ricci 0.8.0

Graph neural network primitives: GCN and hyperbolic GCN on Burn tensors
Documentation

ricci

crates.io Documentation CI

Graph neural network layers, plus Ollivier-Ricci edge curvature (the curvature module) for finding oversquashing bottlenecks.

Quickstart

[dependencies]
ricci = "0.8"
burn = { version = "0.20", default-features = false, features = ["std"] }
burn-ndarray = "0.20"

Hyperbolic distance on the Poincare ball:

use burn::tensor::{backend::Backend, TensorData};
use burn_ndarray::NdArray;
use ricci::PoincareBall;

type B = NdArray<f32>;
let dev = <B as Backend>::Device::default();
let ball = PoincareBall::new(1.0);

let x = burn::tensor::Tensor::<B, 2>::from_data(
    TensorData::new(vec![0.10f32, 0.00, 0.00], [1, 3]), &dev,
);
let y = burn::tensor::Tensor::<B, 2>::from_data(
    TensorData::new(vec![0.00f32, 0.10, 0.00], [1, 3]), &dev,
);
let d = ball.distance(x, y).to_data().to_vec::<f32>().unwrap()[0];
assert!(d >= 0.0);

Geometry

The Poincare ball $\mathbb{B}^d_c = {x \in \mathbb{R}d : c\lVert x \rVert2 < 1}$ with curvature $-c$:

Operation Formula
Distance $d_c(x, y) = \frac{2}{\sqrt{c}} \text{arctanh}\bigl(\sqrt{c}\lVert -x \oplus_c y \rVert\bigr)$
Mobius addition $x \oplus_c y = \frac{(1 + 2c\langle x,y\rangle + c\lVert y\rVert2)x + (1 - c\lVert x\rVert2)y}{1 + 2c\langle x,y\rangle + c2\lVert x\rVert2\lVert y\rVert^2}$
Exp map $\exp_xc(v) = x \oplus_c \bigl(\tanh\bigl(\frac{\sqrt{c}\lambda_xc\lVert v\rVert}{2}\bigr)\frac{v}{\sqrt{c}\lVert v\rVert}\bigr)$
Log map $\log_xc(y) = \frac{2}{\sqrt{c}\lambda_xc}\text{arctanh}(\sqrt{c}\lVert -x \oplus_c y\rVert)\frac{-x \oplus_c y}{\lVert -x \oplus_c y\rVert}$

where $\lambda_xc = \frac{2}{1 - c\lVert x\rVert2}$ is the conformal factor.

API surface

  • ricci::PoincareBall: Poincare ball geometry (project, mobius_add, exp/log maps, distance, parallel transport).
  • ricci::GCNConv: graph convolution (linear projection + adjacency matmul).
  • ricci::HGCNConv: hyperbolic graph convolution on the Poincare ball.
  • ricci::RGCNConv: relational graph convolution (per-relation transforms over an adjacency stack, optional basis decomposition) for typed graphs.
  • ricci::NBFConv: conditional message passing (edge-type representations as forward inputs; indicator-initialized pair representations).
  • ricci::relgraph: the graph of relations (four interaction-type adjacencies over relation nodes, inverses included). All conv layers derive Burn's Module, so they embed in trainable models.
  • ricci::curvature: Ollivier-Ricci edge curvature over an adjacency matrix (lazy-walk alpha, entropic W1).
  • ricci::features: homomorphism-count node features (walk and closed-walk profiles); these separate some graphs that 1-WL message passing cannot.

Inputs are shaped [batch, d] (row-major feature vectors).

Examples

See examples/README.md for runnable examples with captured output.

References

Each entry links to a mechanism-level summary in docs/papers.md.

  • Ollivier. Ricci curvature of Markov chains on metric spaces. Journal of Functional Analysis 256(3), 2009. The edge curvature computed here. notes
  • Topping, Di Giovanni, Chamberlain, Dong, Bronstein. Understanding over-squashing and bottlenecks on graphs via curvature. ICLR 2022. arXiv:2111.14522. Negative curvature marks the bottleneck edges. notes
  • Kipf, Welling. Semi-supervised classification with graph convolutional networks. ICLR 2017. arXiv:1609.02907. GCNConv. notes
  • Ganea, Bécigneul, Hofmann. Hyperbolic neural networks. NeurIPS 2018. arXiv:1805.09112. The Möbius operations behind PoincareBall. notes
  • Chami, Ying, Ré, Leskovec. Hyperbolic graph convolutional neural networks. NeurIPS 2019. arXiv:1910.12933. HGCNConv. notes
  • Cuturi. Sinkhorn distances: lightspeed computation of optimal transportation distances. NeurIPS 2013. arXiv:1306.0895. The entropic W1 solved per edge. notes
  • Dell, Grohe, Rattan. Lovász meets Weisfeiler and Leman. ICALP 2018. arXiv:1802.08876. Homomorphism counts as an expressiveness measure. notes
  • Barceló, Geerts, Reutter, Ryschkov. Graph neural networks with local graph parameters. NeurIPS 2021. arXiv:2106.06707. Hom-count features in practice. notes
  • Schlichtkrull, Kipf, Bloem, van den Berg, Titov, Welling. Modeling relational data with graph convolutional networks. ESWC 2018. arXiv:1703.06103. RGCNConv. notes
  • Zhu, Zhang, Xhonneux, Tang. Neural Bellman-Ford networks: a general graph neural network framework for link prediction. NeurIPS 2021. arXiv:2106.06935. NBFConv. notes
  • Galkin, Yuan, Mostafa, Tang, Zhu. Towards foundation models for knowledge graph reasoning. ICLR 2024. arXiv:2310.04562. relgraph. notes

License

MIT OR Apache-2.0