Radial Distance as a Proxy for Hierarchical Depth

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

Version 1 of 115 papers18 conceptsfrozen 2026-10-04

In traditional machine learning, we usually represent data in Euclidean space—the flat geometry we learned in school. But when we try to map complex, branching structures like social networks, taxonomies, or biological lineages into a flat space, we run into a "crowding problem." Trees expand exponentially as they grow deeper, but the volume of Euclidean space only grows polynomially. To fit a tree into a flat map, we are forced to distort the relationships, crushing the leaves together or stretching the branches.

A growing body of research suggests a more natural home for these structures: hyperbolic geometry. By embedding data into a curved manifold, researchers like Maximilian Nickel and Douwe Kiela have shown that we can capture hierarchical relationships with startling efficiency. The core intuition behind this success is a simple geometric mapping: the radial distance from the origin represents hierarchical depth.

The Geometry of Generality

In a hyperbolic disk (the Poincaré ball model), the space "expands" as you move away from the center. This property mirrors the structure of a tree, where a single root branches into a few nodes, which then branch into many more.

In Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry, Maximilian Nickel and Douwe Kiela demonstrate that the norm of a node's vector—its distance from the origin—effectively represents its "generality." A node at the center of the manifold has the shortest average distance to all other points, making it the natural "root" or "parent." As you move toward the boundary, the space becomes increasingly vast, accommodating the "leaves" or more specialized concepts.

This isn't just a mathematical convenience; it provides a continuous coordinate system for hierarchy. Instead of discrete levels (Level 1, Level 2), we get a smooth gradient of importance or maturity.

Mapping Influence and Popularity

This radial mapping has immediate applications in understanding network influence. In Hyperbolic Graph Neural Networks, Qi Liu and colleagues show that the radial distance naturally encodes a node's centrality. High-importance nodes gravitate toward the center, while peripheral nodes are pushed outward.

This mechanism is particularly powerful in recommendation systems and financial analysis:

  • Recommendation: Naicheng Guo and Ryotaro Shimizu have applied this to session-based and fashion recommendations. In their models, popular or "general" items (like a basic white t-shirt) cluster near the origin, while niche, specialized items reside at the periphery. This allows the model to distinguish between broad trends and specific user tastes based purely on spatial position.
  • Finance: In Transaction Profiling and Address Role Inference in Tokenized U.S. Treasuries, Junliang Luo uses PoincaVec to map blockchain addresses. Institutional treasuries and issuers—the "roots" of the financial network—occupy the center, while retail traders and bots are positioned toward the edges.

Visualizing Biological Time

Perhaps the most striking use of this idea is in the study of cellular development. When cells differentiate, they follow a branching lineage from a single progenitor (the root) to many specialized cell types (the leaves).

Anna Klimovskaia, David López-Paz, and Leon Bottou introduced Poincaré Maps to visualize this process. In their embeddings, the radial distance from the center serves as a "pseudotime" coordinate. The progenitor cells sit at the origin, and as cells mature and differentiate, they move radially outward along different branches.

Unlike Euclidean visualization tools like t-SNE or UMAP, which might scatter these cells based on local similarities, the hyperbolic approach preserves the global structure. It allows a researcher to see both the continuous progression of time (radial distance) and the discrete branching of cell types (angular separation) in a single, low-dimensional view.

Why It Works: The Centrality Constraint

The reason these embeddings are so effective is that they move beyond simple pairwise similarity. Traditional methods like Multi-Dimensional Scaling (MDS) only care about the distance between point A and point B.

In contrast, researchers like Brian Baingana have shown that by explicitly enforcing centrality constraints, we can produce visualizations that are more structurally meaningful. By mapping a node's importance to its distance from the origin, the geometry itself begins to do the work of organization. In hyperbolic space, this "centrality constraint" is not an artificial addition—it is a fundamental property of the manifold's curvature.

By treating the origin as the "top" of the world, hyperbolic embeddings transform undirected similarity data into a directed, navigable hierarchy.

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_radial_distance_as_a_proxy_for_hierarchi_1183bc93,
  title = {Radial Distance as a Proxy for Hierarchical Depth},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/radial-distance-as-a-proxy-for-hierarchical-depth/rdr_1183bc9352704bdc96f5f0e10c710b44},
  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_radial_distance_as_a_proxy_for_hierarchi_1183bc93_v1,
  title = {Radial Distance as a Proxy for Hierarchical Depth},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/radial-distance-as-a-proxy-for-hierarchical-depth/rdr_1183bc9352704bdc96f5f0e10c710b44/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}