Parameters and experiment design

Tune one question at a time. Changing preprocessing, sample selection, neighborhood size, and optimization budget together makes it difficult to explain the result.

Question	Controls
How local should the neighborhood be?	n_neighbors in graph methods, perplexity in t-SNE, k in PHATE
How much optimization work is allowed?	n_epochs in UMAP/NeighborMap, n_iter in several Rust methods
How tightly should UMAP pack nearby points?	min_dist and spread
How does a graph layout start?	NeighborMap(init="pca" or "spectral"); UMAP has its own initialization options
How repeatable is the experiment?	Seed, input order, backend, threads, dependency versions

These controls are not interchangeable across methods. A budget of 160 epochs in NeighborMap has no direct equivalence to 160 iterations in t-SNE.

Compare several seeds

Use the repeated-seed runner to retain each result. Compare runtime alongside neighbor recall and rank trustworthiness, and inspect the coordinates. A smaller runtime is not an improvement if the layout loses the structure your task needs.

Components and supported metrics

NeighborMap and SpectralMap produce exactly two dimensions and use Euclidean distance; they have no n_components or metric argument. Most older Rust constructors accept n_components, but they do not expose UMAP's broad metric selection. Refer to each algorithm page for its exact signature.