Struct UmapSphereProjection 

pub struct UmapSphereProjection { /* private fields */ }

UMAP-style projection that lives on S² and transforms new points by kNN-weighted averaging over the fitted positions.

Implementations

impl UmapSphereProjection

pub fn fit_default(embeddings: &[Embedding]) -> Result<Self, ProjectionError>

Fit with default config and no categories.

pub fn fit_from_graph( graph: &UmapGraph, categories: Option<&[u32]>, radial: RadialStrategy, config: UmapConfig, ) -> Result<Self, ProjectionError>

Optimize from a prebuilt kNN graph. This is the cheap part of UMAP fit — O(N·k·epochs) for the Adam optimizer. The graph is not rebuilt.

Use this when the tuner has already built the graph via UmapGraph::build and is sweeping n_epochs / category_weight.

pub fn fit( embeddings: &[Embedding], categories: Option<&[u32]>, radial: RadialStrategy, config: UmapConfig, ) -> Result<Self, ProjectionError>

Fit with custom config. categories is parallel to embeddings when supplied; pass None to disable the supervised term even if config.category_weight > 0.

Equivalent to UmapGraph::build followed by Self::fit_from_graph — the tuner calls those two halves directly so it can reuse graphs across configs that share n_neighbors; this entry point serves every other caller.

pub fn explained_variance_ratio(&self) -> f64

Post-fit quality: trustworthiness-style kNN recall. For each point, its k nearest neighbors among the fitted 3D positions (k = the graph’s n_neighbors) are intersected with its original-space kNN set; the score is the mean overlap fraction. Bounded [0, 1], where 1.0 means every original neighborhood survives the projection.

Intentionally exposed under the EVR name so the auto-tuner’s MetaModel consumers can compare projection kinds on one scalar. Note the semantics changed: this used to be the fraction of kNN edges shorter than the median random pairwise distance, a bar that random spherical pairs (≈90° apart) made trivially clearable, so scores saturated high and barely discriminated. Recall is rank-meaningful across corpora and more honest when compared against other projection kinds.

Trait Implementations

impl Clone for UmapSphereProjection

fn clone(&self) -> UmapSphereProjection

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl From<UmapSphereProjection> for ConfiguredProjection

fn from(p: UmapSphereProjection) -> Self

Converts to this type from the input type.

impl Projection for UmapSphereProjection

fn project(&self, embedding: &Embedding) -> SphericalPoint

fn project_rich(&self, embedding: &Embedding) -> ProjectedPoint

Project with rich metadata: certainty, intensity, projection magnitude.

fn dimensionality(&self) -> usize