Fitting trees to $\ell_1$-hyperbolic distances

Building trees to represent or to fit distances is a critical component of phylogenetic analysis, metric embeddings, approximation algorithms, geometric graph neural nets, and the analysis of hierarchical data. Much of the previous algorithmic work, however, has focused on generic metric spaces (i.e., those with no a priori constraints). Leveraging several ideas from the mathematical analysis of hyperbolic geometry and geometric group theory, we study the tree fitting problem as finding the relation between the hyperbolicity (ultrametricity) vector and the error of tree (ultrametric) embedding. That is, we define a vector of hyperbolicity (ultrametric) values over all triples of points and compare the $\ell_p$ norms of this vector with the $\ell_q$ norm of the distortion of the best tree fit to the distances. This formulation allows us to define the average hyperbolicity (ultrametricity) in terms of a normalized $\ell_1$ norm of the hyperbolicity vector. Furthermore, we can interpret the classical tree fitting result of Gromov as a $p = q = \infty$ result. We present an algorithm HCCRootedTreeFit such that the $\ell_1$ error of the output embedding is analytically bounded in terms of the $\ell_1$ norm of the hyperbolicity vector (i.e., $p = q = 1$) and that this result is tight. Furthermore, this algorithm has significantly different theoretical and empirical performance as compared to Gromov's result and related algorithms. Finally, we show using HCCRootedTreeFit and related tree fitting algorithms, that supposedly standard data sets for hierarchical data analysis and geometric graph neural networks have radically different tree fits than those of synthetic, truly tree-like data sets, suggesting that a much more refined analysis of these standard data sets is called for.

Paper

References (38)

Scroll for more · 26 remaining

Similar papers

Peer review

Reviewer xUok6/10 · confidence 1/52023-07-06

Summary

The paper introduces a new algorithm, HCCRootedTreeFit, for fitting tree metrics to a given distance matrix. The algorithm is designed to minimize the ℓ1 distortion of the fit. The authors provide a detailed explanation of the algorithm, its theoretical properties, and an extensive experimental evaluation. The results show that the algorithm performs optimally when datasets are close to tree-like and when distortion is measured in the ℓ1 sense. The paper suggests that commonly used datasets, especially in geometric graph neural nets, are not well-represented by trees, indicating the need for more refined geometric notions for learning tasks with these datasets.

Strengths

The algorithm is theoretically sound, thoroughly evaluated, and performs optimally on tree-like datasets. The work provides valuable insights into dataset characteristics and has significant potential for practical applications, particularly in machine learning and data analysis.

Weaknesses

The evaluation of the proposed algorithm is primarily focused on its performance on tree-like datasets. While this is certainly important, it would be beneficial to see how the algorithm performs on a broader range of datasets, particularly those that are not tree-like. This would provide a more comprehensive understanding of the algorithm's performance and its applicability to real-world problems.

Questions

Please solve the weakness above

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer uKfW6/10 · confidence 2/52023-07-07

Summary

The authors consider the tree fitting problem for a given distance. The authors cast the tree fitting problem as finding the relation between the hyperbolicity vector and the error of tree embedding. The authors propose an algorithmic approach with provably tight $\ell_1$ error. The authors also illustrate the advantage of the proposed approach on some “tree-like” datasets.

Strengths

+ By casting the fitting tree problem as finding the relation between the hyperbolicity vector and the error of tree embedding, the authors propose provably tight $\ell_1$ error algorithmic approach. + The authors illustrate the advantage of the proposed approach on some “tree-like” datasets. + Overall, the presentation is good. (It may be better to elaborate more details on the proposed algorithmic approach in Section 3.3)

Weaknesses

+ At the heart of the proposed algorithmic approach (Section 3.3), although the authors summarize their proposed approach in several Algorithms, it is hard to get the ideas how the tree is constructed. It seems the authors focus more on analysis for the proposed algorithm. + It seems better to include discussions about other existing approaches of tree fitting for a given distance. Besides the provable analysis of the proposed algorithmic approach, it is not clear how the proposed approach addresses some limits of existing approaches on tree fitting for a given distance.

Questions

+ Could the authors elaborate the main ideas of the proposed algorithmic approach in Section 3.3? (although some algorithms are given to summarize the ideas, it is better to discuss about it, e.g., how the tree is constructed to fit a given distance, besides the analysis)? + Could the authors discuss how the proposed algorithmic approach improves over some simple approach, such as the QuadTree approach (which is popularly used in the context of optimal transport)? + If possible, please consider the QuadTree approach as a baseline for tree fitting in the experiments. + In Algorithm 1, could the authors explain the “highlyconnected” concept? How is it defined as in the Algorithm 1? + In line 269, why do the negative weights appear? (for tree metric, should the weights be nonnegative?) It seems unclear whether these approaches try to find a closest tree metric for a given distance? Some minor points: + In line 58, “the table below”? which table is mentioned? It is better to have a reference for the mentioned table. + In Theorem 3.4, could the authors comment on the constant $C=4$? Is it a tight result or is there any chance to improve it? --- Thank you for the rebuttal.

