Data and preprocessing

Dense arrays

For the Rust algorithms, supply a finite NumPy float64 matrix of shape (n_samples, n_features). Use at least a few samples and choose a neighbor count smaller than the sample count. Validation differs between older implementations; check inputs before calling the backend.

import numpy as np
from sklearn.datasets import load_digits

X = np.ascontiguousarray(load_digits().data, dtype=np.float64)
assert X.ndim == 2 and X.shape[0] > 2 and X.shape[1] > 0
assert np.isfinite(X).all()

Feature scales influence Euclidean distance. If you standardize or normalize data, fit that preprocessing on the training split and reuse it for evaluation. The published Digits and Fashion-MNIST runs intentionally use raw pixel features; preprocessed results belong in a separately labeled comparison.

Sparse input

UMAP has sparse input paths. The public Rust reducers documented here consume dense matrices. A sparse HNSW implementation also exists internally, but that does not make every reducer sparse-compatible. Estimate dense memory before calling .toarray() on a large matrix.

Labels

Benchmark labels stratify Fashion-MNIST sampling and support evaluation. They do not enter the unsupervised embedding fit. UMAP separately accepts supervised labels via fit_transform(X, y); record that explicitly when comparing results.

Missing values and unusual magnitudes

Resolve NaN and infinity before embedding. NeighborMap and SpectralMap reject non-finite inputs and rescale by a single maximum absolute value internally. That preserves Euclidean neighbor ordering; it does not standardize each feature. The distance kernels use float32 internally in several methods despite float64 Python input, so precision requirements should be validated for your data.