acceptodds
Under review as a conference paper at ICLR 2027

HoloWedge: Quantum Gravity inspired Learning of Hierarchical Embeddings

Abstract

Hyperbolic embeddings provide a natural representation of partially ordered data by encoding hierarchical relationships geometrically. However, existing learning methods suffer from undesirable features, including concentration of embedded data near the boundary, exclusion of the central region of hyperbolic space, and the necessity of pretraining using an external model. We introduce HoloWedge, a new framework for learning embeddings of hierarchical data inspired by theoretical physics, specifically the holographic approach to quantum gravity that relates the interior and the boundary of hyperbolic space. HoloWedge represents each data point by a subregion of hyperbolic space, a so-called wedge, which is bounded by a minimal-area hypersurface. We construct loss functions from areas of minimal surfaces, which encourage the containment relation between wedges to reflect the partial order within the underlying data. In the holographic model of theoretical physics, areas of minimal surfaces are identified with the entropies of the corresponding wedges. Our experiments demonstrate that our quantum gravity-inspired organizing principle outperforms existing approaches for reconstruction tasks in a moderate number of dimensions, while remaining competitive for link prediction tasks. The resulting embeddings are less concentrated at the boundary and utilize the full hyperbolic space. Moreover, the use of entropic ordering in our loss functions reduces sensitivity to initialization and removes the need for pretraining. Our results show that holography-inspired methods can effectively learn hierarchical representations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Hyperbolic Embeddings, Physics-inspired Representation Learning, Hierarchical Data, Entropy, Quantum Gravity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KpyLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for learning hierarchical embeddings in hyperbolic space termed HoloWedge, which is based on the following procedure. Each node is represented by a wedge, which is a hyperbolic half-space associated with a spherical cap on the boundary. The node’s direction in the embedding sets the cap’s center, while its distance from the origin determines the cap’s size. Ancestor-descendant relations are represented by wedge nesting. Inspired by holographic entropy, the method combines a score for the relative arrangement of two wedges with a penalty that encourages parents to have larger wedges than their children. Node embeddings are learned from observed relations and sampled negative pairs using cross-entropy or hinge loss. Experimental results show accurate graph reconstruction and mixed link prediction performance, with less concentration of embedded points near the boundary than some existing Poincare embedding methods.

Critical strengths
  • The paper incorporates holographic ideas into learning hierarchical embeddings.

  • Nesting wedges gives a clear way to represent hierarchy.

  • The method trains from random initialization without pretraining or excluding a region around the center.

  • Reconstruction results are favorable across several datasets and embedding dimensions.

  • The resulting embeddings are less concentrated near the boundary than some existing Poincare embedding methods.

Critical weaknesses
  • The comparisons to baselines are somewhat limited. For example, box embeddings (Boratko et al. NeurIPS 2021) and order embeddings (Vendrov et al ICLR 2016), which are cited in the paper, are not tested.

  • Link prediction results are mixed. On WordNet, the method performs worse than entailment cones in some cases.

  • The benefit of the entropy-inspired score is unclear. There is no comparison with a simpler containment loss using the same wedges.

  • Lower concentration near the boundary is shown, but its practical benefit is not clear. The paper does not show that it improves numerical stability or drives the accuracy gains.

  • There are no theoretical guarantees for hierarchy recovery, embedding accuracy, or optimization convergence.

  • The optimization procedure is not described in detail.

  • Computational complexity and runtime are not reported.

Questions for authors
  • Does the entropy-inspired score improve on a simple containment loss using the same wedges?

  • Does reduced concentration near the boundary improve numerical stability or accuracy?

  • Which optimizer and update rule was used?

  • What are the training time and memory usage and how do they compare with the baselines?

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (Part 1)

Thank you for your comments and the questions.

Does the entropy-inspired score improve on a simple containment loss using the same wedges?

As a first attempt, we in fact investigated a simple containment loss defined on the same RT wedges, before introducing the entropy-inspired score, but were not satisfied with the performance. We therefore decided not to include the results in the manuscript. In an upcoming revision, we will add the material, including a quantitative comparison, to the appendix.