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There are no discussions about the limitations and potential negative societal impact of their work.

Reviewer x4xj5/10 · confidence 3/52023-07-09

Summary

The paper formulate the $l_p$ tree fitting problem introduces a new algorithm, HCCROOTEDTREEFIT, for building trees in hyperbolic space by investigating the relationship between hyperbolicity (ultrametricity) vectors and the error of tree (ultrametric) embedding, which outperforms previous methods both theoretically and empirically.,

Strengths

1) The authors have developed a novel approach to the tree-fitting problem that applies hyperbolic geometry and geometric group theory. 2) The developed algorithm, HCCROOTEDTREEFIT, delivers a tree metric with $l_1$ distortion bounded by polynomial function of the average hyperbolicity, while previous research delivers a tree metric with $l_{\infty}$ distortion only. 3) Comparison not only on performance but also on speed are explored, which makes a more computational sense. 4) Repeated experiments are conducted and standard deviation was analyzed.

Weaknesses

The paper is not well self-contained, needing additional supplementary materials to be complete. Table captions are not very comprehensive, e.g., Table 3 and Table 5, which increase the difficulty to understand the major experimental results. The results lack qualitative examples to show 1) how different the proposed tree-fitting method is compared with previous tree-fitting methods. 2) how different the synthetic datasets is to real datasets.

Questions

As described in the abstract, the proposed method is inspired by geometric group theory, can the authors be more specific on what is inspired the novel approach to the tree-fitting problem using which part of geometric group theory? What does the notation in equation between line 82 and 83 mean? $\begin{pmatrix} X \\\\ 3 \end{pmatrix}$ seems to mean combinatoric number in convention but $X$ here is a set.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The authors does not discussed about limitations, but this method is quite theoretical and is very unlikely to have potential negative societal impact

Reviewer Fg5u5/10 · confidence 1/52023-07-14

Summary

This paper introduces a novel tree fitting algorithm named HCCRootedTreeFit. First, the authors motivate the need for a better tree fitting algorithm by stating that current methods "assume almost nothing about the underlying discrete point set, when, in fact, many real application data sets are close to hierarchical or nearly so". Before introducing their new method, the authors introduce new proxy measures of how tree-like a dataset is. These proxy measures are used later on in several bounds, including in their main theoretical result which guarantees the existence of a tree fitting method with some nice bound on the distortion. In their theoretical analysis, the authors mention a connection between the tree fitting problem and hierarchical correlation clustering (HCC), along with an equivalence result between tree fitting algorithms and ultrametric fitting algorithms. Then, they introduce an adapted HCC problem formulation and three algorithms named HCCTriangle, HCCUltraFit and HCCRootedTreeFit. The first algorithm solves the adapted HCC problem. The second algorithm uses HCCTriangle to solve the ultrametric fitting problem. The last algorithm uses HCCUltraFit to solve the rooted tree fitting problem. In the experiments, the HCCRootedTreeFit algorithm is tested versus various other methods from the literature. For these experiments the authors have used both common and synthetic datasets. They observe that their method underperforms on common datasets, but performs well on synthetic datasets with respect to the $\ell_1$ norm. They attribute this observation to a lack of tree-like structure in the common datasets. They furthermore observe that their method underperforms with respect to the $\ell_\infty$ norm.

Strengths

1. The paper introduces interesting new notions of hyperbolicity and ultrametricity through their hyperbolicity and ultrametricity vectors. 2. The authors derive an interesting connection between the distortion bounds for tree fitting and ultrametric fitting algorithms. 3. The new notion of hyperbolicity is used to show that, contrary to common belief, the common datasets are not very tree-like.

