Choose an algorithm

Squeeze has eleven core reduction methods. Compare an inexpensive linear baseline with methods matching the structure you care about, then measure quality and runtime. The API differences below are part of the contract.

Method	Implementation	Fit new coordinates	Native transform	Main purpose
PCA	Rust	fit_transform; fit	Yes	Linear variance baseline
UMAP	Python/Numba + optional Rust search	fit_transform; fit	Yes, mode-dependent	Neighbor graph layout
t-SNE	Rust	fit_transform	No	Local probability layout
MDS	Rust	fit_transform; from distances	No	Distance fitting
Isomap	Rust	fit_transform	No	Graph geodesics
LLE	Rust	fit_transform	No	Local linear relations
PHATE	Rust	fit_transform	No	Diffusion geometry
TriMap	Rust	fit_transform	No	Triplet constraints
PaCMAP	Rust	fit_transform	No	Pairwise layout
NeighborMap	Experimental Rust	fit_transform	No	Sampled graph refinement
SpectralMap	Experimental Rust	fit_transform	No	Approximate spectral layout

NeighborMap and SpectralMap are Euclidean, two-dimensional methods. The other constructors expose n_components, but that does not establish identical behavior or performance at every dimension. Rust reducers do not implement sklearn's full parameter/cloning interface. Composition wrappers have additional limits when they contain fit-only components.

Start from your requirement

• For a reusable linear projection, use PCA as a baseline.
• For neighborhood visualization, compare UMAP, t-SNE, PaCMAP and NeighborMap.
• For global distances or graph geometry, inspect MDS, Isomap, LLE and PHATE.
• For an inexpensive approximate graph layout, evaluate SpectralMap's quality tradeoff.

There is no supported claim here that every Rust method is faster than every alternative. The Digits and Fashion-MNIST heatmaps show measured configurations, including weak results for the current TriMap implementation. Do not infer large-dataset scalability from these small benchmarks.