Pith. sign in

REVIEW 4 major objections 5 minor 38 references

Even Faster Hyperbolic Random Forests: A Beltrami-Klein Wrapper Approach

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Projecting hyperbolic data into Beltrami-Klein coordinates, training a Euclidean tree, and correcting thresholds with Einstein midpoints reproduces HyperDT's boundaries while cutting training time by orders of magnitude.

desk verdict Clever wrapper idea undermined by a systematic sign error in the printed formulas; the equivalence theorem is false as stated, but the approach should be repairable. read the letter →

arxiv 2506.04360 v1 pith:EFS4Q6XE submitted 2025-06-04 cs.LG

Philippe Chlenski , Itsik Pe'er This is my paper
classification cs.LG
keywords hyperbolicgeometrydecisiontreesrandomforestsBeltrami-KleinmodelLorentzgnomonicprojectionEinsteinmidpointmachinelearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Hyperbolic decision trees (HyperDT) have required their own bespoke implementation because their hyperplane splits do not reduce to simple coordinate thresholds. This paper shows that, in the Beltrami-Klein model of hyperbolic space, geodesic decision boundaries become straight Euclidean hyperplanes, so a hyperbolic tree can be reduced to three steps: project the data with the gnomonic projection, train an ordinary Euclidean decision tree or random forest, then replace each learned threshold with the Einstein midpoint of the two neighboring training points. The authors prove this wrapper produces the same partitions as HyperDT (Theorem 4.5) at the same asymptotic complexity as CART, and report training speedups of over three thousand times on large samples. If correct, this makes hyperbolic tree methods practical and lets them inherit the optimizations of mature Euclidean libraries, including gradient-boosted and oblique tree variants.

What carries the argument

The Beltrami-Klein model of hyperbolic space, obtained by the gnomonic projection $\varphi_K(u_0,\dots,u_d)=(u_1/u_0,\dots,u_d/u_0)$ from the Lorentz model, is the load-bearing object: in this model hyperbolic geodesics appear as straight chords and HyperDT's homogeneous hyperplane splits become ordinary axis-aligned coordinate thresholds (Lemma 4.1). The second component is the Einstein midpoint $m_B(u,v)=(\gamma_u u+\gamma_v v)/(\gamma_u+\gamma_v)$ with $\gamma_x=1/\sqrt{1-K\|x\|^2}$, which the paper argues equals HyperDT's angular midpoint (Lemma 4.3) and is used to recompute decision-boundary thresholds so that they remain hyperbolically equidistant. Threshold invariance (Lemma 4.2) guarantees that any threshold between adjacent training values is equivalent, so the postprocessing step can choose the Einstein midpoint without changing the learned split structure.

What would settle it

Numerically compare the two midpoint formulas on many random hyperboloid pairs with $K=-1$, especially near branch cuts of the arccotangent; any pair where Eq. 15 and the angle of the Einstein midpoint differ beyond machine precision would show the equivalence is not exact. An even more direct check is to train HyperDT and Fast-HyperDT on a dataset with a forced tie-free optimal split at an extreme angle and compare the induced partitions.

Watch

Extended reading notes

Core claim

The central claim is that HyperDT's geodesic-convex splits—homogeneous hyperplanes through the origin in the Lorentz model—are exactly threshold tests on the coordinates of the Beltrami-Klein projection, and that HyperDT's hyperbolic angular midpoint for placing a split boundary coincides with the Einstein midpoint of the projected boundary points in the Klein model. Consequently the entire HyperDT algorithm can be reexpressed as gnomonic preprojection, standard Euclidean tree training, and a postprocessing pass that adjusts thresholds to Einstein midpoints (Algorithms 1–3). The equivalence is exact under identical tie-breaking (Theorem 4.5), and the training complexity remains $O(nd\log n)$ with inference $O(hn)$ (Theorem 4.6). The wrapper also opens the door to gradient-boosted and oblique tree backends, and empirical agreement between HyperDT and Fast-HyperDT ranges from roughly 98% to 100% on train and test predictions, with residual mismatches attributed to tie-breaking and numerical stability at extreme angles.

Load-bearing premise

The whole equivalence hinges on one geometric identity: that the midpoint used by HyperDT coincides with the Einstein midpoint of the two projected boundary points; if those two midpoints ever differ, the wrapper produces different decision boundaries than HyperDT.

