acceptodds
Under review as a conference paper at ICLR 2027

COMET: Commute-Order Mapping of Expected Topology for Faithful Data Visualization

Abstract

Visualization methods such as t-SNE, UMAP and their successors are the standard tools for making high-dimensional data visually interpretable. However, they expose hyperparameters that substantially change the embedding, with no ground truth to guide the choice. The settings are often chosen by visual inspection instead, which may bias the result toward the user's prior expectations. The embeddings also tend to break continuous structure into pieces and to distort both scale and global geometry. We introduce COMET (Commute-Order Mapping of Expected Topology), which embeds the geometry of a neighborhood graph. Commute times on that graph supply distances that follow the manifold, and ranking places them on a common scale with the graph's local similarities. Every neighbor and sampled non-neighbor then receives a target probability in the cross-entropy objective that calibrates the repulsive terms alongside the attractive ones, rather than every non-neighbor sharing the same target probability as in UMAP. COMET scales to large datasets and requires no hyperparameter tuning. Against seven established methods on four synthetic datasets with known topology and four real datasets, it ranks first on of comparisons under three validation measures built to quantify geodesic distance preservation, fragmentation and scale distortion.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
dimensionality reduction, data visualization, manifold learning, commute time distance, neighbor embedding, spectral graph theory, hyperparameter-free methods, validation measures

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_i816OpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes COMET, a tool for high-dimensional data visualization in 2D or 3D. It builds upon existing methods like t-SNE and UMAP by avoiding "subjectively tunable" hyperparameters, separation of continuous structures into separate clusters, and scale/layout distortions. COMET's pipeline begins by building a nearest-neighbor graph, then uses a random walk commute time to record graph distances (approximated by eigenvectors of the graph Laplacian). Nearby neighborhoods are ranked by local similarity and "long-distance" similarity (from randomly sampled points). This ranking than provides target probabilities (whereas UMAP would assign a probability of 0 throughout). COMET fits this layout with a cross-entropy loss and kernel wifths and regularization are calculated from data. Runtime scales linearly with dataset size.

Critical strengths
  • COMET addresses very practical problems with t-SNE and UMAP, which are known to be over-interpreted in certain applications. Tuning until the plot "looks right" is obviously not a sufficient scientific approach, so I appreciate the author's proposed pipeline to address this. Moreover, the central idea of the pipeline is clear and well motivated. It uses the data shape to compute the data repulsion in lower dimensions. Ablation experiments are convincing that removing this modifications leads back to the usual separations and distortions.

  • The method is scalable and thus practical. Authors show it embeds 86k cells in about 20 seconds, with no need for subsampling (unlike PHATE).

Critical weaknesses
  • Weak metrics: First, the headline comparison metrics are proposed by the work itself. Without established measures of global structure, the main claim is weakened in its obvious tie to the scoring design... Authors use an established metric for trustworthiness, and t-SNE beats COMET on every dataset there, sometimes significantly (on Newsgroups, for example). As such, the results on global structure preservation would be significantly strengthened if there was some independent metric that authors could report. In particular, the proposed geodesic rank correlation ρGeo is closely related to DEMaP (Moon et al., 2019), which correlates geodesic distances in the data with distances in the embedding. Second, the quantitative topological metrics are really limited. For one, the only topological check (at least in the main paper) is whether an embedding stays in one piece. Loops (in COIL). and holes (Swiss roll) are qualitatively checked. It would be valuable to include quantiative checks for a paper with "expected topology" in the title... for example with persistent homology, at a minimum.

  • Hyperparameters involved: The claim that COMET avoids hand-tuned hyperparameters seems overstated. COMET uses fixed parameters like neighborhood size, sampled size, spectral dimension, and so on. Sensitivity analyses show that at least one of them noticeably changes results. So it would be important to adjust claims like COMET "removes the analyst's choices". General

Questions for authors

Clarifications

Were the three proposed measures and their settings (e.g., the floor of 5 in N_comp) fixed before COMET was developed? Appendix A reports results at floors of 2 and 15, which suggests the alternatives were seen before the choice was made.

