Most machine learning models assume the world is flat. When we embed a graph into a latent space, we typically use Euclidean geometry, where the shortest path between two points is a straight line and the angles of a triangle always sum to 180 degrees. However, real-world data—especially relational data like social networks, citation graphs, and biological molecules—is rarely "flat." It possesses intrinsic structures like hierarchies and cycles that Euclidean space struggles to represent without significant distortion.
The Geometry of Topology
To understand why flat space fails, consider two common structures:
- Hierarchies (Trees): In a tree, the number of nodes grows exponentially with depth. Euclidean space only grows polynomially with radius. To fit a tree into Euclidean space, you have to "crush" the leaves together. Hyperbolic space, which has constant negative curvature, expands exponentially, making it a natural home for hierarchical data. Emile Mathieu and colleagues demonstrated this with the Poincaré Variational Auto-Encoder, showing that a hyperbolic latent space captures hierarchies far more efficiently than a flat one.
- Cycles and Grids: Conversely, spherical space (positive curvature) is ideal for modeling cyclical or periodic patterns.
The problem is that real-world graphs are "topologically messy." A single graph might contain a family tree (hyperbolic), a grid-like road network (Euclidean), and a set of cyclical chemical bonds (spherical).
Mixing Curvatures
To handle this complexity, researchers like Li Sun have proposed mixed-curvature spaces. Instead of picking one geometry, models like SELFMGNN use a "product manifold"—a combination of hyperbolic, spherical, and Euclidean components.
By projecting different parts of the graph into the component that fits best, these models avoid the "over-squashing" problem described by Li Sun in DeepRicci, where information is lost as it's forced through narrow geometric bottlenecks. Similarly, Quentin Chu extended this logic to classical algorithms, creating Mixed-curvature decision trees that can split data along curved boundaries.
Indefinite Metrics and Ultrahyperbolic Spaces
While product manifolds combine separate spaces, another approach uses Pseudo-Riemannian geometry to create a single, unified space with an "indefinite metric." In these spaces, the "distance" between points can be positive, zero, or even negative.
Bo Xiong and Chuan Zhou have pioneered Pseudo-Riemannian Graph Convolutional Networks. These models use "ultrahyperbolic" geometry to interleave hyperbolic and spherical properties within the same manifold. This allows the model to represent hierarchical and cyclical relations simultaneously without needing to explicitly partition the data into different subspaces. In Ultrahyperbolic Knowledge Graph Embeddings, this approach proved superior for knowledge graphs that contain both strict taxonomies and symmetric, non-hierarchical relations.
The Search for the Geometric Signature
The most significant hurdle in this field is determining the geometric signature: exactly how many hyperbolic vs. spherical dimensions does a specific dataset need?
Currently, this is often a manual process of trial and error. However, new methods are emerging to automate this discovery:
- Distance Metrics: Alvaro Arroyo and Haitz Sáez de Ocáriz Borde proposed using Gromov-Hausdorff distances to compare different product manifolds, providing a mathematical framework to search for the optimal latent configuration.
- Learnable Manifolds: Qiyu Kang introduced HamGNN, which treats the embedding process as a physical system. Instead of pre-defining the geometry, the model learns the "energy landscape" of the graph, effectively allowing the manifold to shape itself around the data's intrinsic topology during training.
By moving from fixed Euclidean grids to flexible, curved, and mixed geometries, these researchers are enabling neural networks to "see" the true shape of complex relational data.