acceptodds
Under review as a conference paper at ICLR 2027

TopoForce: Learning Topology-Revealing Representations across Multiple Scales

Abstract

Topology provides a fundamental description of high-dimensional data, yet the underlying topological organization is typically latent and can contain structures emerging at heterogeneous intrinsic scales. Conventional dimensionality-reduction methods optimize geometric or neighborhood fidelity, which need not preserve collective topology. Persistent-homology (PH)-guided methods use persistent structure inferred from the observed metric as a target for topology preservation, but this structure can become unreliable in high-dimensional noisy observations. We introduce TopoForce, a PH-free graph-based method for learning multiscale topology-revealing representations. TopoForce uses density-normalized multiscale diffusion to infer a pairwise topology scale that characterizes the structural scale relevant to each relationship, and translates these scales into adaptive force-directed interactions. Across controlled simulations, TopoForce recovers topological features spanning heterogeneous scales and remains robust to nonuniform sampling and high-dimensional nuisance variation. Across real-world applications in single-cell genomics, image analysis, and neural population activity, TopoForce reveals cyclic and toroidal organization consistent with known underlying structure. Together, these results demonstrate that TopoForce can reveal latent topological structure across multiple intrinsic scales in high-dimensional data.Our code is available at https://anonymous.4open.science/r/topoforce_2026-34CE/.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topological representation learning, multiscale topology, graph embedding, graph diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PLgLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces TopoForce, a PH-free, graph-based dimensionality-reduction method for learning representations that reveal multiscale topology. The central motivation is that many topology-preserving DR methods use persistent homology (PH) of the observed data as a target. When observed geometry is distorted by noise, nuisance variation, or nonuniform sampling, that target PH can be misleading, so directly preserving it may propagate distortion into the embedding. TopoForce instead builds a neighborhood graph, performs multiscale diffusion, and infers a pairwise topology scale for each pair of points. This scale reflects which diffusion modes most strongly distinguish the pair and how slowly those modes decay.

Critical strengths

I appreciate the move away from PH method, as it has known scalability and performance under noise issues. Experimental setup seems robust, especially the fact that it tests many scenarios with noise and distortion/variation.

Critical weaknesses

Scalability of TopoForce is still a major issue compared to other methods (table 3). It is also unclear how the parameters in Appendix C were chosen, and this may be a reason why TopoForce outperforms other methods? Finally, hhy experiments were not run on multiple seeds escapes me.

Questions for authors
  1. How were hyperparameters for TopoForce chosen? Is the parameter search conducted fair versus all other baseline? More information on this section is crucial to assess TopoForce fairly.
  2. Why was only one seed used for experiments? Is it because the original data is deterministic? What about noise injection?
  3. Is the method easily extendable to higher dimensional data?
4 comments
AuthorsOpenReview ↗

Response to Reviewer PLgL

We thank the reviewer for recognising the motivation behind TopoForce. Geometric distortion can make a loop hard to detect even when the underlying topology is unchanged. TopoForce uses multiscale diffusion on a neighbourhood graph to guide a low-dimensional layout that reveals structures at different scales.

Additional concern: Computational scalability

We agree that computational efficiency is important for practical applications. To improve the scalability of the layout optimisation, we implemented a Barnes--Hut approximation for repulsive force computation, which reduces computational cost by grouping distant points.

We evaluated this acceleration on four synthetic structures (circle, two-loop ring, three unequal loops and torus), keeping the graph, pairwise scales, initialization and iteration budget (1,600 steps) fixed across two seeds. At a Barnes--Hut threshold of 0.5, layout computation achieved a 4.14.1--6.2×6.2\times speed-up, with aligned coordinate differences of only 0.840.84--3.53%3.53\%. Importantly, identity-aware H1H_1 matching on shared 350-point subsets retained all reference targets in the circle, ring and three-loop cases across both seeds. These results demonstrate that the layout stage can be substantially accelerated while maintaining the topological structures recovered by the original method. See Rebuttal Figure R3 hyperlink for the five-seed results.

AuthorsOpenReview ↗

Response to Reviewer PLgL Q1

Q1. Hyperparameter selection and baseline fairness

"How were hyperparameters for TopoForce chosen? Is the parameter search conducted fair versus all other baseline? More information on this section is crucial to assess TopoForce fairly."

We have clarified the parameter-selection procedure and provide the effective configurations used in the experiments. In TopoForce, the neighbourhood size (k=15k=15) and affinity bandwidth (h=0.3h=0.3) are held fixed across the reported experiments. The diffusion reference time tVNEt_{\mathrm{VNE}} is selected adaptively from the knee of the Von Neumann entropy curve. The force coefficients (cattr=1c_{\mathrm{attr}}=1, crep=16c_{\mathrm{rep}}=16, and cgrav=1c_{\mathrm{grav}}=1) are fixed, while the κ\kappa schedule is determined adaptively from the distribution of pairwise topology scales. Thus, the main force-design parameters are not independently tuned for individual benchmark conditions.

