When we visualize a graph, we often imagine nodes scattered across a flat sheet of paper. Early Graph Neural Networks (GNNs) inherited this intuition, embedding nodes into standard flat Euclidean space ($\mathbb{R}^n$). However, real-world networks are rarely flat.
Consider a corporate hierarchy: as you move down the organizational chart, the number of employees grows exponentially. If you try to embed this tree-like structure into a flat space, you quickly run out of room, forcing distant nodes too close together and causing massive distortion. To solve this, researchers turned to hyperbolic space ($\mathbb{H}^n$), which expands exponentially and accommodates hierarchies naturally. Conversely, if your graph contains dense, cyclical friend groups or cliques, spherical space ($\mathbb{S}^n$) provides the perfect wrap-around geometry.
The catch? Real-world graphs are rarely just trees or just cycles. They are a messy patchwork of both. A single social network might contain strict hierarchical corporate ladders, tightly looping friend groups, and regular grid-like road maps. Forcing this topological heterogeneity into any single, uniform geometric space inevitably drops crucial structural information.
To fix this, a new paradigm has emerged: instead of forcing a graph into a one-size-fits-all geometry, models must dynamically adapt their underlying mathematical spaces to match the local shape of the data.
The Limitation of Uniform Curvature
Standard geometric deep learning models assume that a single constant curvature value can describe an entire dataset. As Bachmann et al. point out, this "flat-earth" or single-manifold assumption fails because local graph geometries vary wildly across different regions.
When a model is constrained to a single manifold, it suffers from severe representation distortion. Di Giovanni et al. demonstrated that this structural rigidity forces diverse graph topologies into an ill-fitting geometric framework. For instance, embedding a dense clique into a purely hyperbolic space causes just as much distortion as embedding a tree into a spherical space.
This mismatch degrades performance across a wide variety of graph tasks. Whether trying to detect anomalies with scarce labels (Dong et al.), complete multi-modal knowledge graphs (Gao et al.), or solve complex vehicle routing problems (Liu), constraining the network to a single geometry discards the intrinsic "shape" of the data.
Constructing Mixed-Curvature Product Spaces
The first major breakthrough in tackling topological heterogeneity was the introduction of product manifolds. Instead of choosing between Euclidean, hyperbolic, or spherical spaces, why not combine them?
By taking the Cartesian product of multiple distinct spaces, we can construct a composite latent environment:
$$\mathcal{M} = \mathbb{H}^{d_1} \times \mathbb{S}^{d_2} \times \mathbb{R}^{d_3}$$
Gu et al. pioneered this approach by creating a framework that jointly learns both the optimal curvatures and the allocations of dimensions across these constituent subspaces. If a graph is 70% hierarchical and 30% cyclical, the model automatically allocates more dimensions to the hyperbolic component. This simple shift led to massive reductions in embedding distortion.
Heterogeneous Graph Mixed-Curvature Product Space
/\ (Tree-like) +-----------------------+
/ \ | Hyperbolic Subspace |
/\ /\ | (Hierarchies) |
+-----------------------+
vs. ========> | Spherical Subspace |
| (Cycles/Cliques) |
(O)---(O) (Cyclic) +-----------------------+
| | | Euclidean Subspace |
(O)---(O) | (Flat Grids) |
+-----------------------+
Building on this idea, researchers have scaled these architectures to handle increasingly complex structures. Takeuchi et al. extended core neural operations like matrix-vector multiplication to operate seamlessly across products of multiple hyperbolic spaces. Meanwhile, Cho et al. introduced mixed-curvature Transformers, utilizing kernelized attention mechanisms to let the model automatically discover and adapt to non-Euclidean graph structures without manual geometric tuning.
Adaptive Routing via Mixture of Geometric Experts
While global product spaces offer a richer toolkit, they still apply the same combination of geometries to every node in the graph. However, topological heterogeneity often manifests at the individual node level. A central hub node might sit at the heart of a dense, loopy community (requiring spherical geometry), while a peripheral node sits at the tip of a long, branching tree (requiring hyperbolic geometry).
To address this localized variation, recent work has shifted toward adaptive routing and Mixture of Experts (MoE) frameworks.
Zhu et al. introduced Geometry Interaction Learning (GIL), a framework that embeds graphs in both Euclidean and hyperbolic spaces simultaneously. Crucially, GIL features an interaction module that allows these spaces to communicate, adaptively weighting their contributions for every single node.
Taking this concept further, Guo et al. developed GraphMoRE, which uses a Mixture of Riemannian Experts to assign nodes to personalized combinations of hyperbolic, spherical, and Euclidean spaces based on their immediate neighborhood structures. For graphs that change over time, Yang et al. designed a dynamic graph Transformer that routes nodes to different geometric manifolds on the fly as the network topology evolves, ensuring that the representation space matches the graph's shifting structural needs.
Unifying Geometries in Advanced Manifolds
Instead of stitching separate manifolds together or routing nodes between them, an alternative frontier explores advanced mathematical spaces that inherently unify multiple geometries within a single coordinate system.
One approach utilizes pseudo-Riemannian or ultrahyperbolic geometry. While standard Riemannian manifolds have a strictly positive-definite metric tensor, pseudo-Riemannian manifolds relax this constraint. As Law et al. demonstrated, ultrahyperbolic spaces can simultaneously capture both hierarchical structures and cyclical patterns within a unified space, bypassing the need to manually partition dimensions into separate product spaces.
Another powerful direction leverages matrix manifolds. Cruceru et al. introduced Symmetric Positive Definite (SPD) and Grassmann manifolds for graph embeddings, proving that their non-constant curvature allows for far more faithful representations of complex networks like social graphs and power grids. As Lopez et al. showed, the manifold of SPD matrices naturally unifies both Euclidean and hyperbolic geometries. By developing "gyrocalculus" operations (like addition and scaling) on these matrices, the authors created a continuous space capable of representing diverse structural properties concurrently.
By moving away from rigid, flat assumptions and embracing mixed-curvature spaces, modern graph neural networks can finally respect the true, heterogeneous shapes of real-world data.
Key Foundational Work
Graph Geometry Interaction Learning (2020) — contributed 3 concept elements to this direction.
Constant Curvature Graph Convolutional Networks (2020) — contributed 1 concept elements to this direction.
Ultrahyperbolic Representation Learning (2020) — contributed 1 concept elements to this direction.