Inductive Graph Layout with Implicit Neural Fields

Berfin Inal; Daniel Probst

REVIEW 2 major objections 6 minor 33 references

Inductive Graph Layout with Implicit Neural Fields

T0 review · 2 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A graph drawing can be a fixed-size neural map trained on the layout energy, not a table of N free coordinates.

desk verdict Solid within-graph results and honest limitations, but the cross-graph inductive claim is asserted, not tested; fixable by scoping or by a real cross-graph experiment. read the letter →

arxiv 2608.08876 v1 pith:R7SRMQCJ submitted 2026-08-09 cs.HC cs.LG

Berfin Inal , Daniel Probst This is my paper · ORCID

classification cs.HCcs.LG

keywords graphdrawingimplicitneuralfieldsKamada-Kawaistressinductivelayoutout-of-sampleplacementdiffusionpotentialspivotboundlossconditioning

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

This paper claims that a graph drawing can be produced by a fixed-size neural map instead of a table of N free coordinates. The proposed method, Fling, trains a small network to map each node's diffusion-potential features to a 2-D position, optimising the network's weights on the layout energy itself rather than regressing onto a precomputed drawing. Because the unknown is a weight vector of fixed size, the energy can be evaluated on a sample of nodes, rest lengths can come from a landmark bound in constant time per pair, and a second learned field can approximate the expensive all-pairs majorisation sums at O(|A|N) per step. If the paper is right, layouts become inductive, an unseen node is placed by a single forward pass, and force laws normally considered too expensive, such as full Kamada-Kawai stress and aesthetic terms like node-edge clearance and crossings, become affordable on the same field. The paper reports that the stress variant matches stochastic-gradient stress within a few percent and that the field beats PivotMDS, landmark MDS, and kernel ridge when fitting an energy from a sample of nodes.

What carries the argument

The load-bearing object is the coordinate network Φθ: R^k → $R^{2}$, a two-layer GELU MLP of width 128 that maps a node's diffusion potentials f_v = −log(S_vℓ + ε) to its position; the diffusion potentials come from a personalised PageRank kernel on the graph's normalised adjacency matrix. Three pieces carry the argument. First, the fixed parameter count (about 21k weights) makes the energy a function of a sample rather than of N, so training can use M < N nodes. Second, rest lengths for Fling are supplied by the pivot bound r_ij = max_s |h_is − h_js| from p = 16 BFS pivots, evaluated in O(p) per pair with no distance matrix. Third, a second field Φ̃ψ: $R^{10}$ → $R^{3}$, fit each step at m = 80 anchors and queried at all nodes, learns the two components of the majorisation numerator and the denominator whose quotient is the Jacobi target, replacing the O($N^{2}$) pair sum at O(|A|N) per step. For the other variants the work is done by sampled objectives (pivot batches, random negatives), and for the layout family, λ is injected through learned per-channel gains and shifts rather than concatenated, which the paper shows preserves most of the trade-off range.

What would settle it

On a small-world graph with high symmetry, compute the exact all-pairs majorisation sums and compare them with the anchor-fitted far field's outputs at the same nodes; if the learned sums deviate by more than the anchor sampling noise, or if Fling's stress worsens materially as the zero-bound fraction grows, the O(|A|N) system is solving a distorted objective rather than Kamada-Kawai stress.

Watch

Extended reading notes

Core claim

The central claim is that layout should be a function, not a table: each node v is positioned by x_v = Φθ(f_v), where Φθ is a small GELU MLP and f_v is the node's diffusion potential to a farthest-first landmark set. Training the weights θ on the layout energy, Fling minimises Kamada-Kawai stress without ever forming the N×N distance matrix: rest lengths are obtained from the pivot bound r_ij = max_s |h_is − h_js| over 16 BFS pivots, and a second field Φ̃ψ, fit at 80 resampled anchors per step and queried at every node, supplies the majorisation numerator and denominator. The paper argues this makes the full spring system O(|A|N) per step, gives the drawing an extensible domain (unseen nodes, and in principle whole graphs from the same population, by forward pass), and lets the same parameterisation carry pivot stress, neighbour embedding with clearance and crossing terms, and a λ-conditioned family of layouts from one run. On the nine benchmark graphs, the scale-normalised pivot-stress variant reaches 1.00–1.11 times the stress of full stochastic gradient stress, with median 1.02, while the neighbour-embedding variant draws the least occluded graphs and improves on post-hoc refinement; when the task is fitting the energy from a sample of M = min(500, N/2) nodes, the field leads every matched readout on every tested graph.