In our first approach, for a child cc and a candidate parent pp, we used the score

E(c,p)=max⁡(0,γ+θc−θp),E{(c,p)} = \max(0, \gamma+\theta_c-\theta_p),

together with a hinge containment loss (inspired by the one used by Ganea et al. (2018)):

Lcont=1K+1[E++∑j=1Kmax⁡(0,m−Ej−)],m>0.\mathcal{L}_{\mathrm{cont}} = \frac{1}{K+1} \left[ E^{+} + \sum_{j=1}^{K} \max(0,m-E_{j}^{-}) \right], \qquad m>0.

Here, γ\gamma is the angular separation between the cap centers, θc\theta_c and θp\theta_p are their half-opening angles. This energy vanishes when the child cap is fully contained in the parent cap. Let E+E^{+} denote the energy of the positive pair and let Ej−E^{-}_{j}, with j=1,…,Kj=1,\dots,K, denote the energies of the KK corresponding negative pairs. This encourages positive pairs to satisfy containment, while negative pairs must violate it by a margin. Both the wedge parametrization and the containment condition are identical to those underlying HoloWedge. Although this containment loss performed better than the hyperbolic entailment-cones baseline in the initial reconstruction experiments, the embeddings remained concentrated near the origin of the Poincare ball when trained from random initialization, c.f. the 2d plot using the mammals subtree: https://anonymous.4open.science/r/rebuttal-figures-holowedge-575D/rt_containment_on_mammals_rsgd_2d.png.

This suggested the need for a preliminary distance-based pretraining as in Ganea et al. which we wanted to avoid.

We ran some sample computations, which show that our first approach consistently underperforms HoloWedge, see for example the following table for Mammal reconstruction:

ModelLRMetric2D
HoloWedge (CE)0.0750.075Best Mean Rank ↓\downarrow10.1129±2.0570\mathbf{10.1129 \pm 2.0570}
Best MAP ↑\uparrow0.7282±0.0172\mathbf{0.7282 \pm 0.0172}
⟨∥x∥2⟩\langle \Vert x\Vert^2 \rangle0.9554±0.00140.9554 \pm 0.0014
RT_containment0.30.3Best Mean Rank ↓\downarrow13.615±1.13713.615 \pm 1.137
Best MAP ↑\uparrow0.3796±0.00880.3796 \pm 0.0088
⟨∥x∥2⟩\langle \Vert x\Vert^2 \rangle0.4335±0.00050.4335 \pm 0.0005

Training settings: Results are reported as mean ±\pm standard deviation over three seeds (4242, 4343, 4444).

Starting from this containment baseline, we explored energy scores based on entropic quantities to more directly reflect the physical intuition underlying our approach. The final score presented in the paper yielded the best empirical performance among the variants we tested.

Does reduced concentration near the boundary improve numerical stability or accuracy?

While we do not claim that the lower concentration near the boundary drives accuracy gains (as was suggested in the weaknesses), its impact is twofold. First, one practical application of hyperbolic embeddings is visualization, and lower concentration near the boundary makes it possible to represent the hierarchical structure better, see also Figure 3 in Ganea et al. (ICML 2018). Second, the numerical precision needed for hyperbolic embeddings is set by requiring that (radial) distances can be resolved. For isometric embeddings it is shown in Sala et al. (ICML 2018) that the required numerical precision scales with the logarithm of distance to the boundary. This is of course only a rough upper bound on being able to tell apart two embedded data points. Our embeddings are (by design) not isometric, but we still think that the general point stands and carries over to any embeddings in hyperbolic space.

AuthorsOpenReview ↗

Official Comment by Authors (Part 2)

Which optimizer and update rule was used?

We used RSGD. In an upcoming revision, we will shift to RAdam which was used in the recent paper by Yu et al. (2024 ICLR) which improves our performance slightly over RSGD.

Regarding the update rules, for the RSGD we followed the additive Riemannian-gradient step of Nickel and Kiela (2017). Our optimizer updates each embedding as xt+1=xt−ηtgtRx_{t+1} = x_t - \eta_t g_t^R, where ηt\eta_t is the learning rate and gtRg_t^R is the gradient adjusted for the Poincaré geometry. For RAdam, we are following the HTorch implementation adopted for the Poincaré ball in Shadow cones (Yu et al. (2024)). The following equation summarizes the position update implemented in that code: xt+1=Π[Exp⁡xt(−αtmtvt+ϵ)]x_{t+1} = \Pi\left\lbrack \operatorname{Exp}_{x_t}\left( -\alpha_t \frac{m_t}{\sqrt{v_t}+\epsilon} \right) \right\rbrack, where mtm_t is the moving average of Riemannian gradients, vtv_t is the moving average of their squared Riemannian norms, and αt\alpha_t is the learning rate including bias correction. The optimizer rescales the averaged gradient adaptively, then uses the exponential map Exp⁡xt\operatorname{Exp}_{x_t}. The numerical projection Π\Pi keeps the updated point inside the Poincaré ball, while ϵ>0\epsilon>0 stabilizes the denominator. We will include this information in the upcoming revision.

