UMAP

Fuzzy neighbor graph and optimized layout · Python/Numba with optional Rust search

Squeeze's UMAP implementation descends from umap-learn. It builds a weighted neighborhood graph and optimizes coordinates to represent that graph. Unlike the fit-only Rust reducers, it implements sklearn-style parameter access and fitted transforms. Its broader interface does not apply automatically to other algorithms.

Main controls

Parameter	Default	Role
n_neighbors	15	Neighborhood scale
n_components	2	Output dimension
metric	"euclidean"	Original-space distance
min_dist	0.1	Low-dimensional packing
spread	1.0	Scale paired with min_dist
n_epochs	None	Automatic optimization budget unless set
random_state	None	Reproducible seed when supplied
n_jobs	-1	Worker request; seeded execution can restrict parallelism
use_hnsw	None	Neighbor-backend selection on approximate-search paths

For the complete constructor and its additional fields, use the API reference.

Fit and transform

import numpy as np
from sklearn.datasets import load_digits
from squeeze import UMAP

X = np.asarray(load_digits().data[:120], dtype=np.float64)
model = UMAP(n_neighbors=10, n_epochs=50, random_state=42, n_jobs=1)
training = model.fit_transform(X[:100])
held_out = model.transform(X[100:])
assert training.shape == (100, 2)
assert held_out.shape == (20, 2)

fit(X) returns the estimator; embedding_ holds the fitted coordinates. fit_transform(X, y) can use labels for supervised embedding. Record supervised fits separately from unsupervised benchmark results. inverse_transform approximates original features for supported modes; it does not recover discarded information.

Backend selection and variants

use_hnsw=False selects PyNNDescent for approximate search. True requests Rust HNSW but can fall back for an unavailable backend or unsupported metric. Small-data paths may use direct pairwise calculations. See performance.

UMAP also exposes density, precomputed-neighbor, and supervised options. Related aligned and parametric classes are covered under UMAP variants. Use the installed API's capabilities rather than assuming every upstream example or third-party UMAP integration works with Squeeze.