Which topological properties beyond connectivity does “expected topology” refer to, and does anything in the objective explicitly preserve them?

Were COMET’s fixed defaults (k, s, l = s, κ_t, the fallback threshold) chosen or adjusted using the evaluation datasets?

Suggested future work (satisfying these suggestions would convince me to change my score)

Report at least one established global-structure measure that the authors did not design (e.g., DEMaP, see comment above in Weaknesses).

Quantify topology preservation with some topological metric (like persistent homology), at least on the Swiss roll and COIL-20.

Report sensitivity to k, and subsequently adapt the “no hyperparameter tuning” claims accordingly.

6 comments
AuthorsOpenReview ↗

We thank the reviewer for their constructive feedback

Comment: Weak metrics: First, the headline comparison metrics are proposed by the work itself. Without established measures of global structure, the main claim is weakened in its obvious tie to the scoring design... Authors use an established metric for trustworthiness, and t-SNE beats COMET on every dataset there, sometimes significantly (on Newsgroups, for example). As such, the results on global structure preservation would be significantly strengthened if there was some independent metric that authors could report. In particular, the proposed geodesic rank correlation ρGeo is closely related to DEMaP (Moon et al., 2019), which correlates geodesic distances in the data with distances in the embedding. Second, the quantitative topological metrics are really limited. For one, the only topological check (at least in the main paper) is whether an embedding stays in one piece. Loops (in COIL). and holes (Swiss roll) are qualitatively checked. It would be valuable to include quantiative checks for a paper with "expected topology" in the title... for example with persistent homology, at a minimum.

Answer: We address both points in our answers to the first two requests below.


Comment: Hyperparameters involved: The claim that COMET avoids hand-tuned hyperparameters seems overstated. COMET uses fixed parameters like neighborhood size, sampled size, spectral dimension, and so on. Sensitivity analyses show that at least one of them noticeably changes results. So it would be important to adjust claims like COMET "removes the analyst's choices".

Answer: We agree with all three reviewers that the claim of requiring no hyperparameter tuning was overstated. Therefore, we have revised the wording throughout the manuscript (Abstract, Sections 1, 3.1 and 4, Appendix B.5) to soften the claim. In particular, it should now be clear that while COMET aims to reduce the need for per-dataset tuning it does not eliminate all tunable hyperparameters, and changing the hyperparameters can still affect the results.


Request: Report at least one established global-structure measure that the authors did not design (e.g., DEMaP, see comment above in Weaknesses).

Answer: Thanks for the suggestion. We now report DEMaP for every method and dataset in Appendix B.6 (Table 13). DEMaP and ρGeo\rho_{\mathrm{Geo}} differ only in the LD distances: DEMaP and ρGeo\rho_{\mathrm{Geo}} compare the HD shortest-path distances with the Euclidean distances and the shortest-path distances in the embedding, respectively. We keep ρGeo\rho_{\mathrm{Geo}} as the primary measure because DEMaP penalizes an embedding for reproducing a hole present in the data. For example, the ground-truth embedding of the Swiss roll with a hole scores 1.000 on ρGeo\rho_{\mathrm{Geo}} but 0.991 on DEMaP (Figure 7). For the real datasets both measures select the same best method (Tables 2 and 13). On the synthetic datasets (Tables 1 and 13). they agree on the flat disc and the Swiss roll, while on the hierarchical Gaussians and the tree COMET ranks first on ρGeo\rho_{\mathrm{Geo}} but third on DEMaP, behind SASNE and TriMAP.


Request: Quantify topology preservation with some topological metric (like persistent homology), at least on the Swiss roll and COIL-20.

Answer: Thanks for the suggestion. We added the hole detection score of Damrich et al. (2024), a persistent-homology measure of how clearly a loop stands out in the embedding, to Section 3.2. Appendix A.1 defines and reports the hole detection score for the Swiss roll and COIL-20 (Table 3). Table 3 shows that for the Swiss roll, COMET achieves the highest score by a wide margin. For COIL-20, where each object’s rotation forms a loop, PaCMAP and TriMAP separate these loops most clearly.