To further assess baseline sensitivity, we completed 256 parameter configurations across UMAP, t-SNE, Isomap, PHATE and Diffusion Maps on distorted and nonuniform bracelet and three-loop inputs. The scans cover neighbourhood sizes of 5, 15, 30 and 45, UMAP minimum distances from 0 to 0.9, t-SNE perplexities from 5 to 45, and diffusion times from 0 to 100, using seed 0. The resulting embeddings are available for visual comparison in the baseline hyperparameter sweeps (Rebuttal Figure R2 hyperlink).

These additional scans provide a broader assessment of how baseline representations respond to their respective parameter choices, rather than relying solely on default configurations.

AuthorsOpenReview ↗

Response to Reviewer PLgL Q2

Q2. Random seeds and noise injection

"Why was only one seed used for experiments? Is it because the original data is deterministic? What about noise injection?"

We acknowledge that reporting only one seed in the original submission did not allow the variability arising from random initialization and optimisation to be assessed. We have therefore expanded the evaluation to five algorithmic seeds (0–4), covering 20 synthetic conditions, four real-data conditions and 30 nuisance conditions, with 3,285 valid results across 13 methods. We report individual seed values, means and sample standard deviations.

On distorted three-loop data, TopoForce obtains a TRS of 0.9630±0.00170.9630 \pm 0.0017, compared with 0.7962±0.10460.7962 \pm 0.1046 for DIPOLE, the strongest competing method by mean TRS. On nonuniform three-loop data, the corresponding scores are 0.9779±0.00230.9779 \pm 0.0023 and 0.9565±0.00550.9565 \pm 0.0055. These results demonstrate consistent performance across algorithmic seeds, particularly on the distorted three-loop benchmark.

Updated results for Bracelet and Three loops

Topology recovery score (TRS, Eq. 9) under uniform (U), distorted uniform (UD), and nonuniform (NU) sampling. Each cell gives the mean ± sample SD over optimizer seeds 0–4 on the same input. Bold and italic values indicate the highest and second-highest distinct means, respectively, across all 13 methods before display rounding.

MethodBracelet U (H1H_1)Bracelet UD (H1H_1)Bracelet NU (H1H_1)Three loops U (H1H_1)Three loops UD (H1H_1)Three loops NU (H1H_1)
TopoForce0.970 ± 0.0000.877 ± 0.0000.891 ± 0.0010.975 ± 0.0020.963 ± 0.0020.978 ± 0.002
UMAP0.944 ± 0.0050.836 ± 0.0280.619 ± 0.1070.600 ± 0.1410.668 ± 0.1060.609 ± 0.219
t-SNE0.947 ± 0.0000.445 ± 0.0000.465 ± 0.0000.548 ± 0.0000.454 ± 0.0000.413 ± 0.000
FLE0.509 ± 0.1230.455 ± 0.1450.439 ± 0.1010.510 ± 0.1440.461 ± 0.0940.623 ± 0.131
PHATE0.754 ± 0.0480.567 ± 0.0800.551 ± 0.1190.504 ± 0.0330.463 ± 0.0340.474 ± 0.055
Isomap0.919 ± 0.0000.629 ± 0.0000.851 ± 0.0000.923 ± 0.0000.329 ± 0.0000.936 ± 0.000
DIPOLE0.626 ± 0.0630.643 ± 0.0330.730 ± 0.0170.947 ± 0.0150.796 ± 0.1050.957 ± 0.006
TopoAE++0.868 ± 0.0350.600 ± 0.1590.738 ± 0.2300.845 ± 0.1070.658 ± 0.1330.903 ± 0.027
RTD-AE0.780 ± 0.2140.476 ± 0.2090.889 ± 0.0020.911 ± 0.0010.732 ± 0.0490.830 ± 0.134
Diffusion Maps0.769 ± 0.0000.346 ± 0.0000.466 ± 0.0000.000 ± 0.0000.000 ± 0.0000.323 ± 0.000
Laplacian Eigenmaps0.650 ± 0.0000.329 ± 0.0000.436 ± 0.0000.032 ± 0.0000.000 ± 0.0000.327 ± 0.000
Stress majorization0.916 ± 0.0020.792 ± 0.0050.734 ± 0.0100.908 ± 0.0020.783 ± 0.0010.903 ± 0.006
tsNET0.307 ± 0.0430.265 ± 0.0690.208 ± 0.1000.440 ± 0.1430.389 ± 0.2100.407 ± 0.085

To isolate algorithmic variability from variability introduced during data generation, the five-seed repetitions reuse identical sampled point coordinates and nuisance-noise realisations within each condition. Thus, the reported standard deviations quantify variability arising from initialization and optimisation conditional on the same input observations, rather than variability across independently generated noise realisations. Input identities and effective run configurations are recorded with the results in Rebuttal Figure R3 hyperlink, which also provides the five-seed results.

AuthorsOpenReview ↗

Response to Reviewer PLgL Q3

Q3. Applicability to higher-dimensional data

"Is the method easily extendable to higher-dimensional data?"

TopoForce operates on a neighbourhood graph, and its diffusion and pairwise-scale calculations are therefore not intrinsically restricted to low-dimensional input vectors. We provide empirical evidence through our high-dimensional nuisance experiments, which extend three synthetic structures with up to 2,048 additional Gaussian dimensions. Five algorithmic seeds are evaluated on each fixed input, allowing us to assess whether the recovered structures remain detectable as the input dimensionality increases.

