Crate ricci 

ricci: graph learning primitives on Burn tensors.

Provides Poincare ball geometry and graph convolution layers that run on any Burn backend (ndarray, wgpu, tch, etc.).

• PoincareBall: hyperbolic geometry ops (project, mobius_add, exp/log maps, hyperbolic activation via PoincareBall::hyp_act)
• GCNConv: graph convolutional layer (linear + adjacency matmul)
• HGCNConv: hyperbolic graph convolution on the Poincare ball, with optional activation via HGCNConv::forward_act
• curvature: Ollivier-Ricci edge curvature (the crate’s namesake), the primitive behind curvature-based rewiring of oversquashed graphs
• features: homomorphism-count node features (walks + closed walks), the interpretable lift past 1-WL expressiveness

Re-exports

pub use curvature::ollivier_ricci_curvatures;

pub use curvature::CurvatureConfig;

pub use curvature::CurvatureError;

pub use curvature::EdgeCurvature;

pub use features::hom_profile;

pub use hyperbolic::PoincareBall;

pub use nn::GCNConv;

pub use nn::HGCNConv;

pub use nn::NBFConv;

pub use nn::RGCNConv;

Modules

curvature

Ollivier-Ricci edge curvature for graphs.

features

Homomorphism-count node features: provable expressiveness beyond 1-WL.

hyperbolic

Poincare ball geometry on Burn tensors.

nn

Graph neural network layers on Burn tensors.

relgraph

The graph of relations: four interaction-type adjacencies over relation nodes (Galkin et al., ICLR 2024).