Struct BoostingStep 

pub struct BoostingStep { /* private fields */ }

Available on crate feature alloc only.

A single step in the SGBT boosting sequence.

Owns a TreeSlot and applies variant-aware training repetition. The number of training iterations per sample is determined by the caller (the ensemble orchestrator computes train_count from the configured SGBT variant).

Prediction semantics

Both train_and_predict and predict return the active tree’s prediction before any training on the current sample. This ensures unbiased gradient computation in the boosting loop.

Implementations

impl BoostingStep

pub fn new(tree_config: TreeConfig, detector: Box<dyn DriftDetector>) -> Self

Create a new boosting step with a fresh tree and drift detector.

pub fn new_with_max_samples( tree_config: TreeConfig, detector: Box<dyn DriftDetector>, max_tree_samples: Option<u64>, ) -> Self

Create a new boosting step with optional time-based tree replacement.

pub fn new_with_graduated( tree_config: TreeConfig, detector: Box<dyn DriftDetector>, max_tree_samples: Option<u64>, shadow_warmup: usize, ) -> Self

Create a new boosting step with graduated tree handoff.

pub fn from_slot(slot: TreeSlot) -> Self

Reconstruct a boosting step from a pre-built tree slot.

Used during model deserialization.

pub fn train_and_predict( &mut self, features: &[f64], gradient: f64, hessian: f64, train_count: usize, ) -> f64

Train on a single sample with variant-aware repetition.

Arguments

• features - Input feature vector.
• gradient - Negative gradient of the loss at this sample.
• hessian - Second derivative (curvature) of the loss at this sample.
• train_count - Number of training iterations for this sample:
  • 0: skip training entirely (SK variant or stochastic skip).
  • 1: standard single-pass training.
  • >1: multiple iterations (MI variant).

Returns

The prediction from the active tree before training.

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

Predict without training.

Routes the feature vector through the active tree and returns the leaf value. Does not update any state.

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

Predict with variance for confidence estimation.

Returns (leaf_value, variance) where variance = 1 / (H_sum + lambda).

pub fn predict_smooth(&self, features: &[f64], bandwidth: f64) -> f64

Predict using sigmoid-blended soft routing for smooth interpolation.

See crate::tree::hoeffding::HoeffdingTree::predict_smooth for details.

pub fn predict_smooth_auto(&self, features: &[f64], bandwidths: &[f64]) -> f64

Predict using per-feature auto-calibrated bandwidths.

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

Predict with parent-leaf linear interpolation.

pub fn predict_sibling_interpolated( &self, features: &[f64], bandwidths: &[f64], ) -> f64

Predict with sibling-based interpolation for feature-continuous predictions.

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

Predict with graduated active-shadow blending.

pub fn predict_graduated_sibling_interpolated( &self, features: &[f64], bandwidths: &[f64], ) -> f64

Predict with graduated blending + sibling interpolation.

pub fn n_leaves(&self) -> usize

Number of leaves in the active tree.

pub fn n_samples_seen(&self) -> u64

Total samples the active tree has seen.

pub fn has_alternate(&self) -> bool

Whether the slot has an alternate tree being trained.

pub fn reset(&mut self)

Reset to a completely fresh state: new tree, no alternate, reset detector.

pub fn slot(&self) -> &TreeSlot

Immutable access to the underlying TreeSlot.

pub fn slot_mut(&mut self) -> &mut TreeSlot

Mutable access to the underlying TreeSlot.

Trait Implementations

impl Clone for BoostingStep

fn clone(&self) -> BoostingStep

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BoostingStep

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

Formats the value using the given formatter.