For the output representation, the pairwise force formulation is mathematically applicable to embedding spaces of arbitrary dimension, since the pairwise interactions are defined through Euclidean distances and direction vectors in the embedding space. However, our current implementation and experiments focus on two- and three-dimensional embeddings.

We note that, although the method is not intrinsically restricted by input dimensionality, its performance still depends on constructing a meaningful neighbourhood graph, which can become challenging in high-dimensional settings.

The three-loop results are summarised in Figure 5 hyperlink. For embeddings and persistence diagrams for all three structures, see Appendix Figures A6–A11 hyperlink.

Reviewer_cegEOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents TopoForce, a new method for embedding topological information into manifolds. The authors use the neighbourhood graph to capture topology. They then use a force-based embedding technique to capture the topology. In experiments, the authors show that the proposed method can reconstruct the topology.

Critical strengths
  1. The core idea is simple and works on the examples it targets. TopoForce uses the relaxation times of the diffusion modes separating each pair to choose a pair-specific repulsion scale (Equations 3–4). The numerical gaps are significant on the experiments included in the paper. On the three-loop example, TopoForce gets UD 0.963, compared with 0.765 for the next-best method. Figure 7 also suggests that the pairwise scales are useful, rather than the density correction or the optimisation schedule.

  2. The paper provides enough detail to assess and reproduce the method and is honest about the limitations. The authors also report several cases where TopoForce does not win: torus H1H_1, the uniform sphere, cell cycle, grid cells, and neighbour recall in Table A1.

Critical weaknesses
  1. I am not convinced that the robustness experiments support the claims made from them. The topology recovery score in Equation 9, ℓℓ+b,\frac{\ell}{\ell+b}, is not calibrated against a null model. TRS may also favor the particular geometry produced by TopoForce. It is largely unchanged by a smooth deformation of a correct embedding. TopoForce usually draws loops as thin curves, while UMAP, t-SNE, and the autoencoders produce broader bands. Some of the margin on the bracelet and three-loop examples may therefore reflect the concentration of points around a curve rather than a difference in topology.

    The nuisance experiment also lacks another control. Section 5.4 appends isotropic Gaussian coordinates and embeds the raw vectors but omits PCA. This is a natural baseline for additive high-dimensional noise and is especially relevant because the paper applies PCA before every real-data experiment, reducing the inputs to 50, 6, and 40 dimensions. As written, the synthetic robustness test does not match the preprocessing used in practice.

    There is also too little information to judge the smaller differences in Table 1. The table reports one number per method, without seeds or standard deviations, and Section 5.1 says that the baselines use default hyperparameters. The ablation is likewise based on one benchmark and one run. I could not find a complete specification of the synthetic generators, including the distortion map, nonuniform sampling, noise magnitude, loop radii, and initialisation. Under these conditions, differences such as 0.995 versus 0.990 on the uniform benchmarks are not meaningful.

  2. The closest baselines are missing, and the paper does not measure the geometry lost in producing the cleaner layouts. The pairwise scales in Equations 3–4 are computed from diffusion-map eigenpairs at PHATE’s tVNEt_{\mathrm{VNE}}, but neither Diffusion Maps nor Laplacian Eigenmaps is included in the experiments, despite both being discussed in Section 2. These comparisons are needed to show what the force-directed stage adds beyond the spectral representation it receives.

    TopoForce depends on the neighbourhood graph and the scale matrix SijS_{ij}, and the resulting embeddings often look like standardized drawings of that graph. In Figure 3, for example, the distorted and undistorted bracelets produce nearly identical embeddings, with the original differences in loop size largely removed. The rank normalization in Equation 5 also spreads the force exponents across [0,1][0,1] even when the underlying scales differ only slightly. This may improve the topological score at the cost of metric information. The paper reports no neighbourhood- or distance-preservation metric that would reveal this trade-off. A graph-layout baseline, such as stress majorization or tsNET, would also be useful here.

    The real-data experiments do not provide much additional support. TopoForce is best on only one of the three rows in Table 2. None of the methods recovers the torus in the grid-cell data, although the source study reports one; every method scores at most 0.05 on COIL-20 Object 3; and FLE and Isomap perform better on the cell-cycle data. Scalability is another concern: computing the O(n2)O(n^2) scale matrix takes 2,046 seconds at n=8,000n=8{,}000 in Table 3, while the real datasets evaluated contain at most about 1,500 points. The method may still be usable at larger scales, but the current results do not establish that.

Questions for authors

None

4 comments
AuthorsOpenReview ↗

C1. Topological Recovery

We thank the reviewer for raising this important point. To reveal topology, an embedding should correct distortions that hide loops, for example by making a flattened loop rounder. A round loop, a flattened loop and a square loop all have the same topology. Persistent homology (PH) on point clouds uses distances between sampled points, so geometry affects how clearly it detects that loop. In the noisy examples, flattening brings the loop's H₁ feature closer to the diagonal, making it harder to distinguish from noise (see Rebuttal Figure R4 hyperlink).