What are the training time and memory usage and how do they compare with the baselines?

The training time of our models is larger than the baselines by a factor of three due to our more complex loss function, and the peak RAM usage is comparable. For instance, we obtain the following figures for the reconstruction task on the mammals dataset:

  • Poincaré embeddings: training time 0.326 s/epoch, walltime 193.7 s, peak RAM usage: 381.9 MB
  • Entailment cones: training time 0.310 s/epoch, walltime 177.7 s, peak RAM usage: 380.5 MB
  • HoloWedge CE loss: training time 0.997 s/epoch, walltime 531.0 s, peak RAM usage: 384.9 MB
  • HoloWedge Hinge loss: training time 0.869 s/epoch, walltime 461.9 s, peak RAM usage: 383 MB

These numbers were obtained on a MacBook Pro with an M3 Pro chip while running on the CPU. We will provide further details in the upcoming revision of our paper.

We would also like to comment on the following point.

Limited baselines. The central experimental question of our work was how HoloWedge compares with existing, gradient based, hyperbolic approaches to learning hierarchical representations, which explains our baseline selection. Box and order embeddings, while interesting in their own right utilize Euclidean space. Yu et al. (ICLR 2024) https://openreview.net/forum?id=zbKcFZ6Dbp&noteId=BTgyuPZlSZ provide a list of testruns that is comparable with our results on the Mammals subtree of Wordnet presented in appendix B. Comparing the numbers we see that the performance of our model is better than GBC and VBC, except for GBC in dim 5 at 50% and 90%. For comparison, shadow cones by Yu et al. which improves over Entailment cones also performes worse for d=5, 90%. While we have not included box embeddings in our baseline, we would like to emphasize the conceptual connection to our work (nesting implies hierarchical relation) and since, as explained in the discussion, our model can be seen as spherical embedding on the boundary of hyperbolic space, from a different point of view our HoloWedge and box embeddings indeed seem to be closer related than the difference in underlying geometry suggests.

I thank the authors for the detailed responses to my comments. The responses address well many of comments. In their response, the authors refer to future changes in the paper, which I believe are substantial and need to be reviewed, however, I did not find a revision.

Thank you for your reply. We are just waiting for the results of a few final runs to finalize the revised version of the manuscript. We will upload it as soon as possible.

Reviewer_i9KSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles a problem of embedding of hierarchical data in hyperbolic geometry. Authors claim to base their embedder on anti-de Sitter spacetime. The solution is compared against Poincare embedding in small-sized experiments.

Critical strengths

None

Critical weaknesses
  • This is another paper showing old results that are dressed as something new. Contrary to the popular belief in ML community, the observation that hyperbolic geometry is suitable for modelling hierarchical data was not found by Nickel and Kiela (2017). It was known even before 90's in past century. See, e.g., works by Tamara Munzner. The paper tries to make a similar move by showing a well-established result under a new name.

  • Usage of anti-de Sitter spacetime in this case is superficial, it seems like a clever way to obfuscate what was actually done (it seems we do not benefit at all from the time dimension, from my perspective it is just embedding with the half-planes, and those half-planes, wedges, has already been used before in the literature). In appendix, we are shown that actually we use a trivial case of one snapshot (t=0) which reduces to a well-known result (at least from Munzner's works if we confine ourselves to papers in computer science).

  • Omits a significant portion of literature, even from ICLR: Paper does not cite significant literature on hyperbolic embedders (no citing Tamara Munzner, Frederic Sala, Carlo Cannistraci, Fragkiskos Papadopoulos, Dima Krioukov, Marian Bogunya, Robert Jankowski, Dorota Celińska-Kopczyńska & Eryk Kopczyński,...). Especially, F. Sala, C. Cannistraci or D. Celińska-Kopczyńska had papers in the top ML conferences, including ICLR, which seems a grave omission.

