SNEkhorn: Dimension Reduction with Symmetric Entropic Affinities

Many approaches in machine learning rely on a weighted graph to encode the similarities between samples in a dataset. Entropic affinities (EAs), which are notably used in the popular Dimensionality Reduction (DR) algorithm t-SNE, are particular instances of such graphs. To ensure robustness to heterogeneous sampling densities, EAs assign a kernel bandwidth parameter to every sample in such a way that the entropy of each row in the affinity matrix is kept constant at a specific value, whose exponential is known as perplexity. EAs are inherently asymmetric and row-wise stochastic, but they are used in DR approaches after undergoing heuristic symmetrization methods that violate both the row-wise constant entropy and stochasticity properties. In this work, we uncover a novel characterization of EA as an optimal transport problem, allowing a natural symmetrization that can be computed efficiently using dual ascent. The corresponding novel affinity matrix derives advantages from symmetric doubly stochastic normalization in terms of clustering performance, while also effectively controlling the entropy of each row thus making it particularly robust to varying noise levels. Following, we present a new DR algorithm, SNEkhorn, that leverages this new affinity matrix. We show its clear superiority to state-of-the-art approaches with several indicators on both synthetic and real-world datasets.

Paper

Similar papers

Peer review

Reviewer 8Btp6/10 · confidence 3/52023-07-06

Summary

This paper presents a new dimensionality reduction algorithm, named SNEkhorn. By uncovering the novel links between Entropic affinities (EAs) and Optimal Transport (OT), the authors derive EAs with symmetric doubly stochastic normalization and the fixed row-wise entropy, which is the key to SNEkhorn. Besides, the authors show a dual ascent algorithm to compute this new affinity efficiently. Beyond this theoretical contribution, the authors showcase the benefits of SNEkhorn through numerical experiments on simulated data and real data.

Strengths

- Entropic affinities (EAs) needs to be symmetrized when used in popular Dimensionality Reduction (DR) algorithm t-SNE, but the symmetrization destroys the row-wise constant entropy and stochasticity properties of EAs. To derive EAs that can maintain both symmetry and its own properties is a very natural. The proposed method is novel and simple to describe. - The authors discover novel connections between EAs and OT, and show a dual ascent algorithm to compute EAs from OT. The proposed algorithm provides a new perspective on EAs. - For conventional DR algorithm, the latent affinity is not doubly stochastic, which imposes spherical constraints on the embedding space. The proposed methods avoid this problem. - The simulation study and real data analysis are very sufficient. The authors illustrate the effectiveness from many aspects.

Weaknesses

- There is not any discussion about the computational complexity of proposed method. Although the proposed algorithm performs better than t-SNE and UMAP, how does the computational complexity of proposed method compare with t-SNE or UMAP. Since the computational cost is also important in practice. - Some of the formulas and references in the article are incorrectly hyperlinked.

Questions

- How does the computational complexity for the proposed method compare to t-SNE? Does the proposed method spend the same amount of time as t-SNE while achieve better results? - How to choose the dimensionality of PCA? In the reference the authors mentioned, the dimensionality of PCA is usually set to be 30 instead of 50. Does the dimensionality affects the performance of the algorithm?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer W2ip8/10 · confidence 5/52023-07-07

Summary

This is a very interesting paper about an application of the Sinkhorn algorithm to symmetrize the matrix of entropic affinities in methods of DR like SNE, t-SNE, etc. Strong theoretical contribution. Some experiments to illustrate.

Strengths

The paper is very interesting for its vision, state of the art across multiple domains, strong theoretical contribution, revisiting of existing methods and proposal of a new method. The paper is very well written, not rushed. The symmetrization with Sinkhorn is well done and this a nice "tour de force" like only Brittons can do. The experiments include a grid search on the perplexity, and multiple runs for each perplexity.

Weaknesses

The paper is well written but it is also very/too dense: much information is packed, that is nice, but it takes quite some time to digest and to decode. The notation is sometimes unconventional and difficult to follow, probably to keep it as compact as possible. While the theoretical part is very convincing and interesting for itself, it will only appeal to a very restricted audience of people specialised in the methodological design of advanced methods of DR. Also, the experiments are not sufficient to convince practitioners that the Sinkhorn symmetrization is from now on something necessary to get the best results. The experiments lack broader comparisons with other methods that are in the state of the art (here it looks like the authors have reimplemented t-SNE and UMAP; also, there are methods that do not look for symmetry at all, although this can be detrimental in the presence of many isolated outliers). The quantitative assessment might be partly questionable: the trustworthiness is not the best DR QA indicator; what was the neighbourhood size, by the way? The lack of accelerated version is a possible weakness for dissemination, while NeurIPS would be precisely the ideal launching pad.

Questions

The "crowding effect" is just an intuitive explanation found in LvdM t-SNE paper; would it not be better to refer to something more conventional (empty space phenomenon, norm concentration, or just the reduction in volume from HD to LD?)? In Figure 1, SNEkhorn or t-SNEkhorn ?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

No possible negative societal impact here.