The rationale for our graph-based approach is that, even when the global geometry is distorted, local connectivity relationships may still retain informative structural organisation. TopoForce exploits this information through multiscale diffusion on the neighbourhood graph. Specifically, the diffusion spectrum captures structural organisation at different scales, while the spectral contrasts between individual pairs of observations determine which modes contribute to distinguishing each pair. By weighting these pair-specific contributions according to their relaxation times, we obtain a pairwise structural scale that characterises the multiscale context of each relationship. These scales are then translated into pair-dependent repulsive interactions, allowing relationships at different structural scales to shape the embedding differently.

Thus, TopoForce uses the neighbourhood graph not simply as a structure to be laid out, but as a source of multiscale information that guides the reorganisation of the representation. This approach relies on the neighbourhood graph retaining sufficient information about the underlying structural relationships; it does not assume that topology can be recovered if these relationships have been completely lost.

We address the reviewer's specific technical concerns below.

C1. Validity of Topological Recovery

A central concern is whether the improved topology recovery scores reflect genuine recovery of underlying topological structures, rather than biases arising from the geometry of the learned embeddings. We performed two complementary analyses to address this concern.

First, to assess the contribution of point correspondence to TRS, we performed correspondence permutation tests. These permutations preserve the geometry and persistence diagram of each point cloud while disrupting the correspondence between point identities. We completed 100 permutations for each of 48 dataset–method–seed evaluations. The resulting permutation distributions provide a null reference for assessing the correspondence-dependent component of TRS, distinguishing genuine structural correspondence from scores attributable solely to the geometric appearance and persistence of an embedding (see Rebuttal Figure R5, upper panels hyperlink).

Second, we further examined whether the thin curves typically produced by TopoForce, compared with the broader bands produced by UMAP, t-SNE and autoencoder-based methods, could intrinsically favour PH-based evaluation. The thin/thick ring and open-arc controls are shown in Rebuttal Figure R5, lower panels hyperlink. See Appendix Figures A4–A5 hyperlink for the nonuniform-sampling embeddings and their persistence diagrams.

We agree that PCA provides a natural preprocessing control for additive high-dimensional nuisance variation, particularly given that PCA is applied before our real-data experiments. We therefore conducted an additional PCA-based evaluation covering eight synthetic structures under uniform and nonuniform sampling, resulting in 16 conditions.

For each condition, the input was expanded to 100 dimensions, centred, and reduced by full-SVD PCA to its known native dimension (two or three), without standardisation or whitening. TopoForce, UMAP, t-SNE, FLE, PHATE and Isomap were applied to the PCA coordinates, whereas the three PH-based methods received the raw 100-dimensional inputs. All method runs were completed for the 16 conditions, with evaluations completed for 14.

We performed additional analyses to assess the stability of the reported results, their sensitivity to methodological choices, and the reproducibility of the experimental comparisons.

Multi-seed evaluation.

The expanded results report five algorithmic seeds on fixed inputs, including individual seed values, means and sample standard deviations. On distorted three-loop data, TopoForce and DIPOLE obtain TRS scores of 0.9630 ± 0.0017 and 0.7962 ± 0.1046, respectively. On nonuniform bracelet data, TopoForce and RTD-AE obtain 0.8905 ± 0.0006 and 0.8889 ± 0.0020, respectively. These results show a substantial and consistent advantage on the distorted three-loop benchmark, whereas performance is closely comparable on the nonuniform bracelet (see Rebuttal Figure R3, pp. 2–3 hyperlink). Method coverage is recorded for each condition. See Updated Table 1 for the Bracelet and Three loops scores.

Updated Table 1. Topology recovery score (TRS, Eq. 9) for Bracelet and Three loops under uniform (U), distorted uniform (UD), and nonuniform (NU) sampling. Each cell gives the mean ± sample SD over optimizer seeds 0–4 on the same input. Within each dataset and setting, bold values and † mark the highest and second-highest distinct means, respectively, across all 13 methods before display rounding.

MethodBracelet UBracelet UDBracelet NUThree loops UThree loops UDThree loops NU
TopoForce0.970 ± 0.0000.877 ± 0.0000.891 ± 0.0010.975 ± 0.0020.963 ± 0.0020.978 ± 0.002
UMAP0.944 ± 0.0050.836 ± 0.028†0.619 ± 0.1070.600 ± 0.1410.668 ± 0.1060.609 ± 0.219
t-SNE0.947 ± 0.000†0.445 ± 0.0000.465 ± 0.0000.548 ± 0.0000.454 ± 0.0000.413 ± 0.000
FLE0.509 ± 0.1230.455 ± 0.1450.439 ± 0.1010.510 ± 0.1440.461 ± 0.0940.623 ± 0.131
PHATE0.754 ± 0.0480.567 ± 0.0800.551 ± 0.1190.504 ± 0.0330.463 ± 0.0340.474 ± 0.055
Isomap0.919 ± 0.0000.629 ± 0.0000.851 ± 0.0000.923 ± 0.0000.329 ± 0.0000.936 ± 0.000
DIPOLE0.626 ± 0.0630.643 ± 0.0330.730 ± 0.0170.947 ± 0.015†0.796 ± 0.105†0.957 ± 0.006†
TopoAE++0.868 ± 0.0350.600 ± 0.1590.738 ± 0.2300.845 ± 0.1070.658 ± 0.1330.903 ± 0.027
RTD-AE0.780 ± 0.2140.476 ± 0.2090.889 ± 0.002†0.911 ± 0.0010.732 ± 0.0490.830 ± 0.134
Diffusion Maps0.769 ± 0.0000.346 ± 0.0000.466 ± 0.0000.000 ± 0.0000.000 ± 0.0000.323 ± 0.000
Laplacian Eigenmaps0.650 ± 0.0000.329 ± 0.0000.436 ± 0.0000.032 ± 0.0000.000 ± 0.0000.327 ± 0.000
Stress majorization0.916 ± 0.0020.792 ± 0.0050.734 ± 0.0100.908 ± 0.0020.783 ± 0.0010.903 ± 0.006
tsNET0.307 ± 0.0430.265 ± 0.0690.208 ± 0.1000.440 ± 0.1430.389 ± 0.2100.407 ± 0.085

