Graph Representation Learning in Mixed-Curvature Product Manifolds

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

Version 1 of 138 papers47 conceptsfrozen 2026-10-04

Most machine learning models assume the world is flat. When we embed a graph into a latent space, we typically use Euclidean coordinates—the familiar $x, y, z$ of high school geometry. However, real-world networks are rarely "flat." They are messy, heterogeneous structures containing hierarchies, cycles, and grids all at once.

Forcing these complex topologies into a flat Euclidean space creates distortion. If you try to flatten a globe onto a paper map, the areas near the poles stretch unnaturally; similarly, if you try to flatten a tree-like hierarchy into Euclidean space, the distance between nodes becomes mathematically impossible to preserve.

To solve this, a growing body of research suggests that we should embed graphs into mixed-curvature product manifolds: spaces that combine multiple geometries (Hyperbolic, Spherical, and Euclidean) to match the intrinsic "shape" of the data.

The Geometry of Information

To understand why mixed curvatures matter, we must first look at the three "pure" geometries:

  1. Hyperbolic Space (Negative Curvature): Imagine a space that expands exponentially as you move away from the center. This is the natural home for hierarchies. In a tree structure, the number of nodes grows exponentially with depth; hyperbolic space has the "room" to accommodate this growth without squeezing nodes together.
  2. Spherical Space (Positive Curvature): This is a closed, bounded space. It is ideal for representing cycles and cliques, where entities are interconnected in a ring or a dense cluster.
  3. Euclidean Space (Zero Curvature): This is the standard flat space, best suited for grid-like structures or community clusters with no clear hierarchy.

As Bo Xiong and colleagues point out in Pseudo-Riemannian Graph Convolutional Networks, real graphs are "topologically messy." A social network might have a hierarchical celebrity structure (hyperbolic) alongside tight-knit friend groups (spherical) and regular grid-like professional connections (Euclidean).

Product Manifolds: A Geometric Swiss Army Knife

The core innovation in this field is the use of Product Manifolds. Instead of choosing just one geometry, researchers like Zhen Han in DyERNIE and Li Sun in SELFMGNN construct a representation space as a Cartesian product of different manifolds: $$\mathcal{M} = \mathbb{H}^{n} \times \mathbb{S}^{m} \times \mathbb{E}^{k}$$ In this setup, a single node's embedding is split across these spaces. One part of the vector captures its hierarchical position, another its cyclical relationships, and a third its flat cluster membership. This allows the model to simultaneously capture diverse structural motifs without information loss.

Learning the Shape of the Graph

How does a model know which geometry to use for which part of the graph? There are two primary strategies:

1. Learnable Curvatures and Weights In frameworks like CONGREGATE (Li Sun et al., /paper/art_964459df59814a6e8c9f6ab2829dbede), the curvature of each factor manifold is not fixed. The model treats the curvature as a trainable parameter, allowing the "shape" of the latent space to deform during training to better fit the graph's Ricci curvature.

2. Geometric Routing An alternative approach, R-SGFormer by Ankit Jyothish (/paper/art_38a1990fc16d4fbea6162e2ee97a05bc), uses a "routing" mechanism. Much like a Mixture-of-Experts (MoE) layer, the model analyzes the local topology around a node and routes it to a specialized "expert" manifold (e.g., hyperbolic for a leaf node, spherical for a hub node) to maximize representation quality.

Applications: From Biology to Anomalies

The benefits of mixed-curvature embeddings are most visible in domains where data is inherently multi-faceted:

  • Biological Pathways: Daniel McNeela and Anthony Gitter demonstrated in /paper/art_360e3a272ea042c2a331ffbc3a9d8318 that biological interactions—which involve both hierarchical signaling and cyclical feedback loops—suffer significantly less distortion when embedded in product manifolds compared to flat spaces.
  • Knowledge Graph Completion: Models like IME (Jiapu Wang et al., /paper/art_decdcac742e24928bbb067a7789d0d5b) use "shared and specific" embeddings across Euclidean, Hyperbolic, and Spherical spaces to predict missing facts in temporal knowledge graphs, which often contain a mix of linear chains and complex hierarchies.
  • Anomaly Detection: Karish Grover and Geoffrey J. Gordon introduced CurvGAD (/paper/art_49a5311a0718478c8ea5eb2483628338), which identifies "geometric anomalies"—nodes that don't fit the local curvature of their neighborhood—providing a new signal for fraud or error detection that traditional feature-based methods miss.

Beyond Riemannian Geometry

While most work focuses on standard Riemannian manifolds, some researchers are pushing into even more exotic territories. Aaron Sim and colleagues (/paper/art_f280d93d387043b08a196fb6b622ef64) have explored Pseudo-Riemannian manifolds (spacetimes). By using an indefinite metric—similar to Einstein's general relativity—they can model directed graphs and non-transitive relations (where A relates to B, and B to C, but A does not relate to C) more effectively than any curved Riemannian space.

By matching the geometry of the embedding space to the topology of the data, these methods move us closer to a "physics of information," where the shape of the space itself does the heavy lifting of representation.

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_graph_representation_learning_in_mixed_c_9b8bccb0,
  title = {Graph Representation Learning in Mixed-Curvature Product Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-representation-learning-in-mixed-curvature-product-manifolds/rdr_9b8bccb0cbd140f9810dc211cc76168a},
  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_graph_representation_learning_in_mixed_c_9b8bccb0_v1,
  title = {Graph Representation Learning in Mixed-Curvature Product Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-representation-learning-in-mixed-curvature-product-manifolds/rdr_9b8bccb0cbd140f9810dc211cc76168a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}