconds.

Figures

Figures reproduced from arXiv: 2608.01907 by the authors.

Figure 1
Figure 1. Iterative layouts of communities with more than 800 users in the Orkut online social network [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of classical spring-electrical models and t-FDP models with different weighting schemes. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of different degree-weighting schemes for (a) two [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Validation of the edge-centric negative sampling strategy. (a) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Effect of the number of negative samples [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: (a) Convergence of the silhouette index ( [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Heatmaps employing a pink-to-green colormap illustrate the scores of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Layouts and corresponding runtimes of twelve serial methods for [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Runtime of eight layout methods, together with the GPU imple [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 5 canonical work pages

  1. [1]

    Barnes and P

    J. Barnes and P. Hut. A hierarchical o (n log n) force-calculation algorithm. Nature, 324(6096):446–449, 1986. doi: 10.1038/324446a0 2, 3, 5

    doi:10.1038/324446a0 1986

  2. [2]

    Batson, D

    J. Batson, D. A. Spielman, N. Srivastava, and S.-H. Teng. Spectral sparsi- fication of graphs: theory and algorithms.Communications of the ACM, 56(8):87–94, 2013. doi: 10.1145/2492007.2492029 3

    arXiv 2013

  3. [3]

    J. N. Böhm, P. Berens, and D. Kobak. Attraction-repulsion spectrum in neighbor embeddings.Journal of Machine Learning Research, 23(95):1– 32, 2022. 2

    2022

  4. [4]

    C. Both, N. Dehmamy, R. Yu, and A.-L. Barabási. Accelerating network layouts using graph neural networks.Nature Communications, 14(1):1560, Mar. 2023. doi: 10.1038/s41467-023-37189-2 9

    work page doi:10.1038/s41467-023-37189-2 2023

  5. [5]

    Brandes and C

    U. Brandes and C. Pich. Eigensolver methods for progressive multidi- mensional scaling of large data. InProceedings of the 14th International Symposium on Graph Drawing, pp. 42–53. Springer, 2006. doi: 10.1007/ 978-3-540-70904-6_6 1, 6

    2006

  6. [6]

    D. M. Chan, R. Rao, F. Huang, and J. F. Canny. T-SNE-CUDA: GPU- accelerated t-SNE and its applications to modern data. InProceedings of the 30th International Symposium on Computer Architecture and High Performance Computing, pp. 330–338. IEEE, Campinas, Brazil, 2018. doi: 10.1109/CAHPC.2018.8645912 6

    arXiv 2018

  7. [7]

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. InProceedings of the 37th International Conference on Machine Learning (ICML), pp. 1597–1607. PMLR, 2020. 3

    2020

  8. [8]

    Chimani, C

    M. Chimani, C. Gutwenger, M. Jünger, G. W. Klau, K. Klein, and P. Mutzel. The open graph drawing framework (OGDF). In R. Tamas- sia, ed.,Handbook of Graph Drawing and Visualization, pp. 543–569. Chapman and Hall/CRC, Boca Raton, FL, 2013. 6

    2013

Show all 54 references
  1. [9]

    Damrich and F

    S. Damrich and F. A. Hamprecht. On umap’s true loss function. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, eds.,Advances in Neural Information Processing Systems, vol. 34, pp. 5798–5809. Curran Associates, Inc., 2021. 5

    2021

  2. [10]

    T. A. Davis and Y . Hu. The university of florida sparse matrix collection. ACM Transactions on Mathematical Software, 38(1), art. no. 1, 25 pages, Dec. 2011. doi: 10.1145/2049662.2049663 6

    2011

  3. [11]

    Di Bartolomeo, T

    S. Di Bartolomeo, T. Crnovrsanin, D. Saffo, E. Puerta, C. Wilson, and C. Dunne. Evaluating graph layout algorithms: A systematic review of methods and best practices.Computer Graphics Forum, 43(6):e15073,

  4. [12]

    P. Eades. A heuristic for graph drawing.Congressus numerantium, 42(11):149–160, 1984. 1

    1984

  5. [13]

    Ellson, E

    J. Ellson, E. Gansner, L. Koutsofios, S. C. North, and G. Woodhull. Graphviz–open source graph drawing tools. InProceedings of the 9th International Symposium on Graph Drawing, pp. 483–484. Springer, 2001. 6

    2001

  6. [14]

    Fender, B

    A. Fender, B. Rees, and J. Eaton. Rapids cugraph. InMassive Graph Analytics, pp. 483–493. Chapman and Hall/CRC, 2022. 6

    2022

  7. [15]

    T. M. Fruchterman and E. M. Reingold. Graph drawing by force-directed placement.Software: Practice and experience, 21(11):1129–1164, 1991. doi: 10.1002/spe.4380211102 1, 2, 3, 4, 6

    1991 doi

  8. [16]

    R. Gove. Force-directed graph layouts by edge sampling. InProceedings of the IEEE Symposium on Large Data Analysis and Visualization (LDAV), pp. 1–5. IEEE, 2019. doi: 10.1109/LDA V48142.2019.8944364 3