  • Extremely limited scope of the proposition to hierarchies while there are currently at least 20 embedders to hyperbolic geometries that are capable of embedding both hierarchies and general networks.

  • Limited statistical validity of the results: mostly point estimates are presented

  • the experiments are very limited, important alternatives are missing. See Celińska-Kopczyńska and Kopczyński (ICLR 2026) benchmark on hyperbolic embedders. The paper even does not takes into account TreeRep which also is restricted to hierarchies. What is more, the previous benchmark studies did not show the universal dominance of Nickel & Kiela (2017), so it is not sufficient just to compare against them. Also, the scope of the experiments is very small -- WordNet and an artificial hierarchy are not sufficient.

  • False claims regarding hyperbolic geometry and embedders (based on popular misconceptions in ML community).

Questions for authors

How do you benefit from using anti-de Sitter here? How the time dimension is actually used here? Why your experiments do not compare against popular embedders apart from Nickel& Kiela 2017?

7 comments
AuthorsOpenReview ↗

Official Comment by Authors (Part 1)

Thank you for your comments.

We appreciate the comments on works on hyperbolic embeddings, in particular from other communities, which predates Nickel and Kiela and seems to have been widely overlooked in the ML community. We agree that they deserve more mentioning. In an upcoming revision, we will include references and discussion on the relevant works (Munzner, Sala, etc.) as well as works that establish a connection between communities (Celińska-Kopczyńska and Kopczyński). We also kindly refer you to our answer to reviewer 2vbt regarding the paper by Sala et al. (ICML 2018).

However, the mentioned works seem to focus on metric embeddings. Beyond this, our work addresses shortcomings of subregion based embedders such as Ganea et al. (ICML 2018) and Yu et al. (ICLR 2024) designed for link prediction tasks in any dimension, going beyond the already complicated task of having high quality embeddings of graphs. As already commented in the response to reviewer 2vbt, the example of Nickel and Kiela shows that good embeddings can produce subpar performance on link prediction if the prediction algorithm is built around the symmetric distance measure. We will clarify the positioning of our work relative to hyperbolic embedders and the link prediction task by a more careful phrasing in an upcoming revision.

To answer your questions:

How do you benefit from using anti-de Sitter here?

Using Anti-de Sitter space is certainly not intended to obfuscate our construction. The role of Anti-de Sitter space is to serve as a link between ML and physics, such that a physics-inspired mechanism can be implemented as an embedder for machine learning. We take AdS as a physical model in which information is hierarchically organized in hyperbolic space. Ryu-Takayanagi surfaces are singled out by having an information theoretic interpretation as an entropy. This arises as an insight from AdS/CFT and requires the RT surfaces to be solutions of an equation of motion. Restricting to a constant time slice of AdS happens because Ryu-Takayanagi surfaces are constant time, co-dimension two hypersurfaces. This physics-inspired entropic interpretation of the hyperplanes is one of the new aspects of our approach. We use this feature to build a new energy score involving well known entropic quantities. As we show in our paper, this physics-inspired approach improves performance over previous hyperbolic region based ML algorithms.

Why your experiments do not compare against popular embedders apart from Nickel & Kiela 2017?

We selected our baseline, which consists of both Nickel and Kiela, as well as Ganea et al. specifically with the motivation to show that certain problems of these models (and subsequent work such as Yu et al. (ICLR 2024)) can be overcome while in many cases improving performance. Our work benchmarks both reconstruction and link prediction, while most work on other embedders is focused on high quality embeddings only. Nickel and Kiela was chosen, since they present benchmarks on both, link prediction and reconstruction. Choosing other embedders is ambiguous, since no link-prediction algorithm is specified. Very recently (https://arxiv.org/pdf/2608.07029) a systematic study for link prediction of general networks was undertaken. However, for the link prediction algorithm studied in this paper we expect suboptimal performance for hierarchical data, since, as in Nickel and Kiela, it was based on the symmetric distance measure which empirically seems to have problems correctly capturing asymmetric relationships (c.f. the discussion in Ganea et al.). In addition, our algorithm should be seen as a direct improvement over methods for embeddings of hierarchically ordered data, such as Ganea et al., which itself compares to the popular algorithm of Nickel and Kiela.