Load-bearing premise

For the headline variant, the paper assumes that rest lengths from the 16-pivot landmark bound, combined with majorisation sums learned from 80 anchors, faithfully reproduce exact all-pairs Kamada-Kawai stress; the bound is one-sided, and on small-world graphs it can vanish (it is zero on 11.7% of ego-Facebook pairs, floored at 0.9 hops).

Editorial extensions

If this is right

A graph drawing becomes a map from structural features to positions, so adding a node to a fitted graph costs one forward pass instead of a rerun of the optimiser.
Full Kamada-Kawai stress becomes tractable without its distance matrix: rest lengths from the pivot bound and majorisation sums from the learned far field give O(|A|N) per step.
The same field can carry aesthetic objectives that are normally dropped on large graphs, node-edge clearance and crossing counts, inside the energy, with per-step cost O(κm′).
Conditioning on the energy weight λ yields a continuous family of layouts in one coordinate frame, so trade-offs between stress and neighbourhood preservation can be sampled at inference.
When the energy is fitted from a sample of M nodes, the field beats PivotMDS, landmark MDS, and kernel ridge on the same features on all tested graphs, with the margin growing from M = 100 upward.

Reading between the lines

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

One testable next step is population transfer: train one field on a sample graph and place a much larger independent graph from the same generative family; the paper asserts this should work but does not evaluate it.
The learned-far-field trick is more general than Kamada-Kawai: any pair-specific force that can be written as a function of per-node features could be approximated at O(|A|N), which may extend the same cost saving to other energy models.
Where the pivot bound vanishes, a hybrid that falls back to exact BFS distances for near-zero r_ij, or adds pivots adaptively, could plausibly close the gap between Fling and FlingStress on small-world graphs, since the paper identifies the zero-bound fraction as the failure mode.
Because the unknowns are weights rather than coordinates, the same energy could be trained with M significantly larger than N/2 to refine one map rather than enlarge a table, which would test whether sample complexity really is the right framing.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

I'll get straight to it. Fling is a genuinely new formulation: graph layout as an energy-trained neural field over diffusion features, with fixed parameter count and one-forward-pass placement of new nodes. The paper's central empirical claim survives contact: a field trained on sampled pivot stress beats PivotMDS, landmark MDS, a linear readout, and kernel ridge on the same features/energy across the tested graphs (kernel at 1.02–1.30x, median 1.09), and FlingStress lands within 1.11x of sgd2, median 1.02. That's a real result, and the comparisons are carefully matched. The authors also get credit for disclosing where their own variants lose: Fling, the headline variant carrying the O(|A|N) far-field claim, trails FlingStress on eight of nine graphs, and Appendix B.5 gives an honest mechanism (one-sided pivot bound, vanishing pairs on ego-Facebook, floor at 0.9 hops).

The soft spot is the inductive claim's scope. Section 1 asserts that a field fitted on a sample places a much larger independent sample from the same population. No experiment does that. All fits are single-graph, with held-out nodes or columns only. The stress-test note is right that this is not just missing evidence: the input features are diffusion potentials to landmarks selected on the training graph. For a disjoint graph, there is no defined way to compute those features without either reselecting landmarks, which changes the frame and invalidates the trained map, or aligning to training-graph landmarks, which cannot be computed on a disjoint graph. The abstract's 'extensible domain' and the paper's 'inductive' title lean on that unsupported step. The fix is concrete: either run a real cross-graph experiment with a stated landmark rule and alignment, or explicitly scope 'inductive' to within-graph out-of-sample placement. The rest of the paper's value survives either way.

