Struct StackedEnsemble 

pub struct StackedEnsemble { /* private fields */ }

Available on crate feature alloc only.

Polymorphic model stacking meta-learner using Box<dyn StreamingLearner>.

Combines predictions from heterogeneous base learners through a trainable meta-learner. Uses temporal holdout to prevent information leakage: base predictions are collected before training the bases on each sample.

Note on Clone

StackedEnsemble cannot implement Clone because Box<dyn StreamingLearner> is not Clone. If you need to snapshot the ensemble, serialize it instead.

Implementations

impl StackedEnsemble

pub fn new( base_learners: Vec<Box<dyn StreamingLearner>>, meta_learner: Box<dyn StreamingLearner>, ) -> Self

Create a new stacked ensemble with passthrough disabled.

The meta-learner receives only base learner predictions as features.

Arguments

• base_learners – heterogeneous base models (at least one recommended)
• meta_learner – combiner model trained on base predictions

pub fn with_passthrough( base_learners: Vec<Box<dyn StreamingLearner>>, meta_learner: Box<dyn StreamingLearner>, passthrough: bool, ) -> Self

Create a new stacked ensemble with configurable feature passthrough.

When passthrough is true, the meta-learner receives both base predictions and the original feature vector, enabling it to learn corrections that depend on raw inputs.

Arguments

• base_learners – heterogeneous base models
• meta_learner – combiner model
• passthrough – if true, original features are appended to meta-features

pub fn n_base_learners(&self) -> usize

Number of base learners in the ensemble.

pub fn passthrough(&self) -> bool

Whether original features are passed through to the meta-learner.

pub fn base_predictions(&self, features: &[f64]) -> Vec<f64>

Get predictions from each base learner for inspection.

Returns a vector with one prediction per base learner, in the same order they were provided at construction time.

Trait Implementations

impl Debug for StackedEnsemble

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

Formats the value using the given formatter.

impl StreamingLearner for StackedEnsemble

fn train_one(&mut self, features: &[f64], target: f64, weight: f64)

Train on a single weighted observation using temporal holdout.

1. Collect base predictions before training (temporal holdout).
2. Build meta-features and train the meta-learner on (meta_features, target, weight).
3. Train each base learner on (features, target, weight).

fn predict(&self, features: &[f64]) -> f64

Predict by collecting base predictions and passing them through the meta-learner.

fn n_samples_seen(&self) -> u64

Total number of samples trained on since creation or last reset.

fn reset(&mut self)

Reset all base learners, the meta-learner, and the sample counter.

fn train(&mut self, features: &[f64], target: f64)

Train on a single observation with unit weight.

fn predict_batch(&self, feature_matrix: &[&[f64]]) -> Vec<f64>

Predict for each row in a feature matrix.