Mixed-Curvature Manifolds for Heterogeneous Graph Embeddings

This direction is merged into another direction: product manifold representation. Its last version stays readable and citable here.

Version 1 of 114 papers15 conceptsfrozen 2026-10-04

Most graph representation learning methods assume that the data lives in a uniform geometric space, usually a flat Euclidean one. However, real-world networks are rarely uniform. A single social network might contain hierarchical "tree-like" structures in its periphery, dense "clique-like" communities in its center, and regular "grid-like" patterns in its infrastructure.

Forcing these diverse structures into a single manifold creates distortion. If you try to flatten a tree into a 2D plane, the branches overlap; if you try to wrap a grid around a sphere, the distances stretch. To solve this, a new research direction focuses on embedding graphs into mixed-curvature manifolds that adapt to the local geometry of the data.

The Geometry of Graph Topology

To understand why mixed curvature works, we must first recognize that different graph patterns have natural geometric counterparts:

  • Hyperbolic Space (Negative Curvature): Ideal for hierarchical or tree-like structures. In hyperbolic space, the "volume" grows exponentially with the radius, matching how the number of nodes in a tree grows with depth.
  • Spherical Space (Positive Curvature): Ideal for dense, cyclical clusters or cliques where nodes are highly interconnected and "wrap around" each other.
  • Euclidean Space (Zero Curvature): Ideal for regular, grid-like structures or flat clusters.

As Francesco Di Giovanni and Michael M. Bronstein argue in Heterogeneous manifolds for curvature-aware graph embedding, real networks are "geometric patchworks." A model that uses only one curvature is like trying to map the entire Earth using only a flat piece of paper—you can get the local neighborhood right, but the global structure will be fundamentally warped.

Adaptive Local Selection

One way to handle this heterogeneity is to let the model choose the best space for each node. Feng Ji and colleagues introduced JSGNN (Node-Specific Space Selection via Localized Geometric Hyperbolicity in Graph Neural Networks), which calculates a "hyperbolicity" score for each node’s neighborhood. If a region looks like a tree, the model embeds it in hyperbolic space; if it looks like a grid, it stays in Euclidean space.

Similarly, the Geometry Interaction Learning (GIL) framework by Shichao Zhu and Shirui Pan allows nodes to exist in both Euclidean and hyperbolic spaces simultaneously. An interaction module then adaptively weights these representations, letting the two geometries "talk" to each other to capture mixed structural features that neither could represent alone.

Unified and Continuous Curvature

Rather than picking between discrete spaces, some researchers propose using more flexible, unified mathematical frameworks.

Federico Lopez and Anna Wienhard explore Siegel spaces in Symmetric Spaces for Graph Embeddings: A Finsler-Riemannian Approach. Siegel spaces are a type of symmetric space that can naturally encompass both hyperbolic and Euclidean geometries. By using Finsler metrics (which generalize how we measure distance), they create a representation space that automatically bends to fit the graph's mixed requirements.

Taking this a step further, Chris Ding and Jicong Fan developed Adaptive Riemannian Graph Neural Networks (ARGNN) (Adaptive Riemannian Graph Neural Networks). Instead of choosing a pre-defined manifold, ARGNN learns a unique metric tensor for every node. This allows the geometry to vary continuously across the graph's landscape, effectively "warping" the space locally to match the density and connectivity of the data.

Why It Works: Precision and Generalization

The intuition behind the success of these methods is structural alignment. When the geometry of the embedding space matches the intrinsic geometry of the data:

  1. Distortion drops: Distances between nodes in the embedding space more accurately reflect their true shortest-path distances in the graph.
  2. Efficiency increases: You can represent complex hierarchies in low-dimensional hyperbolic space that would require hundreds of dimensions in Euclidean space.
  3. Generalization improves: As Xinyu Zhao and Qingyun Sun demonstrate in their work on anomaly detection (Zero-shot Generalizable Graph Anomaly Detection with Mixture of Riemannian Experts), models that understand these geometric nuances are better at identifying "out-of-place" structures (anomalies) and generalizing across different types of networks, such as moving from social graphs to financial transaction webs.

By moving away from the "flat earth" assumption of Euclidean GNNs, mixed-curvature methods provide a more faithful lens through which to view the complex, multi-faceted topology of the real world.

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_mixed_curvature_manifolds_for_heterogene_9c1c5607,
  title = {Mixed-Curvature Manifolds for Heterogeneous Graph Embeddings},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mixed-curvature-manifolds-for-heterogeneous-graph-embeddings/rdr_9c1c5607bf3a4f7881bc317cc091fa72},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_mixed_curvature_manifolds_for_heterogene_9c1c5607_v1,
  title = {Mixed-Curvature Manifolds for Heterogeneous Graph Embeddings},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mixed-curvature-manifolds-for-heterogeneous-graph-embeddings/rdr_9c1c5607bf3a4f7881bc317cc091fa72/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}