Struct BayesSearchCV 

#[non_exhaustive]
pub struct BayesSearchCV { /* private fields */ }

Bayesian hyperparameter optimization with cross-validation.

Uses a Tree-structured Parzen Estimator (TPE) to model the objective function and focus evaluations on the most promising hyperparameter combinations.

Algorithm

1. Evaluate n_initial random samples to bootstrap the surrogate model.
2. For each remaining iteration, split observed results at the gamma quantile into “good” and “bad” groups.
3. Build factored 1D kernel density estimates for each group.
4. Draw 100 random candidates and pick the one maximizing l(x) / g(x).
5. Evaluate the chosen candidate and add it to the history.

Examples

use scry_learn::prelude::*;
use scry_learn::search::*;

let mut space = ParamSpace::new();
space.insert("max_depth".into(), ParamDistribution::IntUniform { low: 2, high: 10 });

let result = BayesSearchCV::new(DecisionTreeClassifier::new(), space)
    .n_iter(20)
    .cv(3)
    .fit(&data)
    .unwrap();

println!("Best score: {:.3}", result.best_score());

Implementations

impl BayesSearchCV

pub fn new(model: impl Tunable + 'static, param_space: ParamSpace) -> Self

Create a Bayesian search over the given model and parameter space.

Defaults: 30 iterations, 10 initial random samples, gamma 0.25, 5-fold CV, accuracy scorer, seed 42, non-stratified.

pub fn n_iter(self, n: usize) -> Self

Set the total number of iterations (default: 30).

pub fn n_initial(self, n: usize) -> Self

Set the number of initial random exploration samples (default: 10).

pub fn gamma(self, gamma: f64) -> Self

Set the quantile threshold for splitting good/bad observations (default: 0.25).

pub fn cv(self, k: usize) -> Self

Set the number of cross-validation folds (default: 5).

pub fn scoring(self, scorer: ScoringFn) -> Self

Set the scoring function (default: accuracy).

pub fn seed(self, seed: u64) -> Self

Set the random seed (default: 42).

pub fn stratified(self, stratified: bool) -> Self

Enable stratified k-fold CV (default: false).

When true, uses stratified_k_fold to preserve class proportions in each fold.

pub fn fit(self, data: &Dataset) -> Result<Self>

Run the Bayesian optimization search.

Returns self for chained accessor calls.

pub fn best_params(&self) -> &HashMap<String, ParamValue>

The best parameter combination found.

Panics

Panics if called before fit.

pub fn best_score(&self) -> f64

The best mean CV score achieved.

pub fn cv_results(&self) -> &[CvResult]

All evaluated combinations with their scores.

Trait Implementations

impl Debug for BayesSearchCV

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

Formats the value using the given formatter.