Editorial extensions

If this is right

  • Hyperbolic trees and forests can be trained with highly optimized Euclidean implementations, with reported speedups up to about 3,752 times on 32,768 samples while matching HyperDT's decision boundaries.
  • The same wrapper recipe extends to gradient-boosted and oblique decision trees, making hyperbolic geometry a preprocessing and postprocessing concern rather than a new algorithm.
  • If the midpoint equivalence holds universally, any Euclidean tree that stores per-node thresholds can be made hyperbolic with the same three steps, and future Euclidean tree optimizations transfer automatically.
  • The equivalence is conditional on identical information-gain tie-breaking; in practice the two methods agree on split selection more than 99% of the time, and residual differences are attributed to the original HyperDT's numerical instability near angles of roughly $\pi/4$ and $3\pi/4$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct corollary we draw: any future speedup in Euclidean decision-tree libraries (GPU split finding, streaming trees, learned thresholds) immediately upgrades hyperbolic forests, because the wrapper's extra work is only $O(nd)$ projection plus $O(nh)$ midpoint postprocessing.
  • The tie-breaking sensitivity implies that benchmarks comparing hyperbolic tree variants should report split-selection tie rules; otherwise small accuracy differences may be library artifacts rather than geometric effects.
  • The three-step recipe (project, train, remidpoint) is a template for other threshold-based models: isolation forests and rotation forests, which the paper lists as future targets, would only need a straight-line model and a midpoint correction.
  • If the equivalence in Theorem 4.5 is exact, the hyperbolic random forest is not a new learning algorithm but a geometric pre- and postprocessing of a classical one, which may simplify theoretical analysis and porting of existing tree-based methods.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Fast-HyperDT, a wrapper that rewrites HyperDT as: (i) gnomonic projection of Lorentz-model hyperbolic data to the Beltrami-Klein model, (ii) training a standard Euclidean decision tree or random forest on the projected coordinates, and (iii) postprocessing every split threshold by replacing it with an Einstein midpoint of the two neighboring training values. The central claim is Theorem 4.5, that this wrapper produces exactly the same decision boundaries as HyperDT, together with Theorem 4.6, that it matches CART's O(nd log n) training and O(hn) inference complexity. The paper reports a roughly 3,750x training speedup over the original HyperDT implementation, near-perfect but not exact agreement in split/prediction comparisons, and accuracy/MSE tables comparing Fast-HyperDT with baselines including scikit-learn, oblique trees, LightGBM, and XGBoost.

Significance. If the equivalence theorem is correct, this is a practically valuable and conceptually clean reduction: it turns a specialized hyperbolic tree implementation into a thin preprocessing/postprocessing layer over mature Euclidean tree libraries, with large speedups and easy extensibility to XGBoost, LightGBM, and oblique trees. The paper ships a public implementation, and the main construction is concrete and falsifiable. However, the printed formulas contain a sign error in the inverse gnomonic projection and Einstein midpoint (Eqs. 6 and 9), which breaks the equivalence theorem as stated; the correctness of the central claim is therefore conditional on a systematic correction that is local but load-bearing. Lemma 4.3 also has a genuine proof gap, and the empirical agreement is imperfect in a way that is not reconciled with Theorem 4.5.