A smaller concern: the margin over a kernel ridge is real but modest, and the authors acknowledge it. The learned far field is clever but is evaluated only inside Fling, where it loses to FlingStress; a direct comparison against Barnes-Hut or a low-rank stress solver on larger graphs would strengthen the claim. There are also a few garbled sentences, e.g. in the Limitations section, and typos in the appendix, none affecting the math.

Overall: within-graph results are solid and honestly reported; the cross-graph inductive claim is overreach but cleanly separable from the core method. This paper deserves a serious referee, not a desk reject. I'd cite FlingStress as a within-graph inductive baseline and would bring the paper to reading group.

Referee Report

2 major / 6 minor

Summary. The paper introduces Fling, a family of graph-layout methods that replace the N×2 coordinate table with a fixed-size MLP mapping node features (diffusion potentials to landmarks) to 2D positions, trained directly on layout energies. Three variants are presented: Fling (majorised stress with a learned far field and pivot-bound rest lengths), FlingStress (scale-normalised pivot stress), and FlingVis (neighbour embedding with clearance and crossing terms). The paper claims a fixed parameter count, sample-efficient energy evaluation, out-of-sample placement of unseen nodes by forward pass, sub-quadratic pair-sum treatment, and a λ-conditioned continuum of layouts from one training run. Experiments on nine graphs compare the variants against classical and neural baselines; the sample-fitting comparison (Table 1) reports that FlingStress leads matched closed-form and kernel-ridge baselines for M≥100, and FlingStress approximates sgd2 stress within a few percent.

Significance. If the claims hold, the paper makes a useful conceptual contribution: graph layout as an implicit neural function rather than a coordinate table, enabling within-graph out-of-sample placement, conditioning on the energy weight, and inclusion of aesthetic terms normally dropped for cost. The empirical core is carefully constructed: matched feature blocks, closed-form baselines scored at their better block, exact all-pairs stress evaluation, and extensive ablations in the appendices. The paper is also refreshingly honest about known weaknesses, explicitly stating that no variant trains on the exact stress quantity and that Fling trails FlingStress on eight of nine graphs. The main gap is that the title and abstract advertise cross-graph inductive transfer, but no experiment supports that stronger claim.

major comments (2)

[§1 and Abstract] The abstract and Section 1 assert that inductive placement extends to whole graphs: 'the same holds true for whole graphs drawn from a single population, so a field fitted on a sample places a much larger independent sample by a single forward pass.' No experiment in the paper trains on one graph and evaluates on another from the same population; Table 1, §4.1, and §4.2 are all single-graph fits with held-out nodes or columns. The input features are diffusion potentials to landmarks L selected farthest-first from the training graph (§2.1). For an unseen graph, either the landmarks must be reselected on the new graph, which changes the feature frame and breaks the trained map, or the landmarks are training-graph nodes, for which no diffusion potential can be computed on a disjoint graph. The manuscript defines neither a canonical landmark rule for test graphs nor an alignment procedure. Please either provide a cross-graph transfer experiment or explicitly scope the inductive claim to within-graph out-of-sample placement; the current wording overclaims the main contribution.
[§2.2, Appendix B.5, §4.2] The Fling variant is presented as making the 'full spring system' tractable, but its training objective is not exact Kamada-Kawai stress. Rest lengths are the pivot bound r_ij = max_s |h_is - h_js|, which Appendix B.5 states is one-sided (r_ij ≤ d_ij), vanishes on 11.7% of ego-Facebook pairs (floored at 0.9 hops), and underestimates rest lengths in small-world graphs. The far-field majorisation sums are also learned from m anchors. Section 4.2 reports that Fling trails FlingStress on eight of nine graphs, and the Limitations section concedes the bound is the cause. The abstract and contribution statements should qualify that Fling optimises a bounded surrogate with learned pair sums rather than exact Kamada-Kawai stress, so that readers do not conflate the FlingStress results with the full-system claim.

minor comments (6)

