Module ensemble 

Available on crate feature alloc only.

SGBT ensemble orchestrator – the core boosting loop.

Implements Streaming Gradient Boosted Trees (Gunasekara et al., 2024): a sequence of boosting steps, each owning a streaming tree and drift detector, with automatic tree replacement when concept drift is detected.

Algorithm

For each incoming sample (x, y):

1. Compute the current ensemble prediction: F(x) = base + lr * Σ tree_s(x)
2. For each boosting step s = 1..N:
   • Compute gradient g = loss.gradient(y, current_pred)
   • Compute hessian h = loss.hessian(y, current_pred)
   • Feed (x, g, h) to tree s (which internally uses weighted squared loss)
   • Update current_pred += lr * tree_s.predict(x)
3. The ensemble adapts incrementally, with each tree targeting the residual of all preceding trees.

Re-exports

pub use core::SGBT;

Modules

adaptive

Adaptive learning rate wrapper for SGBT ensembles.

adaptive_forest

Adaptive Random Forest (ARF) for streaming classification.

bagged

Bagged SGBT ensemble using Oza online bagging (Poisson weighting).

config

SGBT configuration with builder pattern and full validation.

core

SGBT core: struct definition, Clone/Debug, and constructors.

diagnostics

Diagnostics for SGBT ensembles.

distributional

Distributional SGBT – outputs Gaussian N(μ, σ²) instead of a point estimate.

lr_schedule

Learning rate scheduling for streaming gradient boosted trees.

moe

Streaming Mixture of Experts over SGBT ensembles.

moe_distributional

Streaming Mixture of Experts over Distributional SGBT ensembles with shadow expert competition.

multi_target

Multi-target regression via parallel SGBT ensembles.

multiclass

Multi-class classification via one-vs-rest SGBT committees.

parallelparallel

Parallel SGBT training with delayed gradient updates.

quantile_regressor

Non-crossing multi-quantile regression via parallel SGBT ensembles.

replacement

TreeSlot: warning/danger/swap lifecycle for tree replacement.

stacked

Polymorphic model stacking meta-learner for streaming ensembles.

step

Single boosting step: owns one tree + drift detector + optional alternate.

variants

SGBT computational variants (Gunasekara et al., 2024).

Type Aliases

DynSGBT

Type alias for an SGBT model using dynamic (boxed) loss dispatch.