major comments (4)
  1. [2.1.2, Eqs. (6) and (9); Algorithm 1] Under the paper's convention K<0, the inverse gnomonic projection and the Einstein midpoint weights use 1-K||x||^2 where the geometry requires 1+K||x||^2. For K=-1 and v=(q,0,...,0) with q^2=0.58, Eq. (6) gives phi^{-1}(v)=(1/sqrt(1.58), q/sqrt(1.58)), whose Minkowski norm is (q^2-1)/(1+q^2) ~ -0.266, not the required -1, so phi^{-1} does not map the Klein ball into the Lorentz model. Consequently the midpoint used in Algorithm 1 is not the hyperbolic geodesic midpoint: for u=(1,0) and v=(cosh 1, sinh 1), the correctly equidistant midpoint has cot(theta)=tanh(1/2) ~ 0.462, whereas Eq. (9) as printed gives a Klein coordinate of about 0.337 and theta ~ 71.3 degrees. Since Algorithm 1 and Lemma 4.3 rely on Eq. (9), Theorem 4.5 does not hold for the formulas as printed. The fix is local (replace 1-K||x||^2 with 1+K||x||^2 throughout), but it must be applied consistently to Eqs. (6), (7), (9), and the postprocessing in Algorithm 1.
  2. [4, Lemma 4.3] The proof of Lemma 4.3 does not establish the claimed equality with Eq. (15). It refers to 'the unique point in L intersect {h0 cos(theta_L)-hd sin(theta_L)=0}' without first restricting to the 2-plane spanned by u and v; in d>2 that intersection is a (d-1)-dimensional submanifold, not a unique point. More importantly, the proof only shows that the Einstein midpoint m_B is equidistant from u and v; it never shows that the angle of m_B satisfies the defining equation of theta_L given in Eq. (15). Uniqueness of hyperbolic midpoints is insufficient because the angular midpoint's status as the geodesic midpoint is exactly what Eq. (15) is supposed to encode. Equations (23)-(24) also contain normalization/convention problems (1/sqrt(K) with K<0, and a negative denominator in Eq. (24)), so the intermediate algebra needs to be rewritten with consistent Lorentz-model normalization.
  3. [5.2, Figure 3] The empirical agreement results are not fully consistent with Theorem 4.5. Figure 3 shows train prediction agreement as low as 99.4%, test prediction agreement as low as 98.4%, and node-value agreement as low as 96.6%; the text reports 22 splits where Fast-HyperDT attains a higher information gain than HyperDT. The theorem's conclusion is exact equality up to tie-breaking, so these mismatches need a quantitative explanation rather than a brief appeal to 'pragmatic factors.' The sign error in Eq. (9) is an obvious candidate explanation, and the authors should rerun the agreement benchmark after correcting it and report whether the mismatches disappear. As written, the experimental section weakens, rather than confirms, the central equivalence claim.
  4. [5.1, Tables 1 and 2] The accuracy and MSE comparisons in Tables 1 and 2 lack error bars or standard deviations over the 100 synthetic benchmarks, so differences of 1-3 points (e.g., Fast-HyperRF vs HyperRF in several rows of Table 1) are not statistically assessable. In addition, the related-work section cites HoroRF (Doorenbos et al., 2023) as the main complementary hyperbolic random forest, but no HoroRF baseline appears in the experiments, making it hard to position Fast-HyperDT against the existing state of the art. At minimum, report means with standard deviations or confidence intervals, and add a HoroRF comparison where feasible.
minor comments (5)
  1. [2.1.1, footnote 1] Footnote 1 contains an unresolved editorial note ('ip: I think we should use arccosh arctan etc. instead of the inverse (negative 1 superscript)') that should be removed or resolved before publication.
  2. [3, Algorithm 2] Algorithm 2 is internally inconsistent: PredictNode computes phi(x)_d = x_d/x0 from the Lorentz input, but line 16 calls PredictNode(X_B[i], T) as though the full Klein projection X_B were precomputed. Clarify which representation is passed.
  3. [4, Lemma 4.3 proof] There are small typos in the lemma's proof: 'v in L^n_K' should be 'v in L^d_K', and Eq. (26) writes m_E where m_B is meant.
  4. [5.3, Section 5.3 ablation note] The note that 'the two ablations actually coincide, as the angular bisector and the average of the Klein coordinates are equal' is not fully derived; the displayed formula 'cot(theta_m) = (u_d/x0 + x_d/x0)/2' appears to contain a typo (u_d/x0 should presumably be u_d/u0) and would benefit from a short derivation.
  5. [2.1.2, Eq. (7)] Eq. (7) inherits the sign error from Eq. (6); after correcting Eq. (6), the distance formula should be re-derived and checked against the standard Klein-model distance for K<0.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the wrapper equivalence is derived from geometric lemmas and validated against the HyperDT baseline, not assumed.

full rationale

The paper's central claim is that Fast-HyperDT, a Beltrami-Klein wrapper around Euclidean decision trees, produces the same decision boundaries as HyperDT. Lemma 4.1 and Lemma 4.2 are self-contained algebraic and combinatorial arguments. Lemma 4.3 attempts an independent geometric proof that the Einstein midpoint is equidistant from the endpoints and then invokes uniqueness of hyperbolic geodesic midpoints, which is cited to an external textbook (Ratcliffe, 2019). The only reliance on the authors' prior work is the definition of HyperDT's angular midpoint (Eq. 15) and its equidistance property, which is a baseline definition rather than the target conclusion. The paper does not assume Fast-HyperDT matches HyperDT; it argues for the equivalence and then checks it empirically (Figure 3). Thus there is no step where the output is inserted as an input, no fitted parameter renamed as a prediction, and no self-citation chain that forces the result. The sign-convention issue in Eqs. 6 and 9 identified by reviewers is a correctness risk, not a circularity, because it does not make the argument assume its conclusion. Score 2 reflects one minor self-citation of the prior HyperDT midpoint formula in the proof of Lemma 4.3, but it is not load-bearing in a circular sense because the equivalence is supported by an attempted derivation and external comparison.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No free parameters are fitted to make the derivations work; curvature K is an input from the hyperbolic model. The main unstated inputs are standard results in hyperbolic geometry and the prior HyperDT midpoint formula. No new entities are introduced.

assumptions (5)
  • standard math The gnomonic projection maps Lorentz-model geodesic hyperplanes to Euclidean (not necessarily axis-aligned) hyperplanes in the Beltrami-Klein model.
    Used throughout Section 3 and Lemma 4.1; it is the geometric basis for the thresholding reformulation.
  • standard math Each point in hyperbolic space has a unique midpoint along a geodesic, implied by geodesic convexity.
    Invoked in the proof of Lemma 4.3 to equate the angular midpoint with the Einstein midpoint; cites Ratcliffe (2019).
  • domain assumption The HyperDT angular midpoint formula (Eq. 15) yields a hyperplane equidistant from the two endpoint points under the hyperbolic distance.
    Taken from the authors' prior paper (Chlenski et al., 2024); not re-derived here, yet it is the target that Fast-HyperDT's Einstein-midpoint postprocessing must reproduce.
  • ad hoc to paper Tie-breaking between splits of equal information gain is identical in HyperDT and Fast-HyperDT.
    Theorem 4.5 assumes it; Section 5.2 shows it is not perfectly satisfied in practice (up to 3.4 percent mismatch), so the equivalence is only approximate at the level of exact implementations.
  • domain assumption For XGBoost and LightGBM, tree splits are axis-parallel and node thresholds are accessible for postprocessing.
    Sections 3.1.2 and 3.1.3; the postprocessing is approximate when subsampling or bagging is enabled, acknowledged with a warning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Even Faster Hyperbolic Random Forests: A Beltrami-Klein Wrapper Approach." pith.science (2026). https://pith.science/paper/EFS4Q6XE

@misc{pith2026250604360,
  author       = {Pith},
  title        = {Pith review of: Even Faster Hyperbolic Random Forests: A Beltrami-Klein Wrapper Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EFS4Q6XE}},
  note         = {Machine review of arXiv:2506.04360}
}
read the original abstract

Decision trees and models that use them as primitives are workhorses of machine learning in Euclidean spaces. Recent work has further extended these models to the Lorentz model of hyperbolic space by replacing axis-parallel hyperplanes with homogeneous hyperplanes when partitioning the input space. In this paper, we show how the hyperDT algorithm can be elegantly reexpressed in the Beltrami-Klein model of hyperbolic spaces. This preserves the thresholding operation used in Euclidean decision trees, enabling us to further rewrite hyperDT as simple pre- and post-processing steps that form a wrapper around existing tree-based models designed for Euclidean spaces. The wrapper approach unlocks many optimizations already available in Euclidean space models, improving flexibility, speed, and accuracy while offering a simpler, more maintainable, and extensible codebase. Our implementation is available at https://github.com/pchlenski/hyperdt.

Figures

Figures reproduced from arXiv: 2506.04360 by the authors.

Figure 1
Figure 1. Original HyperDT decision boundaries visualized in the Poincaré model (top) and in the Beltrami￾Klein model (bottom). The top half of the figure is a reproduction of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. We compare the training time of HyperDT and Fast-HyperDT across different training set sizes. Fast-HyperDT is consistently faster than HyperDT for 8 or more samples; for 32,768 samples, Fast-HyperDT trained on average 3,752× faster. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. We compare the predictions and split values of [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: A Gaussian mixture dataset for which multiple splits attain the same information gain. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 21 canonical work pages

  1. [1]

    Gaussian Beltrami - Klein Model for Protein Sequence Classification : A Hyperbolic Approach

    Sarwan Ali, Haris Mansoor, Prakash Chourasia, Yasir Ali, and Murray Patterson. Gaussian Beltrami - Klein Model for Protein Sequence Classification : A Hyperbolic Approach . In Wei Peng, Zhipeng Cai, and Pavel Skums (eds.), Bioinformatics Research and Applications , pp.\ 52--62, Singapore, 2024. Springer Nature. ISBN 9789819751280. doi:10.1007/978-981-97-5128-0_5

    work page doi:10.1007/978-981-97-5128-0_5 2024

  2. [2]

    Neural Networks are Decision Trees , October 2022

    Caglar Aytekin. Neural Networks are Decision Trees , October 2022. URL http://arxiv.org/abs/2210.05189. arXiv:2210.05189 [cs]

    arXiv 2022

  3. [3]

    Bagnall, M

    A. Bagnall, M. Flynn, J. Large, J. Line, A. Bostrom, and G. Cawley. Is rotation forest the best classifier for problems with continuous features?, April 2020. URL http://arxiv.org/abs/1809.06705. arXiv:1809.06705 [cs]

    arXiv 2020

  4. [4]

    Random spanning forests and hyperbolic symmetry

    Roland Bauerschmidt, Nicholas Crawford, Tyler Helmuth, and Andrew Swan. Random spanning forests and hyperbolic symmetry. Communications in Mathematical Physics, 381 0 (3): 0 1223--1261, February 2021. ISSN 0010-3616, 1432-0916. doi:10.1007/s00220-020-03921-y. URL http://arxiv.org/abs/1912.04854. arXiv:1912.04854 [math]

    work page Pith review arXiv 2021

  5. [5]

    Hyperbolic Geometry in Computer Vision : A Novel Framework for Convolutional Neural Networks , March 2023

    Ahmad Bdeir, Kristian Schwethelm, and Niels Landwehr. Hyperbolic Geometry in Computer Vision : A Novel Framework for Convolutional Neural Networks , March 2023. URL https://arxiv.org/abs/2303.15919v2

    arXiv 2023

  6. [6]

    Multiple Cayley - Klein metric learning

    Yanhong Bi, Bin Fan, and Fuchao Wu. Multiple Cayley - Klein metric learning. PLoS ONE, 12 0 (9): 0 e0184865, September 2017. ISSN 1932-6203. doi:10.1371/journal.pone.0184865. URL https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5608239/

    work page doi:10.1371/journal.pone.0184865 2017

  7. [7]

    Interpreting Neural Networks through the Polytope Lens , November 2022

    Sid Black, Lee Sharkey, Leo Grinsztajn, Eric Winsor, Dan Braun, Jacob Merizian, Kip Parker, Carlos Ramón Guevara, Beren Millidge, Gabriel Alfour, and Connor Leahy. Interpreting Neural Networks through the Polytope Lens , November 2022. URL http://arxiv.org/abs/2211.12312. arXiv:2211.12312 [cs]

    arXiv 2022

  8. [8]

    Random Forests

    Leo Breiman. Random Forests . Machine Learning, 45 0 (1): 0 5--32, October 2001. ISSN 1573-0565. doi:10.1023/A:1010933404324. URL https://doi.org/10.1023/A:1010933404324

    doi:10.1023/a:1010933404324 2001

Show all 38 references
  1. [9]

    Classification and Regression Trees

    Leo Breiman. Classification and Regression Trees . Routledge, New York, October 2017. ISBN 978-1-315-13947-0. doi:10.1201/9781315139470

    2017 doi

  2. [10]

    Hyperbolic Graph Convolutional Neural Networks , October 2019

    Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic Graph Convolutional Neural Networks , October 2019. URL http://arxiv.org/abs/1910.12933. arXiv:1910.12933 [cs, stat]

    2019 arXiv

  3. [11]

    XGBoost : A Scalable Tree Boosting System

    Tianqi Chen and Carlos Guestrin. XGBoost : A Scalable Tree Boosting System . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pp.\ 785--794, August 2016. doi:10.1145/2939672.2939785. URL http://arxiv.org/abs/1603.02754. ar...

