Exploring the Poincaré Ellipsis

Research output: Contributions to collected editions/worksArticle in conference proceedingsResearchpeer-review

Standard

Exploring the Poincaré Ellipsis. / Fadel, Samuel; Paulsen, Tino; Brefeld, Ulf.
20th Machine Learning on Graphs workshop. 2023.

Research output: Contributions to collected editions/worksArticle in conference proceedingsResearchpeer-review

Harvard

Fadel, S, Paulsen, T & Brefeld, U 2023, Exploring the Poincaré Ellipsis. in 20th Machine Learning on Graphs workshop. 20th International Workshop on Mining and Learning with Graphs - MLG 2023, Torino, Italy, 22.09.23.

APA

Fadel, S., Paulsen, T., & Brefeld, U. (2023). Exploring the Poincaré Ellipsis. In 20th Machine Learning on Graphs workshop

Vancouver

Fadel S, Paulsen T, Brefeld U. Exploring the Poincaré Ellipsis. In 20th Machine Learning on Graphs workshop. 2023

Bibtex

@inbook{a6e3936be54a42349ae8f58c0592c778,
title = "Exploring the Poincar{\'e} Ellipsis",
abstract = "Exploring different geometries has shown to be useful for exploiting inherent properties of data at hand, becoming attractive to compute embeddings therein. For example, hyperbolic geometry shows superior performance when embedding hierarchical data. This suggests that the ability to explore different geometries for embedding data might be an important consideration, just as the models and algorithms used to perform the embedding. However, utilising non-Euclidean geometries to embed data can sometimes be a laborious task, as the whole process needs to be first analytically derived for each geometry chosen (e.g., spherical, hyperbolic) and then adapted to it. This discourages the exploration of richer spaces to embed data in. Using the framework of Riemannian manifolds, we explore a new choice of geometry. We consider a modified version of the Poincar{\'e} disc to accept individual (perdimension) curvatures instead of a single global one, resulting in what we refer to as the Poincar{\'e} ellipsis. We experiment with link prediction on graph nodes embedded onto these new spaces, showing the performance implications and highlighting the ease to explore new variants of hyperbolic geometry.",
author = "Samuel Fadel and Tino Paulsen and Ulf Brefeld",
year = "2023",
month = sep,
language = "English",
booktitle = "20th Machine Learning on Graphs workshop",
note = "20th International Workshop on Mining and Learning with Graphs - MLG 2023, MLG 2023 ; Conference date: 22-09-2023 Through 22-09-2023",
url = "https://mlg-europe.github.io/2023/",

}

RIS

TY - CHAP

T1 - Exploring the Poincaré Ellipsis

AU - Fadel, Samuel

AU - Paulsen, Tino

AU - Brefeld, Ulf

N1 - Conference code: 20

PY - 2023/9

Y1 - 2023/9

N2 - Exploring different geometries has shown to be useful for exploiting inherent properties of data at hand, becoming attractive to compute embeddings therein. For example, hyperbolic geometry shows superior performance when embedding hierarchical data. This suggests that the ability to explore different geometries for embedding data might be an important consideration, just as the models and algorithms used to perform the embedding. However, utilising non-Euclidean geometries to embed data can sometimes be a laborious task, as the whole process needs to be first analytically derived for each geometry chosen (e.g., spherical, hyperbolic) and then adapted to it. This discourages the exploration of richer spaces to embed data in. Using the framework of Riemannian manifolds, we explore a new choice of geometry. We consider a modified version of the Poincaré disc to accept individual (perdimension) curvatures instead of a single global one, resulting in what we refer to as the Poincaré ellipsis. We experiment with link prediction on graph nodes embedded onto these new spaces, showing the performance implications and highlighting the ease to explore new variants of hyperbolic geometry.

AB - Exploring different geometries has shown to be useful for exploiting inherent properties of data at hand, becoming attractive to compute embeddings therein. For example, hyperbolic geometry shows superior performance when embedding hierarchical data. This suggests that the ability to explore different geometries for embedding data might be an important consideration, just as the models and algorithms used to perform the embedding. However, utilising non-Euclidean geometries to embed data can sometimes be a laborious task, as the whole process needs to be first analytically derived for each geometry chosen (e.g., spherical, hyperbolic) and then adapted to it. This discourages the exploration of richer spaces to embed data in. Using the framework of Riemannian manifolds, we explore a new choice of geometry. We consider a modified version of the Poincaré disc to accept individual (perdimension) curvatures instead of a single global one, resulting in what we refer to as the Poincaré ellipsis. We experiment with link prediction on graph nodes embedded onto these new spaces, showing the performance implications and highlighting the ease to explore new variants of hyperbolic geometry.

UR - https://mlg-europe.github.io/2023/papers/259.pdf

UR - https://mlg-europe.github.io/2023/

M3 - Article in conference proceedings

BT - 20th Machine Learning on Graphs workshop

T2 - 20th International Workshop on Mining and Learning with Graphs - MLG 2023

Y2 - 22 September 2023 through 22 September 2023

ER -

Recently viewed

Publications

  1. Alltag in den Medien - Medien im Alltag
  2. Subjektivierung und teilhabende Kritik
  3. Neue Ansätze dynamischer Mimikerkennung
  4. Empirische Methoden der Deutschdidaktik
  5. Job maintenance by supported employment
  6. Cost-effective Biodiversity Conservation
  7. Rahmenbedingungen der Schulsozialarbeit
  8. (Re)Produktivität im Dialog mit der Praxis
  9. Anwendungsprogrammierung mit Embedded-SQL
  10. Jenseits von Rente und Vollbeschäftigung
  11. The elephant in the room is really a cow
  12. Innovative Museums- und Besucherforschung
  13. Cool weather tourism under global warming:
  14. Fotojournalismus in der Weimarer Republik
  15. Energiewende? Da war doch etwas ...: Kommentar
  16. Die ökonomischen Kosten des Klimawandels
  17. Energiewende erfordert hohe Investitionen
  18. Learning and Re-learning in Chat-based CSCL
  19. Sexualitäten, Geschlechter und Identitäten
  20. Wann werden Serviceleistungen nachgefragt?
  21. Folgen des Klimawandels für den Tourismus
  22. Eine kluge Energiewende ist nicht zu teuer
  23. The counter-revolution is really televised …
  24. Die wirtschaftlichen Folgen von Bahnstreiks
  25. The Origins of Parliamentary Agenda Control
  26. Improving conservation procurement auctions
  27. Why do German physicians reject managed care?
  28. Verteilungseffekte von Kapazitätsmechanismen
  29. Emotion Perception in Human-Robot Interaction
  30. Die Lüneburger Ziegeleien des 19. Jahrhunderts
  31. Gelebte Inklusion oder große Ungerechtigkeit?
  32. Modeling Efficient Grounding in Chat-based CSCL
  33. Österreich-Tourismus – Zurück zum Wachstumskurs
  34. Antonio Negri. Une philosophie de la subversion
  35. Are Self-Employed Really Happier Than Employees?