Hyperparameter sensitivity.

To assess whether the observed layouts depend on particular neighbourhood or interaction scales, we completed 256 parameter configurations for UMAP, t-SNE, Isomap, PHATE and Diffusion Maps on distorted and nonuniform bracelet and three-loop inputs. The scans cover neighbourhood size, UMAP minimum distance, t-SNE perplexity and diffusion time, using seed 0. The resulting layouts are available in Rebuttal Figure R2 hyperlink.

Component ablations.

To identify the contributions of individual design components, we report the ablation results on the nonuniform bracelet benchmark. The full model, removal of pairwise topology information, removal of density correction, and fixed-κ variant obtain TRS scores of approximately 0.8903, 0.5224, 0.8612 and 0.7918, respectively. Removing pairwise topology information produces the largest decrease, supporting its central role in this example (Figure 7 of the original submission).

AuthorsOpenReview ↗

C4–C5. Spectral and Layout Comparisons

C4. Contribution Beyond Conventional Spectral Embeddings

An important question is whether the improvements of TopoForce arise simply from using spectral information from the neighborhood graph, or whether the pair-specific interactions introduce an additional benefit.

Conventional spectral methods, such as Diffusion Maps and Laplacian Eigenmaps, construct embeddings from a shared set of spectral coordinates. In contrast, TopoForce quantifies the relative contributions of individual diffusion modes separately for each pair of observations, weights these contributions by their relaxation times, and translates the resulting pair-specific structural scales into adaptive repulsive interactions. Thus, rather than using spectral information solely to define embedding coordinates, TopoForce uses it to modulate how different pairs interact during optimisation.

To assess the contribution of this design, we added Diffusion Maps and Laplacian Eigenmaps to the comparison across all 54 conditions, using five seeds and the same observed inputs and output dimensions.

On distorted bracelet data, TopoForce obtains a TRS of 0.8774 ± 0.0003, compared with approximately 0.3464 and 0.3286 for the two spectral methods. On distorted three-loop data, the corresponding scores are 0.9630 ± 0.0017, 0 and 0 (see Updated Table 1 in C3). In contrast, on uniform circles, both spectral methods achieve a score of 1.000, compared with approximately 0.995 for TopoForce. See Rebuttal Figure R3 hyperlink for the five-seed comparison.

These comparisons suggest that the benefit of TopoForce is particularly pronounced when multiple topological structures must be resolved under geometric distortion, rather than representing a universal advantage over conventional spectral embeddings.

C5. Geometric Preservation and Alternative Graph Layouts

We further investigated whether the topological improvements obtained by TopoForce come at the expense of geometric information, and whether they could be explained by alternative graph-layout objectives.

Geometric information retained in the embedding.

We quantified local and global geometric relationships using additional label-based diagnostics on COIL-20. TopoForce achieves neighbor recall@2 of 0.622 ± 0.003 and purity@5 of 0.936 ± 0.002, compared with 0.710 ± 0.003 and 0.957 ± 0.001 for t-SNE. Together with the circular-association results, these measurements reveal a trade-off between preserving local neighborhood relationships and recovering the global circular organization: TopoForce better captures angular organization, whereas t-SNE better preserves local relationships.

Role of rank normalization.

The rank normalization in Eq. (5) expresses diffusion scales as relative interaction priorities, providing a consistent exponent range across datasets. Each endpoint contributes a bounded rank score, and averaging the two scores produces symmetric pairwise interactions.

This relative encoding intentionally prioritizes the ordering of pairwise structural scales rather than reconstructing their absolute magnitudes. Consequently, it discards absolute gaps between scales, meaning that near-tied scales can receive different exponents. The similar bracelet layouts therefore reflect the objective of exposing graph-supported structures through relative structural relationships, rather than preserving the original loop-size ratios.

Comparison with alternative graph-layout methods.

To determine whether the observed improvements could be explained by generic graph-layout objectives, we additionally compared TopoForce with stress majorization and tsNET across all 54 conditions, using five seeds.

On distorted bracelet data, TopoForce, stress majorization and tsNET obtain TRS scores of 0.8774 ± 0.0003, 0.7922 ± 0.0051 and 0.2646 ± 0.0689, respectively. On distorted three-loop data, the corresponding scores are 0.9630 ± 0.0017, 0.7827 ± 0.0005 and 0.3894 ± 0.2098 (see Updated Table 1 in C3). See Figure 3 hyperlink for the distorted-input embeddings and Rebuttal Figure R3 hyperlink for the five-seed comparison.