Reviewer s51e7/10 · confidence 3/52023-07-27

Summary

The paper presents a novel approach to dealing with entropic affinities (EAs) used in machine learning for dimensionality reduction tasks, specifically in the popular t-SNE algorithm. It addresses the limitations of current symmetrization methods applied to EAs, which can compromise the entropy and stochasticity properties of the affinity matrix. The proposed method uses optimal transport to achieve a natural symmetrization, leading to a new affinity matrix. This new matrix is then leveraged in a new dimensionality reduction algorithm called SNEkhorn, which is demonstrated to outperform state-of-the-art methods on both synthetic and real-world datasets.

Strengths

1. The paper introduces a novel approach to symmetrizing entropic affinities by formulating them as an optimal transport problem. This brings a fresh perspective to the problem and could open up new avenues for research and application in machine learning. 2. The proposed symmetrization method claims to maintain the constant entropy and stochasticity properties of the affinity matrix while being computationally efficient through dual ascent. This combination of robustness and efficiency is essential for practical applications. 3. The development of the SNEkhorn algorithm based on the new affinity matrix offers a practical implementation of the proposed approach. The demonstrated superiority of SNEkhorn over other baseline methods on various datasets further strengthens the paper's claims. 4. The paper evaluates the proposed SNEkhorn algorithm on both synthetic and real-world datasets, providing a comprehensive assessment of its performance and general applicability.

Weaknesses

Clarity of Presentation: While the abstract provides a high-level overview, some concepts, such as entropic affinities, optimal transport, and dual ascent, might be challenging for readers not already familiar with the domain. The paper should provide a clear and concise introduction to these concepts, ideally with intuitive explanations and illustrative examples. For example, what is P and C on paper 3 "Symmetric Entropy-Constrained Optimal Transport" Robustness Analysis: Since the proposed method claims to be robust to varying noise levels, it would be valuable to include a thorough analysis of its performance under different levels of noise in the datasets instead of two level. This could strengthen the claim of robustness and highlight the algorithm's practicality in real-world scenarios.

Questions

1. What sensitivity analysis was performed on the hyperparameters introduced by the new symmetrization approach? How robust is the SNEkhorn algorithm to changes in these parameters? 2. Can you provide additional insights into the interpretability of the reduced representations obtained by SNEkhorn? How well does it preserve meaningful structure in the data compared to other methods? Is the visualization in Figure 6 a special case? 3.Could you elaborate on the limitations or potential failure cases of the proposed approach? What are the scenarios where SNEkhorn might not perform as well or could face challenges?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

While the paper proposes a novel approach to symmetrizing entropic affinities and introduces the SNEkhorn algorithm with promising results, it is essential to consider the potential limitations and shortcomings of the work. Some possible limitations include: Scalability: Dimensionality reduction algorithms often face challenges with scalability when dealing with large datasets. The paper should address the computational efficiency and scalability of the proposed SNEkhorn algorithm, especially when applied to high-dimensional and massive datasets commonly encountered in real-world applications. Parameter Sensitivity: It is essential to investigate the sensitivity of hyper parameters and assess how they impact the performance of the SNEkhorn algorithm, e.g. perplexity.

Reviewer WhX76/10 · confidence 4/52023-08-03

Summary

Existing DR approaches employ EAs after heuristic symmetrisation (symmetric-SNE). This leads to less faithful embeddings with low silhouette scores. This paper avoids such heuristic symmetrisation by enforcing symmetrisation in a related, new OT based formulation. Towards this goal, firstly, the EA problem is equivalently written as a (semi-relaxed) OT problem, where the transport plan recovers the EAs. Then symmetrisation is explicitly enforced as constraints on the transport map in this OT problem. Using optimality conditions it is shown that such EAs, maintain the crucial row-wise entropic equalities (prop4). As a result EAs are symmetric, doubly stochastic, and satisfy the entropic equalities, making the organically superior to the existing heuristics. Using the proposed EAs, a DR formulation is proposed (leading to so-called SNEkhorn). Details of solving this problem are presented. The proposed embeddings are empirically compared to the state-of-the-art wrt. spectral clustering and dimension reduction. The improvements are significant.

Strengths

1. The organisation and write-up are well polished and makes it an easy read. 2. Connections made between EAs and OT are interesting, especially in the light of the symmetrisation issue with existing DR approach. 3. Empirical improvements over baselines are impressive.

Weaknesses

..

Questions

1. It seems a primary reason for having Q_Z^{ds} in SNEkhorn is to enable fast objective-gradient computation via sinkhorn. Is this true? 2. Will it be more appropriate to consider other affinities like \tilde{Q}_Z etc. in place of Q_Z^{ds} ? If \tilde{Q}_Z leads to spherical embeddings, can there be applications where this is desirable? Any discussion on this and alternatives may help the reader.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations were discussed in the concluding section.

Reviewer WhX72023-08-18

Thanks for the reply. I would like to keep my score after reading the rebuttal and going through other reviews.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC