Native Classification in Hyperbolic Space

This direction is merged into another direction: hyperbolic embedding. Its last version stays readable and citable here.

Version 1 of 127 papers33 conceptsfrozen 2026-10-04

Most machine learning models assume that data lives in a "flat" Euclidean space. However, many real-world datasets—social networks, phylogenetic trees, and language taxonomies—are inherently hierarchical. When we try to force these branching structures into Euclidean space, we encounter a fundamental geometric mismatch: the number of nodes in a tree grows exponentially with depth, but the volume of Euclidean space only grows polynomially with radius. This leads to "crowding," where the representation of the hierarchy becomes distorted.

Hyperbolic geometry offers a solution. Because hyperbolic space has constant negative curvature, its volume grows exponentially with its radius, providing a "natural home" for hierarchical data. To fully exploit this geometry, researchers have moved beyond simple embeddings to develop native classification mechanisms that operate directly within the manifold.

The Geometry of the Decision Boundary

In standard neural networks, classification is performed by a linear layer that defines a flat hyperplane to separate classes. In hyperbolic space, a "straight" line is a geodesic—the shortest path between two points—which appears curved to our Euclidean eyes.

To perform classification natively, Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann introduced Hyperbolic Neural Networks (HNNs). They redefined the basic building blocks of deep learning—like multinomial logistic regression (MLR)—using Möbius gyrovector spaces. In this framework, a decision boundary is not a flat plane but a gyroplane (or hyperbolic hyperplane). These boundaries respect the underlying curvature, allowing the model to partition the space in a way that aligns with the branching structure of the data.

Why the Classifier Geometry Matters

It is tempting to think that one could simply embed data into hyperbolic space and then use a standard Euclidean classifier on the resulting vectors. However, Federico Lopez and Michael Strube 0001 demonstrated that the geometry of the final classification layer is often more critical than the geometry of the encoder.

When classifying fine-grained entities, replacing a hyperbolic classifier with a Euclidean one leads to a significant drop in accuracy, even if the rest of the network remains hyperbolic. This is because Euclidean hyperplanes cannot effectively capture the "parent-child" relationships inherent in the embedding. As Simon Weber and colleagues noted, moving to the Poincaré ball helps "flatten" the parent bias that plagues Euclidean models, leading to better generalization and more reliable confidence scores.

Efficient Decision Logic in the Lorentz Model

While the Poincaré ball is intuitive for visualization, the Lorentz (or Hyperboloid) model is often preferred for computation. In this model, hyperbolic space is represented as a sheet of a hyperboloid in a higher-dimensional Minkowski space.

Philippe Chlenski, Ethan Turok, and Itsik Pe'er leveraged this to create fast decision tree algorithms. By defining decision boundaries as the intersection of the hyperboloid with homogeneous hyperplanes passing through the origin, they achieved constant-time decision logic. The classification of a point is determined simply by the sign of the Minkowski inner product between the data point and the hyperplane normal. This approach allows for native classification without the need to project data back into Euclidean space, which would introduce distortion.

Beyond Geodesics: Horospheres and Ideal Boundaries

Standard geodesic boundaries are not always optimal. Xiran Fan, Chun-Hao Yang, and Baba C. Vemuri pointed out that geodesic-based optimization can be non-convex and difficult to solve. They proposed HoroSVM, which uses horospheres—surfaces whose normal geodesics all converge asymptotically to the same point—as decision boundaries. This shift transforms the classification task into a geodesically convex optimization problem, guaranteeing a global optimum. This concept was further extended to ensemble learning by Lars Doorenbos and Pascal Mettes to create Hyperbolic Random Forests.

Another powerful mechanism involves the "edge" of the space. Mina Ghadimi Atigh and Pascal Mettes introduced Hyperbolic Busemann Learning, which places class prototypes on the ideal boundary (the infinite limit of the space). By measuring the distance to these "points at infinity" using Busemann functions, the model can discover hierarchical structures even when explicit label taxonomies are unavailable.

Applications and Impact

The shift toward native hyperbolic classification has yielded significant gains across diverse fields:

  • Genomics: Raiyan R. Khan and colleagues showed that Hyperbolic CNNs outperform massive DNA language models on genomic tasks by better representing the tree-like nature of genetic evolution.
  • Computer Vision: Pascal Mettes demonstrated that hyperbolic MLR allows for superior few-shot learning and uncertainty quantification.
  • Audio: Darius Petermann and Jonathan Le Roux used hyperbolic manifolds to capture hierarchical relationships between sound sources, providing a "free" measure of model uncertainty.

By aligning the geometry of the classifier with the geometry of the data, these methods ensure that the structural information preserved during feature extraction is fully utilized in the final decision.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_native_classification_in_hyperbolic_spac_f75ee150,
  title = {Native Classification in Hyperbolic Space},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/native-classification-in-hyperbolic-space/rdr_f75ee150f70840a7b0bee970e98d4a79},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_native_classification_in_hyperbolic_spac_f75ee150_v1,
  title = {Native Classification in Hyperbolic Space},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/native-classification-in-hyperbolic-space/rdr_f75ee150f70840a7b0bee970e98d4a79/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}