The two additional methods use the documented union-kNN graph protocol, with their treatment of disconnected inputs recorded.

AuthorsOpenReview ↗

C6. Real Data and Scalability

Real-world evaluation.

To provide a more direct assessment of whether the learned representations reflect independently supported biological and physical organisation, we evaluated absolute Fisher–Lee circular correlation against COIL image rotation and a measured-FUCCI-fluorescence angular reference, using eight methods and five seeds.

On the full COIL-20 dataset, macro-averaging over 20 objects gives a circular correlation of 0.4507 ± 0.0199 for TopoForce, compared with 0.3720 ± 0.0285 for t-SNE and 0.3268 ± 0.0296 for UMAP. On FUCCI, the raw-coordinate correlations are approximately 0.3865 for TopoForce, 0.5091 for PHATE and 0.5089 for Laplacian Eigenmaps.

These results provide a task-specific picture of performance: TopoForce captures circular organisation particularly well in COIL-20, but does not outperform all alternatives across real-world datasets. Object 3 in COIL-20 remains challenging, and the grid-cell experiment remains an unsuccessful torus-recovery case. See Figure 6 hyperlink for the COIL-20, FUCCI and grid-cell embeddings.

Computational scalability.

We further separated total fitting time from layout optimisation time to clarify the computational cost and the potential benefits of acceleration. Existing single-run measurements extend to 16,000 points, with a total fitting time of approximately 9,089 seconds. The 2,045.57-second measurement at 8,000 points also refers to total fitting time (Table 3 of the original submission).

In a controlled Barnes–Hut study on four clean inputs, we obtained 4.1–6.2× acceleration of the layout stage at a threshold of 0.5, using two seeds and a fixed 1,600-step budget. This timing includes tree construction and layout updates but excludes graph and spectral preprocessing. See Rebuttal Figure R1 hyperlink for the layout timings and embeddings.

These results demonstrate the potential to accelerate the layout stage, while the cost of graph and spectral preprocessing remains an important consideration for scaling to larger datasets.

Reviewer_z5vtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TopoForce, a dimensionality-reduction method aimed at revealing topological structures in high-dimensional point clouds. It constructs a nearest-neighbor graph and realizes a diffusion operator on it. For each pair of observations, it defines a "scale" as a weighted average of the relaxation times of the diffusion modes that distinguish the pair. These scales are then used to determine the exponent of a pair-dependent repulsive force in a force-directed graph layout. The method is evaluated on synthetic datasets with known topology and on three real-world applications. The authors report reasonable results, particularly on synthetic datasets containing multiple cycles at different scales.

Critical strengths

The paper is written in a clear way.The underlying idea is reasonable and interesting. The experimental evaluation covers a reasonably-broad range of (mostly synthetic) instances.

Critical weaknesses
  1. The method combines several choices which are all, in essence, heuristic (pairwise spectral scales, rank normalization, adaptive repulsion, coarse-to-fine optimization). A convincing explanation of why they should recover topology is completely missing. Comparisons with simpler spectral alternatives are also missing, and I find this the main weakness of the work. The trend of coming up with a seemingly-reasonable operator/architecture with not theoretical justification and no extensive comparison to the literature must, simply put, stop.

  2. Reiterating on the previous point, the experimental results should be stronger. Baselines use default parameters, results are based on single runs, and improvements on real data are mixed despite heavy computational costs.

Questions for authors
  • Can you justify the construction of the pairwise topology scales and their translation into repulsive forces, beyond the reported ablation results? Why should these choices be preferable to simpler alternatives?

  • How robust are the computational results across different random seeds? Would the conclusions change if the competing methods were (as they should have been) appropriately tuned?

4 comments
AuthorsOpenReview ↗

Response to Reviewer z5vt

Q1A: The method combines several choices that feel like heuristic methods without a convincing explanation why they would work. The trend of using reasonable methods without theoretical justification should be depreciated.

Response: We thank the reviewer for the frank criticism, and we deeply regret that we left the reviewer an unfortunate impression that TopoForce is a combination of heuristics. It is an oversight in our writing that we failed to fully disclose the idea and the theory behind TopoForce. We are grateful for having a chance to rectify our error in this response and we greatly appreciate the reviewer for the earnest criticism. We truly believe your comment helps us reflect on our writing and improve the rigor of our work. We would also like to take this opportunity to share that we totally agree with the reviewer that all methods should clearly state the reasoning behind their design choices and provide theoretical basis of their architecture. This should be a standard of the entire field and we salute to the reviewer for holding us to it.

Above all, we would like to take the opportunity to clarify that TopoForce is not a combination of a series of heuristic practices and there is a central theme that connects all of its design choices. We again apologize that our writing failed to clearly deliver this message and we hope that this response could rectify this mistake.

The premise: topological detectability remains coupled to embedding geometry.

Although ground-truth topology is invariant under continuous deformation, its detection from finite point clouds is inherently dependent on the geometry of their representation. Persistent homology (PH), for example, identifies topological features by tracking the emergence and disappearance of connected components, loops and higher-dimensional structures across a filtration induced by pairwise distances. Its robustness to small perturbations does not, however, make it independent of the input geometry: substantial geometric distortions can alter the filtration and obscure the persistence of genuine topological features. Consequently, even when the underlying topology remains unchanged, it may become difficult to detect from a geometrically distorted representation.