    2016

  4. [12]

    Fully Hyperbolic Neural Networks , March 2022

    Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully Hyperbolic Neural Networks , March 2022. URL http://arxiv.org/abs/2105.14686. arXiv:2105.14686 [cs]

    2022 arXiv

  5. [13]

    Fast hyperboloid decision tree algorithms, March 2024

    Philippe Chlenski, Ethan Turok, Antonio Moretti, and Itsik Pe'er. Fast hyperboloid decision tree algorithms, March 2024. URL http://arxiv.org/abs/2310.13841. arXiv:2310.13841 [cs]

    2024 arXiv

  6. [14]

    Khan, Kaizhu Du, Antonio Khalil Moretti, and Itsik Pe'er

    Philippe Chlenski, Quentin Chu, Raiyan R. Khan, Kaizhu Du, Antonio Khalil Moretti, and Itsik Pe'er. Mixed-curvature decision trees and random forests, February 2025 a . URL http://arxiv.org/abs/2410.13879. arXiv:2410.13879 [cs]

    2025 arXiv

  7. [15]

    Manify: A Python Library for Learning Non - Euclidean Representations , March 2025 b

    Philippe Chlenski, Kaizhu Du, Dylan Satow, and Itsik Pe'er. Manify: A Python Library for Learning Non - Euclidean Representations , March 2025 b . URL http://arxiv.org/abs/2503.09576. arXiv:2503.09576 [cs] version: 1

    2025 arXiv

  8. [16]

    Large- Margin Classification in Hyperbolic Space , June 2018

    Hyunghoon Cho, Benjamin DeMeo, Jian Peng, and Bonnie Berger. Large- Margin Classification in Hyperbolic Space , June 2018. URL http://arxiv.org/abs/1806.00437. arXiv:1806.00437 [cs, stat]

    2018 arXiv

  9. [17]

    Hyperbolic Random Forests , August 2023

    Lars Doorenbos, Pablo Márquez-Neila, Raphael Sznitman, and Pascal Mettes. Hyperbolic Random Forests , August 2023. URL http://arxiv.org/abs/2308.13279. arXiv:2308.13279 [cs]

    2023 arXiv

  10. [18]

    Oblique Decision Tree in Python , 2021

    ECNU . Oblique Decision Tree in Python , 2021. URL https://github.com/zhenlingcn/scikit-obliquetree. Publication Title: GitHub repository

    2021

  11. [19]

    Privileged Bases in the Transformer Residual Stream , 2023

    Nelson Elhage, Robert Lasenby, and Christopher Olah. Privileged Bases in the Transformer Residual Stream , 2023. URL https://transformer-circuits.pub/2023/privileged-basis/index.html

    2023

  12. [20]

    Xiran Fan, Chun-Hao Yang, and Baba C. Vemuri. Horospherical Decision Boundaries for Large Margin Classification in Hyperbolic Space , June 2023. URL http://arxiv.org/abs/2302.06807. arXiv:2302.06807 [cs, stat]

    2023 arXiv

  13. [21]

    Hyperbolic Neural Networks , June 2018

    Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic Neural Networks , June 2018. URL http://arxiv.org/abs/1805.09112. arXiv:1805.09112 [cs, stat]

    2018 arXiv

  14. [22]

    Bekkers, and Danica Kragic

    Alejandro García-Castellanos, Aniss Aiman Medbouhi, Giovanni Luca Marchetti, Erik J. Bekkers, and Danica Kragic. HyperSteiner : Computing Heuristic Hyperbolic Steiner Minimal Trees . January 2025. URL http://arxiv.org/abs/2409.05671. arXiv:2409.05671 [cs]

    2025 arXiv

  15. [23]

    LightGBM : A Highly Efficient Gradient Boosting Decision Tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. LightGBM : A Highly Efficient Gradient Boosting Decision Tree . 2017

    2017

  16. [24]

    Khan, Philippe Chlenski, and Itsik Pe'er

    Raiyan R. Khan, Philippe Chlenski, and Itsik Pe'er. Hyperbolic Genome Embeddings . October 2024. URL https://openreview.net/forum?id=NkGDNM8LB0

    2024

  17. [25]

    Isolation Forest

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation Forest . In 2008 Eighth IEEE International Conference on Data Mining , pp.\ 413--422, December 2008. doi:10.1109/ICDM.2008.17. URL https://ieeexplore.ieee.org/document/4781136. ISSN: 2374-8486

    2008

  18. [26]

    Werner, and Dongmian Zou

    Yidan Mao, Jing Gu, Marcus C. Werner, and Dongmian Zou. Klein Model for Hyperbolic Neural Networks , October 2024. URL http://arxiv.org/abs/2410.16813. arXiv:2410.16813 [cs]