Request: Report sensitivity to k, and subsequently adapt the “no hyperparameter tuning” claims accordingly.

Answer: We added a sensitivity analysis of k in Appendix B.4. Table 11 varies k from 5 to 50 on the Swiss roll and the hierarchical Gaussians and reports the runtime on the C. elegans atlas dataset. From k=10 to k=30, ρGeo\rho_{\mathrm{Geo}} stays between 0.975 and 0.986 on the Swiss roll and between 0.933 and 0.980 on the Gaussians, and the default k=15 is within 0.006 of the best k on both. The largest change occurs at k=5 on the Gaussians, where the 5-NN graph is disconnected and ρGeo\rho_{\mathrm{Geo}} falls to 0.884. Repeating the analysis for UMAP (Figure 6), COMET’s results vary less across k than UMAP’s on four of the five measures. The “no hyperparameter tuning” claims have been adapted accordingly (Abstract, Sections 1, 3.1 and 4, Appendix B.5).

AuthorsOpenReview ↗

We thank the reviewer for their constructive feedback

Question: Which topological properties beyond connectivity does “expected topology” refer to, and does anything in the objective explicitly preserve them?

Answer: The name refers to how the objective’s targets are constructed. “Expected” refers to the expected ranks (Equation 4) that assign target probabilities to the sampled non-neighbors. “Topology” refers to the k-nearest-neighbor graph from which those targets are derived. The graph encodes topological structure such as holes indirectly: paths between points on opposite sides of a hole must go around it, which lengthens the commute times between them. We want to clarify that no term in the objective preserves such structure explicitly. It is preserved only implicitly, through the commute-time targets.


Question: Were COMET’s fixed defaults (k, s, l = s, κ_t, the fallback threshold) chosen or adjusted using the evaluation datasets?

Answer: kk and ss were both fixed at 15 before evaluation, and l=sl = s was chosen based on the motivation given in the final paragraph of Section 2.3. κtκ_t and the fallback threshold are numerical-stability safeguards, set during the development of COMET on a wide range of datasets that included some of the evaluation datasets.

Thank you for the detailed response and additional experiments. The DEMaP results, persistent-homology evaluation, and sensitivity analysis address my main requests. I also appreciate the changes to the hyperparameter claims and clarifications about topology preservation.

I still have a few unanswered questions:

  • How were the evaluation measures and their settings chosen, particularly the component-count floor of 5? Were these adjusted after inspecting comparative results, and do the alternative floors change the conclusions?
  • For the numerical safeguards developed using some evaluation datasets, could you clarify whether the choices were based only on numerical stability, or also on the quality of the resulting embeddings?

Also, I would appreciate if the main text could better present the new clearer picture of where COMET performs best versus other methods, particularly the DEMaP results on trees and hierarchical Gaussians, the loop-detection results on COIL-20, and the lower trustworthiness scores.

Reviewer_Yj3YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This manuscript proposes COMET, a manifold-aware dimensionality-reduction method that combines a locally adaptive k-NN kernel with global commute-time distances approximated by a truncated spectral decomposition of the graph Laplacian. It ranks local similarities and global commute-time distances on a common scale and optimizes a rank-calibrated cross-entropy objective. This manuscript reports strong empirical results on four synthetic and four real datasets against seven baselines.

Critical strengths
  1. The method is conceptually well motivated: commute-time distances provide a principled manifold-aware global geometry, while the rank-based formulation elegantly places local and global quantities on a comparable scale.

  2. The data-derived kernel scale and calibrated attractive/repulsive terms are technically sound and reduce reliance on ad hoc tuning.

  3. The empirical results are strong and consistent: COMET ranks first on 21 of 24 comparisons and attains the lowest scale distortion on every real dataset.