This distinction motivates a different objective for representation learning. Rather than faithfully preserving the observed geometry, which may itself obscure the underlying structure, we seek to reorganise the representation so that latent topological relationships become more clearly expressed and detectable. In this sense, geometric reorganisation is not merely a means of improving visualisation, but a means of facilitating more reliable topological analysis. This is the central premise of TopoForce: to learn a topology-revealing representation in which multiscale topological structures can be more readily identified by subsequent topological analysis.

The intuition: Force-directed Layout Embedding (FLE) helps derive an ideal geometrical layout with bespoke force-to-distance mapping.

The idea of TopoForce is inspired by natural phenomena in math and physics. In physics, we observe that with attraction and repulsion, objects and particles naturally form ideal geometries such as circles and spheres. For instance, the Asteroid Belt of the Solar system or the ring of the Saturn, naturally form circular shape with just gravity and centrifugal force. Alternatively neutron stars are spherical due to gravity and quantum degeneracy pressure. This is known as the Principle of Minimum Potential Energy and isotropic symmetry. Following this theory, when we apply an isotropic repellent force between every node pair on a single non-circular, uniform-density loop, restrained by the linkage between connecting nodes, the noose will naturally expand into a circular geometry. As a generalization of this setting, the linkage restraint could be replaced with an attractive force, which forms the basis of Force-directed Layout Embedding. For any loop of non-circular geometry, without density homogeneity, with attractions applied between local neighbors and isotropic repellent force applied everywhere, FLE would always arrive at ideal circular or spherical geometries that are friendly to Persistent Homology detection algorithms.

Challenge: Force interference between local and global topological structures.

The simple, rudimentary isotropic attractive and repellent force configuration, while effective at expanding a simple loose loop into a perfect circle, breaks down when the topology is complex, such as the beaded bracelet topology in Theory Figure 1 hyperlink. This is because when there are hierarchical topological structures, such as the global loop of all beads and the local points on each individual bead, the aggregated repulsive force from the global structure overwhelms the repulsive force received from local data points, as there are far more particles from the global scale than particles within the local structure. Fig. 1 shows three perfect geometries (2 beaded bracelets and 1 chained-circle). If we would like to maintain the geometry with the general FLE framework by designing a force configuration that maintains homeostasis and the stationary position of each particle, we would have to elevate the repellent force between particles in local circles while suppressing the repellent received from global structures. If we tie the repulsive force to distances, we can observe that short-distance particles would need a much greater repulsive force to maintain the general shape of the particle distribution.

However, the most ideal force configuration is strongly specific to each topology (mid column of Theory Figure 1 hyperlink). While the general trend of smaller local structures requiring stronger repulsive forces is shared across the three cases, their force-to-distance mappings are all different, even between the two beaded bracelet examples. If we exchange the force configurations between the examples, we observe that the particles will converge to new configurations where the perfect geometry is destroyed.

Inspiration: Hierarchical neighborhoods in graphs are natural divisions for local and global topological structures.

Diffusion Maps approximate the diffusion distance by Euclidean distance in a truncated spectral embedding. With all modes, the squared distance is

\bigl(\psi_k(i)-\psi_k(j)\bigr)^2.$$ On the bracelet, modes 1 and 2 retain the global ring and compress the small loops. Modes 8 and 9 reveal finer loop structure and fold the global ring (see [**Theory Figure 6 hyperlink**](https://anonymous.4open.science/api/repo/topoforce_2026-34CE/file/rebuttal%20figure/theory6.pdf)). The spectrum thus carries information at both scales. TopoForce combines that information for each pair through $$S_{ij}(t)=2\,\frac{\int_t^\infty D_s^2(i,j)\,\mathrm{d}s}{D_t^2(i,j)}.$$ This quantity measures how long diffusion continues to distinguish the pair. TopoForce uses it to adapt the range of pairwise repulsion, allowing global and local structure to shape the same embedding. **Main insight: The geometry of the global topology is preserved while force configuration is tailored to local structures.** While the general trend of particles in local structures requiring much stronger repulsive forces than particles in global structures persists across particle distribution shapes, the exact ratio is case dependent and non-universal. We use a parameterized function family $$F^{\mathrm{rep}}(\widehat r) \propto \left(\frac{\kappa}{\widehat r}\right)^{1-\widehat g}$$ to reflect this, where $\kappa$ controls the contrast between the global and local repulsion strengths. A small $\kappa$ would still have a larger global-to-local repulsion ratio, while a large $\kappa$ would have a smaller ratio, which would lead to a stronger expansion of a local structure to a circle. Still, the perfect choice of $\kappa$ remains elusive. [**Theory Figure 2 hyperlink**](https://anonymous.4open.science/api/repo/topoforce_2026-34CE/file/rebuttal%20figure/theory2.pdf) shows the expansion outcome of a 10-bead bracelet under fixed $\kappa$. The particle positions are randomly initialized. When $\kappa$ is small, the expansion is biased towards the global structure with squashed local loops. The local circles expand as $\kappa$ increases. But when $\kappa$ becomes too large, the global structure (the outer circle) collapses. However, the global structure collapse only occurs when particles are randomly initialized. When particles are initialized to a local-circle-quashed but global circle maintained shape, i.e., the $\kappa = 1$ shape, then a large $\kappa$ would not collapse the global circle but rather it focuses only on expanding the local circle. This insight forms the corner stone for TopoForce, which prioritizes the expansion of the global structure with a small $\kappa$ before progressing to local structures with large $\kappa$.

