Differentiable Hyperbolicity

Official implementation of the paper “Bridging Arbitrary and Tree Metrics via Differentiable Gromov Hyperbolicity” (accepted to NeurIPS 2025). The library provides differentiable estimators of Gromov hyperbolicity and several tree-metric fitting algorithms for graph data.

• Paper: Bridging Arbitrary and Tree Metrics via Differentiable Gromov Hyperbolicity (NeurIPS 2025). The code in this repository was used to produce the experiments in the paper.
• If you build on this work, please cite the paper above.

Installation

pip install -e .

What’s inside

• Differentiable hyperbolicity computation with smooth max/min (differentiable_hyperbolicity/delta.py).
• Tree fitting methods: DeltaZero (differentiable), Neighbor Joining, TreeRep, HCCRootedTreeFit, Gromov tree construction, and Layering approximation (differentiable_hyperbolicity/tree_fitting_methods/).
• Experiment scripts used in the paper (expes/) and preprocessed distance matrices (datasets/).

Quick start

import torch
import networkx as nx
from differentiable_hyperbolicity.delta import compute_hyperbolicity
from differentiable_hyperbolicity.tree_fitting_methods.deltazero import deltazero_tree

# Build a toy graph and its distance matrix
graph = nx.random_geometric_graph(20, 0.35, seed=0)
distances = nx.floyd_warshall_numpy(graph)
distances_t = torch.tensor(distances, dtype=torch.float32)

# Fit a tree metric with the differentiable DeltaZero method
tree = deltazero_tree(
    distances=distances_t,
    root=0,
    lr=0.1,
    n_batches=1,
    batch_size=64,
    scale_delta=1e-2,
    distance_reg=1.0,
    num_epochs=200,
    gpu=False,
    verbose=False,
)

Reproducing paper experiments

• Datasets used in the paper are stored in datasets/.
• Benchmark scripts live in expes/; for example:

python expes/benchmark.py --dataset cora --methods hcc gromov treerep layering --num_expe 20

Repository structure

.
├── differentiable_hyperbolicity/   # Core library (delta computation, utilities, tree fitting)
├── datasets/                       # Preprocessed distance matrices and raw graphs
└── expes/                          # Experiment scripts and results

License

This project is licensed under the MIT License.