    2019

  9. [17]

    R. Gove. A random sampling o(n) force-calculation algorithm for graph layouts.Computer Graphics Forum, 38(3):739–751, 2019. doi: 10.1111/ cgf.13724 3, 5

    2019

  10. [18]

    Hangan, S

    D. Hangan, S. Kobourov, and J. Miller. Bridging graph drawing and di- mensionality reduction with stochastic stress optimization.arXiv preprint arXiv:2605.00641, 2026. 2

    2026 arXiv

  11. [19]

    Y . F. Hu. Efficient and high quality force-directed graph drawing.The Mathematica Journal, 10:37–71, 2005. 3, 6

    2005

  12. [20]

    Hubert and P

    L. Hubert and P. Arabie. Comparing partitions.Journal of Classification, 2(1):193–218, 1985. doi: 10.1007/BF01908075 7

    1985 doi

  13. [21]

    Jacomy, T

    M. Jacomy, T. Venturini, S. Heymann, and M. Bastian. Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software.PloS one, 9(6):e98679, 2014. doi: 10. 1371/journal.pone.0098679 1, 2, 3, 4, 6

    2014

  14. [22]

    Kamada and S

    T. Kamada and S. Kawai. An algorithm for drawing general undirected graphs.Information Processing Letters, 31(1):7–15, 1989. doi: 10.1016/ 0020-0190(89)90102-6 1, 2

    1989

  15. [23]

    J. F. Kruiger, P. E. Rauber, R. M. Martins, A. Kerren, S. Kobourov, and A. C. Telea. Graph layouts by t-sne.Computer Graphics Forum, 36(3):283– 294, 2017. doi: 10.1111/cgf.13187 2, 4, 6

    2017 doi

  16. [24]

    B. Lee, C. Plaisant, C. S. Parr, J.-D. Fekete, and N. Henry. Task taxonomy for graph visualization. InProceedings of the AVI Workshop on BEyond Time and Errors: Novel Evaluation Methods for Information Visualization (BELIV), pp. 1–5, 2006. doi: 10.1145/1168149.1168168 1

    2006

  17. [25]

    Leskovec and A

    J. Leskovec and A. Krevl. SNAP Datasets: Stanford large network dataset collection.http://snap.stanford.edu/data, June 2014. 6, 9

    2014

  18. [26]

    G. C. Linderman, M. Rachh, J. G. Hoskins, S. Steinerberger, and Y . Kluger. Fast interpolation-based t-sne for improved visualization of single-cell rna-seq data.Nature Methods, 16(3):243–245, 2019. doi: 10.1038/s41592 -018-0308-4 6

    2019 doi

  19. [27]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville. Umap: Uniform manifold approxi- mation and projection for dimension reduction, 2020. doi: 10.48550/arXiv .1802.03426 2, 3

    2020 doi arXiv

  20. [28]

    Meidiana, S.-H

    A. Meidiana, S.-H. Hong, P. Eades, and D. Keim. A quality metric for visualization of clusters in graphs. InProceedings of the 27th International Symposium on Graph Drawing and Network Visualization (GD), pp. 125–

  21. [29]

    A. Noack. Energy models for graph clustering.Journal of Graph Algo- rithms and Applications, 11(2):453–480, 2007. doi: 10.7155/jgaa.00154 1, 2, 3, 4, 6

    2007 doi

  22. [30]

    Onnela, J

    J.-P. Onnela, J. Saramäki, J. Hyvönen, G. Szabó, D. Lazer, K. Kaski et al. Structure and tie strengths in mobile communication networks. Proceedings of the National Academy of Sciences, 104(18):7332–7336,

  23. [31]

    Y . Onoue. Graph drawing stress model with resistance distances.IEEE Transactions on Visualization and Computer Graphics, pp. 1–10, 2026. doi: 10.1109/TVCG.2026.3694437 2, 6

    2026

  24. [32]

    A. v. d. Oord, Y . Li, and O. Vinyals. Representation learning with con- trastive predictive coding.arXiv preprint arXiv:1807.03748, 2018. doi: 10.48550/arXiv.1807.03748 3

    2018 doi arXiv

  25. [33]

    F. V . Paulovich, A. Arleo, and S. van den Elzen. When dimensionality re- duction meets graph (drawing) theory: Introducing a common framework, challenges and opportunities.Computer Graphics Forum, 44(3):e70105,

  26. [34]

    M. K. Rahman, M. H. Sujon, and A. Azad. Batchlayout: A batch-parallel force-directed graph layout algorithm in shared memory. InProceedings of the 2020 IEEE Pacific Visualization Symposium (PacificVis), pp. 16–25. IEEE, 2020. doi: 10.1109/PacificVis48177.2020.3756 3, 6

    2020

  27. [35]

    M. K. Rahman, M. H. Sujon, and A. Azad. Force2vec: Parallel force- directed graph embedding. InProceedings of the IEEE International Conference on Data Mining (ICDM), pp. 442–451. IEEE, 2020. doi: 10. 1109/ICDM50108.2020.00053 2, 3, 6