κ\kappa progression matters: Not getting trapped in local minima

A central reason that a large fixed κ\kappa would destroy a global structure is not necessarily that the local repellent force would overwhelm the global repellent force. It is simply that the global repellent force becomes too insignificant such that the total energy of the system becomes irresponsive to global structure shifts, to the point that opposing local circles could be placed in close vicinities in the representation space yet the total energy of the system remains small, even though there are still non-zero weak repulsion forces. However, the repulsion between distant particles in the global structure is too weak that any updates would result in an increase in system energy before it reaches the global minima.

To counter this phenomenon, TopoForce borrows the idea from Simulated Annealing, where the optimization algorithm starts at an initial solution with a high virtual "temperature", to allow it to escape from a local minimum. In the case of TopoForce, since a small κ\kappa would drive the particles to form a better global shape, and a larger κ\kappa is not inclined to disrupt the global shape due to weakened-but-not-vanished repulsion, the progressive optimization of TopoForce would gradually move from global to local topological structures, thus achieving the joint expansion of both global and local cycles. By leveraging the global structure persistent property of force expansion under large κ\kappa, the progressive learning scheme is not sensitive to the topology of the data and achieves good all-around performance in many topologies ranging from beaded bracelets to chained circles and beyond.

Q1B: a comparison against spectral methods is needed. Response: We apologize for the oversight. Indeed a comparison against simpler spectral methods is important. We have originally compared against spectral methods and due to their prioritization of global structures in top components, we assumed their defects are known and omitted in our original submission. We have since amended this mistaken. The results are supplied in the updated result figure. As shown in this figure, both DiffusionMap and Laplacian Eigenmaps would prioritize the global structure of the topology, which is expected as previously explained, and omit the representation of local loops.

AuthorsOpenReview ↗

Response to Reviewer z5vt Q2

Q2: Robustness across random seeds and baseline parameter sensitivity

Reviewer question: "How robust are the computational results across different random seeds? Would the conclusions change if the competing methods were (as they should have been) appropriately tuned?"

We thank the reviewer for raising these important concerns regarding experimental robustness and comparison fairness. We have substantially expanded the evaluation to examine both variability across random initializations and sensitivity to baseline hyperparameters.

Robustness across random seeds.

We acknowledge that the single-seed results in the original submission did not allow the variability arising from random initialization and optimisation to be assessed. We have therefore repeated the evaluation using five algorithmic seeds (0–4) across 20 synthetic conditions, four real-data conditions and 30 nuisance conditions, yielding 3,285 valid results across 13 methods and 54 conditions. All five repetitions within each condition use identical input observations and, where applicable, identical nuisance-noise realisations, thereby isolating variability due to algorithmic initialization and optimisation. We report individual seed values, means and sample standard deviations.

The results show that TopoForce is particularly consistent on the distorted three-loop benchmark, achieving a TRS of 0.9630±0.00170.9630 \pm 0.0017, compared with 0.7962±0.10460.7962 \pm 0.1046 for DIPOLE, the strongest competing method by mean TRS. On nonuniform bracelet data, TopoForce and RTD-AE achieve closely comparable scores of 0.8905±0.00060.8905 \pm 0.0006 and 0.8889±0.00200.8889 \pm 0.0020, respectively. Thus, the multi-seed evaluation confirms that the substantial advantage on distorted multiloop data is not attributable to a favourable random initialization, while also showing that improvements are not universal across all conditions (Multi-seeds figure hyperlink).

Baseline parameter sensitivity.

To examine whether the comparisons depend on the default hyperparameters used in the original submission, we additionally conducted parameter sensitivity scans for UMAP, t-SNE, Isomap, PHATE and Diffusion Maps on distorted and nonuniform bracelet and three-loop inputs. We completed 256 configurations in total, varying neighbourhood sizes (5, 15, 30 and 45), UMAP minimum distances (0–0.9), t-SNE perplexities (5–45), and diffusion times (0–100), using seed 0. The resulting embeddings provide a visual assessment of how the baseline representations change across these parameter ranges, rather than relying exclusively on their default configurations (Parameters tuning figure hyperlink).

Together, these additional analyses provide a broader assessment of the stability of TopoForce and the sensitivity of competing methods to parameter choices. While the baseline scans are not an exhaustive search for the optimal hyperparameters of every method, they allow us to assess whether the observed differences are specific to the default configurations used in the original comparison.

Comments on the paper

1 comment

Dear Reviewers,

As the discussion period draws to a close, please read and acknowledge the authors’ rebuttal within the next 24 hours. I also encourage you to engage in the discussion and share any remaining questions or concerns.

Thank you for your time and thoughtful reviews.

Best,

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.