AuthorsOpenReview ↗

Official Comment by Authors (Part 2)

We would also like to comment on additional points raised.

Note that reported numbers are not simply point estimates. For both reconstruction and link prediction, our results are obtained by averaging over three random seeds. We will provide statistics in the upcoming revision of our paper. In the case of link prediction, the standard error for all models is below 2% in F1 on the larger datasets and below 9% on the smaller datasets. For the reconstruction task, the standard error of MAP is below 1.3% on both the larger and smaller datasets. The standard error of mean rank is below 1.2 rank units on the larger dataset and 0.12 rank units on the smaller datasets.

Since the "critical weaknesses" paragraph suggests that we are making false claims regarding hyperbolic geometry, we would like to respectfully ask, which of our claims you think are false.

Similarly, our work addresses and improves on specific issues in the work of Ganea et al. While it is correct that many hyperbolic embedders exist, we disagree with the claim that we are "showing a well-established result under a new name." Instead, we have proposed a physics-inspired approach for hierarchical embeddings which performs well on both reconstruction and link prediction.

Reviewer_i9KSOpenReview ↗

Response

Thank you for the detailed response, but I am keeping my score.

  • How do you benefit from using anti-de Sitter here? -- I do not see using time dimension in your proposition. And in my opinion, if we advance to adS, this is where it shines.

  • The role of Anti-de Sitter space is to serve as a link between ML and physics, such that a physics-inspired mechanism can be implemented as an embedder for machine learning <-- I am surprised. That link has been well-established for many, many years. The whole community of studies on complex networks and ML exists in Physics, they even publish in ML conferences. Also, related de Sitter space-time has been much better included in the modeling of networks by Krioukov et al. [https://www.nature.com/articles/srep00793] and you are not even citing them, nor any paper from complexity studies and embedders from physics.

  • Very recently (https://arxiv.org/pdf/2608.07029) a systematic study for link prediction of general networks was undertaken. However, for the link prediction algorithm studied in this paper we expect suboptimal performance for hierarchical data, since, as in Nickel and Kiela, it was based on the symmetric distance measure which empirically seems to have problems correctly capturing asymmetric relationships (c.f. the discussion in Ganea et al.). <-- that paper compares about a dozen of embedders, including changes for link-prediction in more than one embedder. Also, it is not true that only N+K directly wrote about link-prediction task. E.g., HyperLink by Kitsak et al. (KVK in the benchmark paper you provided in the response) is specifically designed to account for missingness. It seems your study needs a lot of improvement when it comes to previous research and benchmarks for your proposition.

Reviewer_2vbtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents HoloWedge, a new hand-designed energy function on nested regions (Eq 8-11) that trains hyperbolic embeddings more stably from the origin with random initialization. Motivated by AdS/CFT and the Ryu-Takayanagi formula, it represents data points as hyperbolic wedges and encodes hierarchies through containment. The training objective leverages an entropy-based score and a hinge penalty to ensure ancestor wedges encompass their descendants. The experiments compared their method with some baselines like Entailment Cones in reconstruction and link prediction on WordNet and Hearst datasets.

Critical strengths

S1. Clear motivation from physics and cleanly flown derivations. Math and implementation details are checkable.

S2. Competitive-to-strong reconstruction results on WordNet datasets.

Critical weaknesses

W1. Important literature not cited and compared against. Since the paper talked about reconstruction, Sala et al. (ICML 2018) give a combinatorial construction that embeds a tree in hyperbolic space with arbitrarily low distortion, with no optimization. On WordNet it reaches a MAP of 0.989 in two dimensions, while the paper’s best WordNet Nouns MAP is 0.906 at d = 20 (Table 1). I understand that the experimental setups could be different, so that the two numbers are not directly comparable. But losing this important baseline makes the claimed improvement less reliable.

W2. In Table 1, HoloWedge’s ⟨‖x‖²⟩ (0.75–0.86) is higher than Entailment Cones’ (0.67), meaning its embeddings are on average pushed further to the boundary, which contradicts their claim that the method mitigates crowding of embeddings.

Questions for authors

Q1. Can you compare your method with the deterministic construction of "Representation Tradeoffs for Hyperbolic Embeddings", Sala et al. (ICML 2018)?

Q2. Can you explain in more details, preferably ablation studies, why your method has higher ‖x‖² than baselines?