[Limitations] The sentence 'To avoid the O(N^2) costs more than sampling noise.' is incomplete and should be rephrased to state which approximations do or do not average out with more steps.
[Appendix B.5] The sentence 'The looseness ranks with the deficit of Fling against FlingStress without predicting it' is unclear; please specify whether the bound's looseness is merely correlated with the deficit or whether the authors intend a monotone ranking.
[Appendix B.3] The metric name 'Crlosslessnes' should be 'Crosslessness'.
[§4.2] The text says a rare fold occurs in 'approximately 6.4% of random seeds' on dwt_1005, but Table C.1 shows one of three seeds is anomalous; please clarify how the 6.4% rate is derived and how it relates to the three-seed mean.
[Figure 2] The caption describes 'The grey band represents 60 independent fits' while the body text says 'Sixty independent fits... shown as a wide gray line'; please reconcile whether the independent fits are displayed as a band or a line.
[Author affiliation] The affiliation contains the typo 'Wageningen Univeristy & Research'; it should read 'Wageningen University & Research'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Fling trains on explicit surrogates and is scored on the exact all-pairs stress it never trains on.

full rationale

The derivation chain is self-contained. Fling and FlingStress are trained on approximate or sampled energies (pivot-bound majorised stress with a learned far-field, and scale-normalised pivot stress over |Q| columns), while every reported stress value is the exact all-pairs scale-normalised stress; the paper states 'No variant trains on the exact quantity' (Appendix B.3). The evaluation is therefore a stricter objective than the training target by construction, not the same quantity renamed. The closed-form scale a in FlingStress is 'recomputed from current positions with every step and is then held fixed for the duration of that step,' so it is not a fitted constant smuggled into the result. The lambda-conditioned field is fit only at endpoints and then sampled at unseen lambda values: 'Five of the seven panels and 194 further samples, shown as the black line, were never presented to the optimiser,' which is genuine interpolation rather than fitting-then-reporting the same points. The matched-budget comparison trains the field, linear readout, and kernel ridge on the same sampled energy and features, with the kernel 'choosing its bandwidth, learning rate, and budget by the same rule used by the field'; the field's lead is an empirical margin, not an identity. No load-bearing self-citation or imported uniqueness theorem appears; citations (PageRank diffusion, Barnes-Hut, FiLM, loss-conditional training, Muon, sgd2) are external and used as prior art. The only questionable assertion is the §1 extrapolation that a field fitted on one graph 'places a much larger independent sample' from the same population; no experiment evaluates cross-graph transfer, and the landmark-based diffusion feature frame is graph-specific. That is an unsupported extrapolation and a correctness risk, but it is not a circular reduction of the claimed result to its own inputs, so it does not raise the circularity score.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claims rest on standard triangle-inequality bounds, several smoothness and approximation assumptions about neural surrogates and diffusion features, and a set of hand-chosen hyperparameters. No new physical entities are introduced; the learned far field is a standard surrogate network, and the label-keyed probes are input features.

free parameters (8)

pivot_count = 16
Rest-length bound r_ij = max_s |h_is - h_js| uses p=16 farthest-first pivots; bound tightness and stress depend on it (Appendix B.5).
anchor_count = 80
Learned far field is fit at m=80 anchors per step; defines the O(|A|N) step cost and the surrogate quality.
target_pivots = 400 (capped to 76/239 on small graphs)
FlingStress's exact BFS distance targets; per-step batch is min(128,|Q|).
diffusion_landmarks = 64
Input feature dimension k=64; no ablation of this count is reported.
random_walk_length = not specified
The diffusion potential depends on a walk length K that is never assigned a value in the paper.
restart_probability = 0.05
Sets the diffusion kernel S in §2.1.
pivot_bound_floor = 0.9 hops
Pairs with r_ij=0 are floored at 0.9 hops, which sets rest lengths on 11.7% of ego-Facebook pairs (Appendix B.5).
network_architecture = 128x2 GELU MLP
Default network; Table D.2 shows a flat plateau but it is chosen by hand.

assumptions (6)

standard math Triangle inequality gives r_ij = max_s |h_is - h_js| <= d_ij for the pivot bound.
Used in §2.2 and B.5 to justify the rest-length approximation.
domain assumption Diffusion potentials are a smooth, informative function of the K-hop neighbourhood, giving unseen nodes features in the same frame.
Load-bearing for inductive placement; stated in §2.1 and validated only on the nine benchmark graphs.
domain assumption A second network trained at m anchors per step accurately predicts the majorisation sums at all nodes.
Stated in §2.2; the far field's error is disclosed as not averaging out in the Limitations section.
domain assumption A network trained on M sampled nodes generalizes to the N-M held-out nodes of the same graph.
The basis of §4.1's sample-fitting experiment; empirically validated on the test graphs but not guaranteed beyond them.
ad hoc to paper Graphs drawn from a single population share a common feature-to-position map.
Section 1 claims induction to whole graphs from a population, but no cross-graph experiment is run.
domain assumption FiLM conditioning on lambda lets the network interpolate a continuum of layouts between two energies.
§2.5 and Appendix C.4 show the interpolation reaches about three quarters of the range of independently fitted fields.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Inductive Graph Layout with Implicit Neural Fields." pith.science (2026). https://pith.science/paper/R7SRMQCJ

@misc{pith2026260808876,
  author       = {Pith},
  title        = {Pith review of: Inductive Graph Layout with Implicit Neural Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7SRMQCJ}},
  note         = {Machine review of arXiv:2608.08876}
}

abstract

A graph layout is normally a table of $N$ free coordinates. We optimise a function with a fixed number of parameters instead. This gives a drawing a sample complexity and an extensible domain. Force-directed algorithms remain the standard tools for graph drawing. The most accurate among them minimise stress in the Kamada-Kawai formulation by directly optimising the node coordinates, at a full objective cost of $O(N^2)$ in time and space. Here, we propose Fling (Field Layout via Implicit Neural Geometry), a small neural network mapping the distances of each node to a set of landmarks, positioning it in the plane by training on the layout energy. The full spring system then becomes tractable without its distance matrix, as rest lengths follow from a landmark bound in constant time per pair while a second network learns the majorisation sums from exact anchor rows, at $O(|\mathcal{A}|N)$ per step for $|\mathcal{A}|\ll N$ anchors. Unlike neural drawers that read the graph by message passing, we represent the drawing as a function of node features. An unseen node costs one forward pass, where sparse and low-rank majorisation remain transductive. As the unknowns are weights rather than coordinates, the energy only requires a small fraction of the nodes, and a field fitted that way outperforms PivotMDS, landmark MDS, and a kernel ridge trained on the same energy and features, when the task is fitting the energy of a graph from a sample of its nodes. In addition, the same parameterisation enables a stochastic pivot stress variant, an aesthetics-optimised variant carrying a neighbour-embedding energy with node-edge clearance and crossing terms on the same field, and conditioning on the weight between two energies gives a whole layout family from one run.

Figures

Figures reproduced from arXiv: 2608.08876 by the authors.

**Figure 1.** The Fling pipeline. All variants read the diffusion potentials fv as input and drive a coordinate network Φθ to node positions. They only differ in the energy the weights are trained on, which is all-pairs stress majorisation for Fling, scale-normalised pivot stress for FlingStress, and a neighbour-embedding energy with sampled negatives plus node-edge and crossing terms for FlingVis. The second field Φψ replaces th… view at source ↗

**Figure 2.** The shaded panels show the two λ the field was fitted at. The five in between, together with the black line in the line plots, sample at unseen λ, The grey band represents 60 independent fits and the red line the field resampled at nine λ values. A family of drawings from a single run. Conditioning the field on the weight λ between two energies makes the blend between the energies a continuous function, so it can be… view at source ↗

**Figure 3.** draws nine benchmark graphs by the three Fling variants and four baselines [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: scores FlingVis on lesmis and cora against baselines. On both, FlingVis draws the least occluded graph of any method. Adding the two aesthetic terms to a FlingVis variant without them (base) decreases occlusion while keeping purity and stress stable. Introducing the te…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

[1]

Graph drawing by force-directed placement

Thomas MJ Fruchterman and Edward M Reingold. Graph drawing by force-directed placement. Software: Practice and experience, 21(11):1129–1164, 1991. 1

work page 1991
[2]

Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software.PloS one, 9(6):e98679, 2014

Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software.PloS one, 9(6):e98679, 2014

2014
[3]

Graph drawing by stress majorization

Emden R Gansner, Yehuda Koren, and Stephen North. Graph drawing by stress majorization. InInternational Symposium on Graph Drawing, pages 239–250. Springer, 2004

work page 2004
[4]

A maxent-stress model for graph layout

Emden R Gansner, Yifan Hu, and Stephen North. A maxent-stress model for graph layout. IEEE transactions on visualization and computer graphics, 19(6):927–940, 2012. 1

work page 2012
[5]

An algorithm for drawing general undirected graphs

Tomihisa Kamada, Satoru Kawai, et al. An algorithm for drawing general undirected graphs. Information processing letters, 31(1):7–15, 1989. 1

work page 1989
[6]

Graph drawing by stochastic gradient descent.IEEE transactions on visualization and computer graphics, 25(9):2738–2748,

Jonathan X Zheng, Samraat Pawar, and Dan FM Goodman. Graph drawing by stochastic gradient descent.IEEE transactions on visualization and computer graphics, 25(9):2738–2748,

work page
[7]

Drawing large graphs by low-rank stress majorization

Marc Khoury, Yifan Hu, Shankar Krishnan, and Carlos Scheidegger. Drawing large graphs by low-rank stress majorization. InComputer Graphics Forum, volume 31, pages 975–984. Wiley Online Library, 2012. 1

work page 2012
[8]

A sparse stress model

Mark Ortmann, Mirza Klimenta, and Ulrik Brandes. A sparse stress model. InInternational symposium on graph drawing and network visualization, pages 18–32. Springer, 2016. 1

work page 2016

Show all 33 references

[9]

Accelerating network layouts using graph neural networks.Nat

Csaba Both, Nima Dehmamy, Rose Yu, and Albert-László Barabási. Accelerating network layouts using graph neural networks.Nat. Commun., 14(1560):1560, March 2023. ISSN 2041-1723. doi: 10.1038/s41467-023-37189-2. 2, 12

2023 doi
[10]

Deepgd: A deep learning framework for graph drawing using gnn.IEEE computer graphics and applications, 41(5):32–44, 2021

Xiaoqi Wang, Kevin Yen, Yifan Hu, and Han-Wei Shen. Deepgd: A deep learning framework for graph drawing using gnn.IEEE computer graphics and applications, 41(5):32–44, 2021. 2, 12

2021
[11]

Core-gd: A hierarchical framework for scalable graph visualization with gnns.arXiv preprint arXiv:2402.06706, 2024

Florian Grötschla, Joël Mathys, Robert Veres, and Roger Wattenhofer. Core-gd: A hierarchical framework for scalable graph visualization with gnns.arXiv preprint arXiv:2402.06706, 2024. 12

2024 arXiv
[12]

Word2vecgd: Neural graph drawing with cosine-stress optimization.arXiv preprint arXiv:2509.17333, 2025

Minglai Yang and Reyan Ahmed. Word2vecgd: Neural graph drawing with cosine-stress optimization.arXiv preprint arXiv:2509.17333, 2025. 2, 12

2025
[13]

Nnp-net: Accelerating t-sne graph drawing for large static and dynamic graphs by neural networks.IEEE Transactions on Visualization and Computer Graphics, 2026

Ilan Hartskeerl, Tamara Mchedlidze, Simon van Wageningen, Peter Vangorp, and Alexandru Telea. Nnp-net: Accelerating t-sne graph drawing for large static and dynamic graphs by neural networks.IEEE Transactions on Visualization and Computer Graphics, 2026. 2, 12

2026
[14]

Kruiger, Paulo Eduardo Rauber, Rafael Messias Martins, Andreas Kerren, Stephen G

Johannes F. Kruiger, Paulo Eduardo Rauber, Rafael Messias Martins, Andreas Kerren, Stephen G. Kobourov, and Alexandru Cristian Telea. Graph layouts by t-sne.Computer Graph- ics Forum, 36, 2017. URLhttps://api.semanticscholar.org/CorpusID:1235159. 2

2017
[15]

Multivariate Graph Layouts using NNP-NET

Ilan Hartskeerl, Florian Grötschla, and Alexandru Telea. Multivariate Graph Layouts using NNP-NET . In Claudio D. G. Linhares, Andreas Kerren, Rafael Martins, Amilcar Soares, Alessio Arleo, Stef van den Elzen, and Fernando V . Paulovich, editors,EuroVis Workshop on GDxDR. The ...

2026 doi
[16]

GNN-Based Inductive Dimensionality Reduction on UMAP Graphs

Florian Grötschla, Simon van Wageningen, and Alexandru Telea. GNN-Based Inductive Dimensionality Reduction on UMAP Graphs. In Claudio D. G. Linhares, Andreas Kerren, Rafael Martins, Amilcar Soares, Alessio Arleo, Stef van den Elzen, and Fernando V . Paulovich, editors,EuroVis ...

2026 doi
[17]

Visualizing large-scale and high- dimensional data

Jian Tang, Jingzhou Liu, Ming Zhang, and Qiaozhu Mei. Visualizing large-scale and high- dimensional data. WWW ’16, page 287–297, Republic and Canton of Geneva, CHE, 2016. International World Wide Web Conferences Steering Committee. ISBN 9781450341431. doi: 10.1145/2872427.2883...

2016
[18]

Drgraph: An efficient graph layout algorithm for large-scale graphs by dimensionality reduction.IEEE Transactions on Visualization and Computer Graphics, 27:1666–1676, 2020

Minfeng Zhu, Wei Chen, Yuanzhe Hu, Yuxuan Hou, Liangjun Liu, and Kaiyuan Zhang. Drgraph: An efficient graph layout algorithm for large-scale graphs by dimensionality reduction.IEEE Transactions on Visualization and Computer Graphics, 27:1666–1676, 2020. URL https: //api.semant...

2020
[19]

Impred: An improved force-directed algorithm that prevents nodes from crossing edges.Computer Graph- ics Forum, 30(3):1071–1080, 2011

Paolo Simonetto, Daniel Archambault, David Auber, and Romain Bourqui. Impred: An improved force-directed algorithm that prevents nodes from crossing edges.Computer Graph- ics Forum, 30(3):1071–1080, 2011. doi: https://doi.org/10.1111/j.1467-8659.2011.01956. x. URL https://onli...

2011
[20]

Multicrite- ria scalable graph drawing via stochastic gradient descent, (sgd)2(sgd)2.IEEE Transactions on Visualization and Computer Graphics, 28(6):2388–2399, 2022

Reyan Ahmed, Felice De Luca, Sabin Devkota, Stephen Kobourov, and Mingwei Li. Multicrite- ria scalable graph drawing via stochastic gradient descent, (sgd)2(sgd)2.IEEE Transactions on Visualization and Computer Graphics, 28(6):2388–2399, 2022. doi: 10.1109/TVCG.2022. 3155564. 3

2022 doi
[21]

Scaling personalized web search

Glen Jeh and Jennifer Widom. Scaling personalized web search. InProceedings of the 12th international conference on World Wide Web, pages 271–279, 2003. 3

2003
[22]

J. H. Barnes and Piet Hut. A hierarchical o(n log n) force-calculation algorithm.Nature, 324: 446–449, 1986. URLhttps://api.semanticscholar.org/CorpusID:4267861. 3

1986
[23]

Courville

Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron C. Courville. Film: Visual reasoning with a general conditioning layer. InAAAI Conference on Artificial Intelligence,
[24]