Critical weaknesses
  1. Scalability evidence is limited. The largest experiment is n=86,024; there is little stress testing on much larger n and higher ambient dimension. Additional scalability experiments, including memory consumption and end-to-end preprocessing cost, would strengthen the claims.

  2. Related work is insufficient. The paper should more comprehensively review global-structure-preserving projection methods, such as Isomap, and explicitly position COMET relative to them. Comparative experiments on their ability to preserve global structure should also be added.

  3. The suitability of Ncomp for datasets containing multiple manifolds deserves clarification. For COIL-20 and the hierarchical Gaussians, the clear cluster structure may reflect intrinsically separate manifolds, so Ncomp could conflate faithful cluster separation with manifold fragmentation. In such cases, computing Ncomp separately within each cluster may be preferable; otherwise, the connectivity preservation of the embedding could be underestimated.

  4. The claim of requiring no hyperparameter tuning is overstated. Although several key hyperparameters are data-derived, others—such as the neighborhood size for k-NN graph construction and the number of sampled non-neighbors—still cannot be set automatically and can substantially affect the results.

  5. The ablation comparing spectral and PCA initialization is insufficient to isolate the effect of the objective. Before attributing the final embedding quality to COMET's commute-ordering objective, the authors should report geodesic rank correlation, Ncomp, and Dscale for the raw PCA embedding and the raw spectral (commute-time) embedding themselves. Otherwise, if PCA already performs well on these global measures, the ablation cannot distinguish the contribution of the objective from that of the initialization, and the conclusion is confounded. Adding these initialization-only baselines and comparing them with the optimized embeddings would make the ablation substantially more convincing.

Questions for authors

No Questions

3 comments
AuthorsOpenReview ↗

We thank the reviewer for their constructive feedback

Comment 1: Scalability evidence is limited. The largest experiment is n=86,024; there is little stress testing on much larger n and higher ambient dimension. Additional scalability experiments, including memory consumption and end-to-end preprocessing cost, would strengthen the claims.

Answer: Appendix B.8 now reports COMET on EMNIST ByClass: 814,241814{,}241 images, around 10 times more points than the atlas, in d=784d = 784 rather than d=100d = 100 dimensions. Table 14 reports the wall-clock time of each stage of the fit, on nested subsamples from 16,00016{,}000 points to the full set, together with the peak memory footprint. COMET embeds the full set in 282 s with a peak memory footprint of 12.4 GB. Its runtime grows as n0.94n^{0.94}, and memory grows linearly at about 15 KB per point.


Comment 2: Related work is insufficient. The paper should more comprehensively review global-structure-preserving projection methods, such as Isomap, and explicitly position COMET relative to them. Comparative experiments on their ability to preserve global structure should also be added.

Answer: We added a review of PCA, Isomap and spectral methods to the introduction (Section 1, paragraph 2), together with a detailed comparison of COMET with PCA and Isomap in Appendix B.9 using the five validation measures. In short, COMET's geodesic rank correlation is the highest or within 0.016 of the best classical method on seven of the eight datasets, and PCA has the lowest scale distortion on six of the eight, as expected of a linear projection. COMET has the highest or tied-highest trustworthiness and continuity on every dataset except the flat disc, which PCA recovers exactly. Isomap cannot run on the atlas dataset due to memory limits.


Comment 3: The suitability of Ncomp for datasets containing multiple manifolds deserves clarification. For COIL-20 and the hierarchical Gaussians, the clear cluster structure may reflect intrinsically separate manifolds, so Ncomp could conflate faithful cluster separation with manifold fragmentation. In such cases, computing Ncomp separately within each cluster may be preferable; otherwise, the connectivity preservation of the embedding could be underestimated.

Answer: We thank the reviewer for raising this point, and have added a clarifying sentence to Section 3.2 (Manifold Fragmentation). NcompN_{\text{comp}} does not measure connectivity in absolute terms, but relative to the high-dimensional data. It is computed at k⋆k^\star, the smallest neighborhood size at which the high-dimensional kk-NN graph is connected (with a floor of 5), so k⋆k^\star is already large enough to join any separate clusters in the data. In COIL-20, for example, the well-separated objects connect only at k0=52k_0 = 52, and the embedding is evaluated at the same neighborhood size. NcompN_{\text{comp}} therefore equals 1 whenever the embedding is connected at the same neighborhood size as the data, and exceeds 1 only when the embedding separates its pieces more than the data does. Faithful cluster separation is not penalized; only separation beyond that present in the data is.

AuthorsOpenReview ↗

We thank the reviewer for their constructive feedback

Comment 4: The claim of requiring no hyperparameter tuning is overstated. Although several key hyperparameters are data-derived, others—such as the neighborhood size for k-NN graph construction and the number of sampled non-neighbors—still cannot be set automatically and can substantially affect the results.

Answer: We agree with all three reviewers that the claim of requiring no hyperparameter tuning was overstated. Therefore, we have revised the wording throughout the manuscript (Abstract, Sections 1, 3.1 and 4, Appendix B.5) to soften the claim. In particular, it should now be clear that while COMET aims to reduce the need for per-dataset tuning it does not eliminate all tunable hyperparameters, and changing the hyperparameters can still affect the results.


Comment 5: The ablation comparing spectral and PCA initialization is insufficient to isolate the effect of the objective. Before attributing the final embedding quality to COMET's commute-ordering objective, the authors should report geodesic rank correlation, Ncomp, and Dscale for the raw PCA embedding and the raw spectral (commute-time) embedding themselves. Otherwise, if PCA already performs well on these global measures, the ablation cannot distinguish the contribution of the objective from that of the initialization, and the conclusion is confounded. Adding these initialization-only baselines and comparing them with the optimized embeddings would make the ablation substantially more convincing.

Answer: The reviewer suggests reporting ρGeo\rho_{\mathrm{Geo}}, NcompN_{\text{comp}} and DscaleD_{\text{scale}} for the raw PCA and commute-time embeddings, so that the contribution of COMET's objective can be separated from that of the initialization. We thank the reviewer for this suggestion, but believe the current ablation already makes this separation. The ablation in Appendix B.3 (Figure 5, Table 10) is a 2×22 \times 2 factorial design: each objective, COMET's objective (Equation 6) and the same objective with every non-neighbor target probability set to zero, is run from each initialization, spectral and PCA. The comparison that isolates the objective is therefore between two configurations that start from the identical layout and differ only in their target probabilities. Because the initialization is held fixed within each pair, the differences between them cannot be attributed to it, however well the initialization itself scores.

Reviewer_Yj3YOpenReview ↗

Response to the rebuttal

Thank you for the response. Most of my concerns have been addressed. However, the revision weakens the claimed contribution — that COMET requires no hyperparameter tuning — without introducing other important contributions to compensate. I therefore keep my score.

Reviewer_kSejOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper advances graph-based visualization by combining local neighborhood similarities with global ranks derived from a truncated, regularized commute-time embedding. Neighbor ranks and estimated non-neighbor ranks determine target probabilities in a binary cross-entropy objective and provide differentiated global interactions. Experiments on synthetic and real datasets show promising results of the proposed method. However, although COMET performs well with a single fixed configuration, its claim of parameter-freeness is overstated, as consequential hyperparameters were fixed rather than eliminated.

Critical strengths
  • the intuition of combining local neighborhood similarities with global ranks makes sense.
  • strong performance across synthetic and real datasets, with solid ablation analyses.
Critical weaknesses
  • While strong performance is demonstrated on proposed measures, it is not clear whether the method improves general visualization quality/usefulness.
  • The parameter-freeness claim is overstated. There are still multiple hyperparamesters that remain fixed, and sensitivity results show some of them actually affect embedding quality. Neighborhood size, for example, is likely to be more sensitive in COMET (than in UMAP and others) as it controls both the graph used to compute global commute-time distances and the boundary between the two different similarity rules between neighbors & non-neighbors. Other potential hyperparameters (even though they are fixed in experiments) that COMET introduces include the spectral dimension, spectral regularization threshold, the minimum-distance bridges, and the rate of decay (σ\sigma), etc.
Questions for authors
  • Can you vary neighborhood size (k), regularization threshold, and rate of decay (σ\sigma)?
  • Why should the spectral dimension l=s preserve the relevant distance ordering? The fact that s difference vectors span at most s dimensions does not imply that the first s Laplacian eigenvectors capture that span across anchors.

The authors' response has addressed my questions above.

2 comments
AuthorsOpenReview ↗

We thank the reviewer for their constructive feedback.

Comment: While strong performance is demonstrated on proposed measures, it is not clear whether the method improves general visualization quality/usefulness.

Answer: We thank the reviewer for this point. Since visualization quality/usefulness has no single ground truth, the paper evaluates several complementary measures, including the established trustworthiness and continuity (Tables 4–5). To strengthen this evaluation, the revision adds two further established measures that we did not design: DEMaP (Moon et al., 2019), which measures geodesic distance preservation (Section 3.2, Loop Preservation; Swiss roll paragraph of Section 3.3; COIL-20 paragraph of Section 3.4; Appendix A.1, Table 3), and the persistent-homology hole detection score of Damrich et al. (2024), which measures loop preservation (Section 3.2, Loop Preservation; Swiss roll paragraph of Section 3.3; COIL-20 paragraph of Section 3.4; Appendix A.1, Table 3).


Comment: The parameter-freeness claim is overstated. There are still multiple hyperparamesters that remain fixed, and sensitivity results show some of them actually affect embedding quality. Neighborhood size, for example, is likely to be more sensitive in COMET (than in UMAP and others) as it controls both the graph used to compute global commute-time distances and the boundary between the two different similarity rules between neighbors & non-neighbors. Other potential hyperparameters (even though they are fixed in experiments) that COMET introduces include the spectral dimension, spectral regularization threshold, the minimum-distance bridges, and the rate of decay ( ), etc.

Answer: We agree with all three reviewers that the claim of requiring no hyperparameter tuning was overstated. Therefore, we have revised the wording throughout the manuscript (Abstract, Sections 1, 3.1 and 4, Appendix B.5) to soften the claim. In particular, it should now be clear that while COMET aims to reduce the need for per-dataset tuning it does not eliminate all tunable hyperparameters, and changing the hyperparameters can still affect the results.


Question: Can you vary neighborhood size (k), regularization threshold, and rate of decay?

Answer: All three can be varied, but the defaults are intended to work across datasets without per-dataset tuning. For kk, we have added a sensitivity analysis (Appendix B.4), described in our answer on neighborhood size below. The decay rate of the rational kernel is not a separate parameter: it is set to r1=k+(n−k)/(s+1)r_1 = k + (n-k)/(s+1), the expected rank of the first non-neighbor, so it follows from kk and ss. The regularization term ϵ\epsilon bounds the condition number of LsymL_{sym} by κt\kappa_t, and κt=5000\kappa_t = 5000 is chosen large enough that ϵ\epsilon is zero unless the kk-NN graph is close to disconnected. So κt\kappa_t only acts as a numerical safeguard in this case rather than as a modelling parameter.


Question: Why should the spectral dimension l=s preserve the relevant distance ordering? The fact that s difference vectors span at most s dimensions does not imply that the first s Laplacian eigenvectors capture that span across anchors.

Answer: We thank the reviewer for this question and agree. The dimension count shows only that ss dimensions could represent the ss difference vectors of a single anchor, not that the first ss Laplacian eigenvectors span them, since that subspace differs across anchors. We have revised Section 2.3 to state this explicitly and to present l=sl = s as a heuristic rather than a derivation.


Question: Neighborhood size, for example, is likely to be more sensitive in COMET (than in UMAP and others) as it controls both the graph used to compute global commute-time distances and the boundary between the two different similarity rules between neighbors & non-neighbors.