3 comments

Thank you for your comments and questions.

Q1: Can you compare your method with the deterministic construction of "Representation Tradeoffs for Hyperbolic Embeddings", Sala et al. (ICML 2018)?

We are planning on amending our paper by adding a discussion on deterministic embedding algorithms.

The work of Sala et al. (ICML 2018) focuses on metric embeddings with distances obtained by (approximately) mapping a graph to a tree which is subsequently embedded into hyperbolic space. Our paper additionally addresses embeddings of partially ordered data for link prediction (i.e., a generalization task). While for reconstruction, we expect their model to outperform ours in both embedding quality (in low dimensions) and training time (in any dimension), we expect it to have the same problems as Nickel and Kiela (NeurIPS 2017) for link prediction, provided we use Nickel and Kiela's distance based algorithm. The reason is that it uses the symmetric distance function and thus has problems capturing asymmetric relationships. In general, comparing our results on link prediction to Sala et al. would require additional input to their model, to determine when two nodes have a hierarchical relationship. Also note that there is an overlap between the authors of Sala et al. and the Shadow Cones paper by Yu et al. (ICLR 2024), which does not compare against Sala et al., presumably for this reason.

Looking into the question regarding MAP in a bit more detail, we re-ran our model with a RAdam optimizer for three seeds (before we used approximate RSGD) and found a MAP of 0.982(1), which is surprisingly close to their value.

Q2. Can you explain in more details, preferably ablation studies, why your method has higher ⟨∥x∥2⟩\langle\Vert x\Vert^2\rangle than baselines?

The difference in smaller average radius of the embedded points of Ganea et al. (ICML 2018) compared to HoloWedge can be attributed to an additional step in their training, namely to the rescaling factor applied on the coordinates after pretraining. Note that while the algorithm of Ganea et al. crucially depends on this step, our algorithm bypasses it completely.

To reproduce the results of Ganea et al. for comparison we used a rescaling factor of 0.7 for entailment cones, as was done in the original paper. In the table below, we show the results for ⟨∥x∥2⟩\langle\Vert x\Vert^2\rangle for different values of this rescaling factor, while keeping all other parameters constant. These results were obtained from the reconstruction task on the mammals dataset in five dimensions for one random seed. All other parameters are chosen like in our paper. We note that for larger rescaling factors, the average squared norm also increases, and without rescaling, the embeddings are pushed much further to the boundary than in our model. For reference, our models reach ⟨∥x∥2⟩=0.841\langle\Vert x\Vert^2\rangle=0.841 for the CE loss and ⟨∥x∥2⟩=0.732\langle\Vert x\Vert^2\rangle=0.732 for the hinge loss in the same setting. These values are obtained without any additional rescaling; it is rather an inherent property of our models that the points are arranged closer to the center compared to the entailment cones model without rescaling.

Rescaling factor0.50.60.70.80.91.0
⟨∥x∥2⟩\langle\Vert x\Vert^2\rangle0.4660.5590.6520.7450.8380.936

Thank you for the experiments on the rescaling factor results. It partially resolved my questions. However, there is no amended manuscript to review now. So I'd keep my evaluation.

Thank you for the comment. We have just posted the revision of the paper. We hope it will resolve the remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Resubmission Posted

We thank the reviewers for their constructive comments, which we believe have helped to make the paper stronger. In the new revision, we have:

  • reworked the introduction to position our work more clearly where we intended, namely contributing to the development of embedding methods which are also suitable for link prediction

  • extended the number of references throughout the paper. We have mostly focused on methods for embedding hierarchical data and have decided to address additional uses of hyperbolic space for embedding on general graphs by referring to comprehensive comparisons, such as (Celinska-Kopczynska and Kopczynski, 2026)

  • reworked the related work section, discussing deterministic embedders and explaining our choice of baselines

  • added more detailed information about the used optimizers in appendix A

  • added an analysis on the effect of hyperparameter choice on the radial distribution of datapoints to appendix A

  • added information about training time and memory usage in appendix A

  • updated our results in the main text with results obtained using RAdam (this decision was made after submission, but before the review process started). For completeness and transparency, all previously reported RSGD results are retained in appendix B

  • added information about statistics in the table captions as well as given more detailed information in appendix B

  • added appendix F containing a discussion on a containment-based energy function

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.