Struct NearestNeighborMetaModel 

pub struct NearestNeighborMetaModel { /* private fields */ }

The simplest useful meta-model: given a new corpus, return the best_config of the training record whose corpus-feature vector is closest in z-score-normalized Euclidean distance (scale features log-compressed first — see [to_model_space]).

• Works with N ≥ 1 records.
• Deterministic; no hyperparameters.
• Degenerate features (zero variance across training records) are dropped from the distance computation at fit time so they don’t divide by zero or dominate via raw-scale inflation.
• Training sets mixing multiple metric_names are stratified to the dominant metric at fit time — see Self::records for what was actually retained.

Implementations

impl NearestNeighborMetaModel

pub fn new() -> Self

pub fn records(&self) -> &[MetaTrainingRecord]

Borrow the training records — useful for introspecting what the model considers the nearest-neighbor candidate pool (after the dominant-metric stratification applied by fit).

pub fn rank_candidates(&self, features: &CorpusFeatures) -> Vec<(usize, f64)>

Distance from a given feature vector to every stored record, sorted ascending. Returned as (record_index, distance) pairs.

pub fn predict_blended( &self, features: &CorpusFeatures, k: usize, ) -> PipelineConfig

k-NN prediction with per-knob aggregation, instead of copying the single nearest record’s config wholesale.

Takes the k nearest records (clamped to the training-set size), picks the majority projection_kind (ties break toward the nearest record), and sets each tuned knob to the median of the top-k values — kind-specific knobs aggregate over kind-matching neighbors only.

Blended knobs: routing.num_domain_groups, routing.low_evr_threshold, bridges.threshold_base, bridges.threshold_evr_penalty, bridges.overlap_artifact_territorial, inner_sphere.min_evr_improvement, plus the kind-specific laplacian.* / umap.* knobs. Everything else (bridges.balanced_affinity_quantile, bridges.min_evr_for_classification, routing.group_routing_alpha, the remaining inner_sphere.* knobs, spatial.*, min_category_size) is inherited nearest-neighbor from the closest record.

With k = 1 this is exactly MetaModel::predict. Larger k trades sharpness for robustness against a single poorly-tuned or mislabeled record. Panics like predict if fit has not been called with at least one record.

Trait Implementations

impl Clone for NearestNeighborMetaModel

fn clone(&self) -> NearestNeighborMetaModel

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for NearestNeighborMetaModel

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for NearestNeighborMetaModel

fn default() -> Self

Returns the “default value” for a type.

impl MetaModel for NearestNeighborMetaModel

fn fit(&mut self, records: &[MetaTrainingRecord])

Fit on a training set. Replacing any prior state.

fn is_fitted(&self) -> bool

True once fit has been called with at least one usable record. Self::predict panics when this is false, so Result-returning boundaries (e.g. SphereQLPipeline::new_from_metamodel) check this first. No default impl on purpose: every model must answer for its own notion of “fitted” rather than inherit a guess.

fn predict(&self, features: &CorpusFeatures) -> PipelineConfig

Predict the config that should work best on a corpus with the given profile. Panics if fit has not been called with at least one record — callers should treat MetaModel as a trained object and front-load fit, or check Self::is_fitted when the training state isn’t statically known.

fn name(&self) -> &str

Short name for logs and model comparison.