You only train once: Loss-conditional training of deep networks

Alexey Dosovitskiy and Josip Djolonga. You only train once: Loss-conditional training of deep networks. InInternational Conference on Learning Representations, 2020. URL https://api.semanticscholar.org/CorpusID:214278158. 6

2020
[25]

Eigensolver methods for progressive multidimensional scaling of large data

Ulrik Brandes and Christian Pich. Eigensolver methods for progressive multidimensional scaling of large data. InInternational Symposium Graph Drawing and Network Visualization,
[26]

Sparse multidimensional scaling using landmark points

Vin De Silva and Joshua B Tenenbaum. Sparse multidimensional scaling using landmark points. Technical report, technical report, Stanford University, 2004. 6

2004
[27]

Learning continuous image representation with local implicit image function

Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8624–8634, 2021. doi: 10.1109/CVPR46437.2021.00852. 7

2021
[28]

normalized stress

Kiran Smelser, Jacob Miller, and Stephen G. Kobourov. “normalized stress” is not nor- malized: How to interpret stress correctly.2024 IEEE Evaluation and Beyond - Method- ological Approaches for Visualization (BELIV), pages 41–50, 2024. URL https://api. semanticscholar.org/Cor...

2024
[29]

Purchase

Helen C. Purchase. Metrics for graph drawing aesthetics.J. Vis. Lang. Comput., 13:501–516,
[30]

table on M rows

Keller Jordan, Yuchen Jin, Vlado Boza, Jiacheng You, Franz Cesista, Laker Newhouse, and Jeremy Bernstein. Muon: An optimizer for hidden layers in neural networks, 2024. URL https://kellerjordan.github.io/posts/muon/. 18 11 Inductive Graph Layout with Implicit Neural Fields A G...

2024
[2002]

URLhttps://api.semanticscholar.org/CorpusID:5895591. 12
[2006]

URLhttps://api.semanticscholar.org/CorpusID:17293104. 6
[2017]

URLhttps://api.semanticscholar.org/CorpusID:19119291. 6

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Graph drawing by force-directed placement

Thomas MJ Fruchterman and Edward M Reingold. Graph drawing by force-directed placement. Software: Practice and experience, 21(11):1129–1164, 1991. 1

work page 1991

[2] [2]

Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software.PloS one, 9(6):e98679, 2014

Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software.PloS one, 9(6):e98679, 2014

2014

[3] [3]

Graph drawing by stress majorization

Emden R Gansner, Yehuda Koren, and Stephen North. Graph drawing by stress majorization. InInternational Symposium on Graph Drawing, pages 239–250. Springer, 2004

work page 2004

[4] [4]

A maxent-stress model for graph layout

Emden R Gansner, Yifan Hu, and Stephen North. A maxent-stress model for graph layout. IEEE transactions on visualization and computer graphics, 19(6):927–940, 2012. 1

work page 2012

[5] [5]

An algorithm for drawing general undirected graphs

Tomihisa Kamada, Satoru Kawai, et al. An algorithm for drawing general undirected graphs. Information processing letters, 31(1):7–15, 1989. 1

work page 1989

[6] [6]

Graph drawing by stochastic gradient descent.IEEE transactions on visualization and computer graphics, 25(9):2738–2748,

Jonathan X Zheng, Samraat Pawar, and Dan FM Goodman. Graph drawing by stochastic gradient descent.IEEE transactions on visualization and computer graphics, 25(9):2738–2748,

work page

[7] [7]

Drawing large graphs by low-rank stress majorization

Marc Khoury, Yifan Hu, Shankar Krishnan, and Carlos Scheidegger. Drawing large graphs by low-rank stress majorization. InComputer Graphics Forum, volume 31, pages 975–984. Wiley Online Library, 2012. 1

work page 2012

[8] [8]

A sparse stress model

Mark Ortmann, Mirza Klimenta, and Ulrik Brandes. A sparse stress model. InInternational symposium on graph drawing and network visualization, pages 18–32. Springer, 2016. 1

work page 2016