    2020

  28. [36]

    Recht, C

    B. Recht, C. Re, S. Wright, and F. Niu. HOGWILD!: A lock-free ap- proach to parallelizing stochastic gradient descent. In J. Shawe-Taylor, R. Zemel, P. Bartlett, F. Pereira, and K. Weinberger, eds.,Advances in Neural Information Processing Systems, vol. 24. Curran Associates, Inc.,

  29. [37]

    P. J. Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis.Journal of Computational and Applied Mathematics, 20:53–65, 1987. doi: 10.1016/0377-0427(87)90125-7 2, 7

    1987 doi

  30. [38]

    Saket, P

    B. Saket, P. Simonetto, and S. Kobourov. Group-level graph visualization taxonomy, 2014. doi: 10.48550/arXiv.1403.7421 4

    2014 doi arXiv

  31. [39]

    Van Der Maaten

    L. Van Der Maaten. Accelerating t-sne using tree-based algorithms.The journal of machine learning research, 15(1):3221–3245, Jan. 2014. 2

    2014

  32. [40]

    van der Maaten and G

    L. van der Maaten and G. Hinton. Visualizing data using t-SNE.Journal of Machine Learning Research, 9(86):2579–2605, 2008. 7

    2008

  33. [41]

    C. Walshaw. A multilevel algorithm for force-directed graph drawing. In Proceedings of the 8th International Symposium on Graph Drawing, pp. 171–182. Springer, 2000. doi: 10.1007/3-540-44541-2_17 3

    doi

  34. [42]

    M. Xue, Y . Wang, Z. Wang, L. Zhu, L. Cui, Y . Chen et al. AutoFDP: Automatic force-based model selection for multicriteria graph drawing. IEEE Transactions on Visualization and Computer Graphics, 32(2):1554– 1568, 2026. doi: 10.1109/TVCG.2025.3631659 9

    2026

  35. [43]

    M. Xue, Z. Wang, F. Zhong, Y . Wang, M. Xu, O. Deussen et al. Taurus: Towards a unified force representation and universal solver for graph layout.IEEE Transactions on Visualization and Computer Graphics, 29(1):886–895, 2023. doi: 10.1109/TVCG.2022.3209371 2

    2023

  36. [44]

    Yang and J

    J. Yang and J. Leskovec. Defining and evaluating network communities based on ground-truth. InProceedings of the ACM SIGKDD Workshop on Mining Data Semantics, MDS ’12, art. no. 3, 8 pages. Association for Computing Machinery, New York, NY , USA, 2012. doi: 10.1145/2350190 .235...

    2012 doi

  37. [45]

    H. D. Young, R. A. Freedman, T. Sandin, and A. L. Ford.University Physics, vol. 9. Addison-Wesley Reading, MA, 1996. 3

    1996

  38. [46]

    J. X. Zheng, S. Pawar, and D. F. M. Goodman. Graph drawing by stochastic gradient descent.IEEE Transactions on Visualization & Computer Graph- ics, 25(09):2738–2748, Sept. 2019. doi: 10.1109/TVCG.2018.2859997 5

    2019

  39. [47]

    Zhong, M

    F. Zhong, M. Xue, J. Zhang, F. Zhang, R. Ban, O. Deussen et al. Force- directed graph layouts revisited: a new force based on the t-distribution. IEEE Transactions on Visualization and Computer Graphics, 30(7):3650– 3663, 2023. doi: 10.1109/TVCG.2023.3238821 1, 2, 3, 4, 6, 7

    2023

  40. [48]

    F. Zhou, S. Malher, and H. Toivonen. Network simplification with minimal loss of connectivity. InProceedings of the IEEE International Conference on Data Mining (ICDM), pp. 659–668. IEEE, 2010. doi: 10.1109/ICDM. 2010.133 3

    2010 doi

  41. [49]

    M. Zhu, W. Chen, Y . Hu, Y . Hou, L. Liu, and K. Zhang. DRGraph: An efficient graph layout algorithm for large-scale graphs by dimensionality reduction.IEEE Transactions on Visualization and Computer Graphics, 27(2):1666–1676, 2020. doi: 10.1109/TVCG.2020.3030447 1, 2, 3, 4, 6, 7

    2020

  42. [50]

    Zinkevich, M

    M. Zinkevich, M. Weimer, L. Li, and A. Smola. Parallelized stochastic gradient descent.Advances in Neural Information Processing Systems, 23,

  43. [138]

    doi: 10.1007/978-3-030-35802-0_10 1, 7

    Springer, 2019. doi: 10.1007/978-3-030-35802-0_10 1, 7

    2019 doi

  44. [2007]

    doi: 10.1073/pnas.0610245104 6

    doi

  45. [2024]

    doi: 10.1111/cgf.15073 1

    doi

  46. [2025]

    doi: 10.1111/cgf.70105 2

    doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.