    2024 arXiv

  19. [27]

    Hyperbolic Manifold Regression , May 2020

    Gian Maria Marconi, Lorenzo Rosasco, and Carlo Ciliberto. Hyperbolic Manifold Regression , May 2020. URL http://arxiv.org/abs/2005.13885. arXiv:2005.13885 [cs, stat]

    2020 arXiv

  20. [28]

    HEAT : Hyperbolic Embedding of Attributed Networks , May 2019

    David McDonald and Shan He. HEAT : Hyperbolic Embedding of Attributed Networks , May 2019. URL http://arxiv.org/abs/1903.03036. arXiv:1903.03036 [cs]

    2019 arXiv

  21. [29]

    A Wrapped Normal Distribution on Hyperbolic Space for Gradient - Based Learning , May 2019

    Yoshihiro Nagano, Shoichiro Yamaguchi, Yasuhiro Fujita, and Masanori Koyama. A Wrapped Normal Distribution on Hyperbolic Space for Gradient - Based Learning , May 2019. URL http://arxiv.org/abs/1902.02992. arXiv:1902.02992 [cs, stat]

    2019 arXiv

  22. [30]

    Hyperbolic Voronoi diagrams made easy

    Frank Nielsen and Richard Nock. Hyperbolic Voronoi diagrams made easy. In 2010 International Conference on Computational Science and Its Applications , pp.\ 74--80, 2010. doi:10.1109/ICCSA.2010.37. URL http://arxiv.org/abs/0903.3287. arXiv:0903.3287 [cs]

    2010 arXiv

  23. [31]

    Collins, David J

    Mohammad Norouzi, Maxwell D. Collins, David J. Fleet, and Pushmeet Kohli. CO2 Forest : Improved Random Forest by Continuous Optimization of Oblique Splits , June 2015. URL http://arxiv.org/abs/1506.06155. arXiv:1506.06155 [cs]

    2015 arXiv

  24. [32]

    Ratcliffe

    John G. Ratcliffe. Hyperbolic Geometry . In John G. Ratcliffe (ed.), Foundations of Hyperbolic Manifolds , pp.\ 52--96. Springer International Publishing, Cham, 2019. ISBN 9783030315979. doi:10.1007/978-3-030-31597-9_3. URL https://doi.org/10.1007/978-3-030-31597-9_3

    2019 doi

  25. [33]

    Computational Complexity Analysis of Decision Tree Algorithms

    Habiba Muhammad Sani, Ci Lei, and Daniel Neagu. Computational Complexity Analysis of Decision Tree Algorithms . In Max Bramer and Miltos Petridis (eds.), Artificial Intelligence XXXV , pp.\ 191--197, Cham, 2018. Springer International Publishing. ISBN 9783030041915. doi:10.100...

    2018 doi

  26. [34]

    Random Forest regression for manifold-valued responses, February 2017

    Dimosthenis Tsagkrasoulis and Giovanni Montana. Random Forest regression for manifold-valued responses, February 2017. URL http://arxiv.org/abs/1701.08381. arXiv:1701.08381 [stat]

    2017 arXiv

  27. [35]

    D. C. Wickramarachchi, B. L. Robertson, M. Reale, C. J. Price, and J. Brown. HHCART : An Oblique Decision Tree , April 2015. URL http://arxiv.org/abs/1504.03415. arXiv:1504.03415 [stat]

    2015 arXiv

  28. [36]

    D. C. Wickramarachchi, B. L. Robertson, M. Reale, C. J. Price, and J. A. Brown. A reflected feature space for CART . Australian & New Zealand Journal of Statistics, 61 0 (3): 0 380--391, 2019. ISSN 1467-842X. doi:10.1111/anzs.12275. URL https://onlinelibrary.wiley.com/doi/abs/...

    2019 doi

  29. [37]

    Towards kernelizing the classifier for hyperbolic data

    Meimei Yang, Qiao Liu, Xinkai Sun, Na Shi, and Hui Xue. Towards kernelizing the classifier for hyperbolic data. Frontiers of Computer Science, 18 0 (1): 0 181301, August 2023. ISSN 2095-2236. doi:10.1007/s11704-022-2457-y. URL https://doi.org/10.1007/s11704-022-2457-y

    2023 doi

  30. [38]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.