embedor 0.1.1

Curvature-based stochastic neighbor embeddings for cluster-preserving visualization

Official Implementation of the paper: "EmbedOR: Provable Cluster-Preserving Visualizations with Curvature-Based Stochastic Neighbor Embeddings"

Getting Started

Installation

Create and activate a virtual environment:

python3 -m venv env
source env/bin/activate

Upgrade pip and install the published package from PyPI:

pip install --upgrade pip
pip install embedor

requirements.txt tracks the third-party packages imported from src/, including the optional biological-data and visualization helpers in src/data/data.py and src/vis.py.

Quick Start

The repository currently exposes a Python API rather than a checked-in CLI script. A minimal end-to-end example is:

from src.data.data import moons
from src.embedor import EmbedOR
from src.plotting import plot_data_2D

dataset = moons(n_points=1000, noise=0.05)
X = dataset["data"]
labels = dataset["cluster"]

embedding = EmbedOR(
	nng_params={"mode": "nbrs", "n_neighbors": 15},
	edge_weight="orc",
	layout="torch",
	seed=42,
).fit_transform(X)

plot_data_2D(embedding, color=labels, title="EmbedOR embedding")

Project Layout

• src/embedor.py contains the main EmbedOR estimator.
• src/data/data.py contains synthetic data generators and dataset loaders.
• src/plotting.py contains lightweight plotting helpers for embeddings and graphs.
• src/vis.py contains additional visualization and RNA-velocity utilities.
• src/utils/ contains graph construction, curvature, layout, and timing helpers.

If you want a scriptable workflow, add a small runner that imports EmbedOR from src.embedor and passes either a NumPy array or a precomputed adjacency matrix into fit/fit_transform.