Answer: This is an interesting hypothesis, so we tested it directly. Appendix B.4 varies kk from 5 to 50 on the Swiss roll and the hierarchical Gaussians for both COMET and UMAP (Table 11, Figure 6), and Section 3.1 now summarizes the results. We found the opposite effect in these tests: COMET's results vary less across kk than UMAP's on four of the five validation measures, with the largest change at k=5k=5, where the graph of the hierarchical Gaussians is disconnected. Section 4 now states that kk affects the results and that its default was validated only on the datasets studied.

Thanks for the author's effort into the rebuttal. It resolved most of my concerns, and I'd be happy to raise my score.

Comments on the paper

1 comment

Summary of revisions

Each item is tagged with the reviewers whose comments it addresses.

New results

  • [Yj3Y] Added a comparison of COMET with the classical methods PCA and Isomap on all eight datasets, reporting the five validation measures and fit times (Appendix B.9, Tables 15–16; summarized in Section 4, paragraph 1).
  • [i816] Added DEMaP as a complementary measure of geodesic distance preservation for every method and dataset (Appendix B.6, Table 13). Section 3.2, paragraph 3, explains how it relates to ρGeo\rho_{\mathrm{Geo}}, and the results are discussed in the final paragraph of Section 3.3 and the second-to-last paragraph of Section 3.4.
  • [i816] Added the hole detection score (Damrich et al., 2024), a persistent-homology measure of how clearly the hole in the Swiss roll and the rotation loops in COIL-20 are preserved (Section 3.2, Loop Preservation; Appendix A.1, Table 3). The results are discussed in the Swiss roll paragraph of Section 3.3 and the COIL-20 paragraph of Section 3.4.
  • [kSej, i816] Added a sensitivity analysis to the neighborhood size kk from 5 to 50, repeated for UMAP for comparison (Appendix B.4, Table 11, Figure 6; summarized in Section 3.1, paragraph 2).
  • [Yj3Y] Added a scalability experiment on EMNIST (n=814,241n = 814{,}241, d=784d = 784), reporting the runtime of each stage and the peak memory up to the full dataset (Appendix B.8, Table 14).
  • [i816] Added an analysis of how the floor on the neighborhood size used by NcompN_{\text{comp}} affects the component counts, at floors from none to 15 (Appendix A, Figure 4; referenced in Section 3.2, Manifold Fragmentation).

Revised claims and clarifications

  • [kSej, Yj3Y, i816] The claim of requiring no hyperparameter tuning is softened throughout (Abstract, Sections 1, 3.1 and 4, Appendix B.5), and a new limitation in Section 4, paragraph 3, states that kk and ss affect the results.
  • [i816] The headline result is now 22 of 26 comparisons, up from 21 of 24, because the hole detection score was added.
  • [Yj3Y] Section 2.6 now states that the spectral embedding becomes superlinear when the spectral gap narrows with nn.
  • [kSej] The setting l=sl = s is now presented as a heuristic, with its limits stated explicitly (Section 2.3, final paragraph).
  • [Yj3Y] NcompN_{\text{comp}} is clarified as measuring connectivity relative to the high-dimensional data (Section 3.2, Manifold Fragmentation).

Related work

  • [Yj3Y] Section 1, paragraph 2, now reviews classical distance-preserving methods (PCA, Isomap, spectral methods) and positions COMET relative to them.

We also made minor edits for clarity and consistency throughout.

Response to all reviewers

We thank the reviewers for the constructive feedback.

Comment: The claim of requiring no hyperparameter tuning is overstated.

Answer: We agree with all three reviewers that the claim of requiring no hyperparameter tuning was overstated. Therefore, we have revised the wording throughout the manuscript (Abstract, Sections 1, 3.1 and 4, Appendix B.5) to soften the claim. In particular, it should now be clear that while COMET aims to reduce the need for per-dataset tuning it does not eliminate all tunable hyperparameters, and changing the hyperparameters can still affect the results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.