Weaknesses

The paper is very difficult to read. Due to the very large number of definitions and technical results, the paper reads more like a collection of statements than a paper. Much of the actual content seems to be put in the appendix, adding to the feeling that the paper is not at all self-contained. Moreover, the paper appears to use many tricks from different papers without any description of these tricks. For example, in line 233, the authors refer to some paper to obtain a tree fit without actually explaining the actual procedure. The structure of the paper also makes it rather difficult to follow, with several forward- and backward references spread throughout the paper. As an example, Subsection 3.2 appears to state several results regarding algorithms that have not yet been introduced. Then, in the next subsection, these algorithms are introduced and following these algorithms, the authors quickly throw in a proof of a result from the previous Subsection. Another issue with the paper is that they first formulate the tree (ultrametric) fitting problem, but then, after quickly mentioning a connection to hierarchical correlation clustering, seem to actually solve an adapted version of this HCC problem. However, the introduction of their adapted HCC problem is rather unclear and uses, for example, a 'number of disagreement edges' term that is not properly defined in the paper. Also, it is not really clear to me what the connection with this problem and the original problem is. Due to these issues I cannot adequately judge the validity of the many technical results of this paper. There is also an issue regarding the newly proposed algorithm within the context of the experimental results. The authors show that their method underperforms on common datasets with respect to both the $\ell_1$ norm and the $\ell_\infty$ norm. Moreover, the method is quite slow compared to two of the other methods. It therefore appears that this method is only useful in synthetic settings.

Questions

1. What is the exact connection between the tree fitting problem and the HCC problem? 2. How is the term $|E_t \Delta E(P_t)|$ defined precisely?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

Mostly yes. They do not really address the fact that their method does not seem to have a proper use case, but they do point out its weaknesses in the experiments.

Reviewer Fg5u2023-08-14

I would like to thank the authors for their clarifications. I understand that including the proofs and details of the algorithms is difficult given the page limit. However, I do strongly recommend the authors to: 1. add some explanations whenever results from other papers are introduced and applied; 2. restructure the paper to improve readability and remove forward- and backward references when possible.

Reviewer xUok2023-08-18

Thanks for the author's reply. May I ask in what application scenarios the author's method can be used and what kind of problems it can solve?

Reviewer x4xj2023-08-18

Thank you for the author's response. I would like to discuss further the last three points: * This question pertains to **qualitative** examples. It would be beneficial to include visualizations to demonstrate the superiority of the proposed method, especially since hyperbolic geometry is a mainstream tool in the visualization community [c1]. * Gromov's original work in Geometric Group Theory is cited in almost all the related works. However, my inquiry regarding **being specific** wasn't adequately addressed. * In chapter 1.2 of Stanley’s "Enumerative Combinatorics" [c2] and a more recent version [c3], the convention is **NOT about subsets**. Instead, it is also used for combinatorial numbers. Could the authors clarify the origin of this convention? [c1] Lamping J, Rao R, Pirolli P. A focus+ context technique based on hyperbolic geometry for visualizing large hierarchies[C]//Proceedings of the SIGCHI conference on Human factors in computing systems. 1995: 401-408. [c2] Stanley R P. Enumerative combinatorics, v. 1[J]. Wadsworth and Brooks/Cole Mathematics Series, Monterey, California, 1986. [c3] Stanley R P. Enumerative Combinatorics Volume 1 second edition[J]. Cambridge studies in advanced mathematics, 2011.

Authorsrebuttal2023-08-18

Strictly speaking we are not using the "group" portion of the phrase "geometric group theory." One of the origins of the field was to study the Cayley graphs of finitely generated groups. If one endows such Cayley graphs with metrics and considers these graphs as metric spaces, sometimes they are hyperbolic spaces (and the groups are referred to as hyperbolic groups). The machinery for considering both graphs and hyperbolic spaces is useful, however, despite the disconnection from a group.

Authorsrebuttal2023-08-21

In chapter 1.2 of Stanley’s "Enumerative Combinatorics" (We both checked [c2] and [c3]), they stated that "Now define $\binom{S}{k}$ (sometimes denoted $S^{(k)}$ or otherwise, and read "S choose k") to be the set of all $k$-element subsets (or $k$-subsets) of $S$. And "then" they defined the usual combinatorial number. While we understand this convention is not widely accepted as the number convention, we have found that several references already used this notation, probably due to its convenience.

Reviewer x4xj2023-08-21

When talking about a **convention**, it should be widely adopted, especially in the textbook[c1,c2] mentioned in the rebuttal. In this textbook, the notation appears for combinatorial subset **only once**, while it is used for combinatorial number **more than 1000 times**. While I disagree with the overstatement in the rebuttal , I recognize its convenience and would stop the discussion on this minor point.

Area Chair pGit2023-08-18

Further discussion needed?

Dear reviewer, Please let us know what you think about the rebuttal and whether you have any points you want to raise. Thank you, AC

Reviewer uKfW2023-08-18

Thank you for the rebuttal, I have no other raised points.

Reviewer xUok2023-08-20

Thanks for the author's response.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC