Struct TreeConfig 

pub struct TreeConfig {

    pub max_depth: usize,
    pub n_bins: usize,
    pub lambda: f64,
    pub gamma: f64,
    pub grace_period: usize,
    pub delta: f64,
    pub feature_subsample_rate: f64,
    pub seed: u64,
    pub leaf_decay_alpha: Option<f64>,
    pub split_reeval_interval: Option<usize>,
    pub feature_types: Option<Vec<FeatureType>>,
    pub gradient_clip_sigma: Option<f64>,
    pub monotone_constraints: Option<Vec<i8>>,
    pub leaf_model_type: LeafModelType,
}

Configuration for a single streaming decision tree.

These parameters control tree growth, regularization, and the statistical confidence required before committing a split decision.

Defaults

Parameter	Default
max_depth	6
n_bins	64
lambda	1.0
gamma	0.0
grace_period	200
delta	1e-7
feature_subsample_rate	1.0

Fields

max_depth: usize

Maximum tree depth. Deeper trees capture more interactions but risk overfitting on streaming data. Default: 6.

n_bins: usize

Number of histogram bins per feature. More bins give finer split resolution at the cost of memory and slower convergence. Default: 64.

lambda: f64

L2 regularization parameter (lambda). Penalizes large leaf weights, helping prevent overfitting. Appears in the denominator of the leaf weight formula: w = -G / (H + lambda). Default: 1.0.

gamma: f64

Minimum split gain threshold (gamma). A candidate split must achieve gain > gamma to be accepted. Higher values produce more conservative trees. Default: 0.0.

grace_period: usize

Minimum number of samples a leaf must accumulate before evaluating potential splits. Also controls when bin edges are computed from observed feature values. Default: 200.

delta: f64

Hoeffding bound confidence parameter (delta). Smaller values require more statistical evidence before committing a split, producing more conservative but more reliable trees. The bound guarantees that the chosen split is within epsilon of optimal with probability 1-delta. Default: 1e-7.

feature_subsample_rate: f64

Fraction of features to consider at each split evaluation. 1.0 means all features are evaluated; smaller values introduce randomness (similar to random forest feature bagging). Default: 1.0.

seed: u64

Random seed for feature subsampling. Default: 42.

Set by the ensemble orchestrator to ensure deterministic, diverse behavior across trees (typically config.seed ^ step_index).

leaf_decay_alpha: Option<f64>

Per-sample decay factor for leaf statistics.

Computed from leaf_half_life as exp(-ln(2) / half_life). When Some(alpha), leaf gradient/hessian sums and histogram bins are decayed by alpha before each new accumulation. None (default) means no decay.

split_reeval_interval: Option<usize>

Re-evaluation interval for max-depth leaves.

When Some(n), leaves at max depth will re-evaluate potential splits every n samples, allowing the tree to adapt its structure over time. None (default) disables re-evaluation.

feature_types: Option<Vec<FeatureType>>

Per-feature type declarations (continuous vs categorical).

When Some, categorical features use one-bin-per-category binning and Fisher optimal binary partitioning. None (default) treats all features as continuous.

gradient_clip_sigma: Option<f64>

Per-leaf gradient clipping threshold in standard deviations.

When Some(sigma), leaf-level EWMA gradient statistics are tracked and incoming gradients are clamped to mean ± sigma * std_dev. None (default) disables clipping.

monotone_constraints: Option<Vec<i8>>

Per-feature monotonic constraints: +1 = increasing, -1 = decreasing, 0 = free.

Candidate splits violating monotonicity are rejected. None (default) means no constraints.

leaf_model_type: LeafModelType

Leaf prediction model type.

Controls how each leaf computes its prediction:

• ClosedForm (default): constant leaf weight w = -G / (H + lambda).
• Linear: online ridge regression with AdaGrad optimization, learning a local w . x + b surface. Optional decay for concept drift. Recommended for low-depth trees (depth 2–4).
• MLP: single hidden layer neural network per leaf. Optional decay for concept drift.
• Adaptive: starts as closed-form, auto-promotes to a more complex model when the Hoeffding bound (using delta) confirms it is statistically superior. No arbitrary thresholds.

Implementations

impl TreeConfig

pub fn new() -> Self

Create a new TreeConfig with default parameters.

Equivalent to TreeConfig::default(), but provided as a named constructor for clarity in builder chains.

pub fn max_depth(self, max_depth: usize) -> Self

Set the maximum tree depth.

pub fn n_bins(self, n_bins: usize) -> Self

Set the number of histogram bins per feature.

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

Set the L2 regularization parameter (lambda).

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

Set the minimum split gain threshold (gamma).

pub fn grace_period(self, grace_period: usize) -> Self

Set the grace period (minimum samples before evaluating splits).

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

Set the Hoeffding bound confidence parameter (delta).

pub fn feature_subsample_rate(self, rate: f64) -> Self

Set the feature subsample rate.

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

Set the random seed for feature subsampling.

pub fn leaf_decay_alpha(self, alpha: f64) -> Self

Set the per-sample decay factor for leaf statistics.

pub fn leaf_decay_alpha_opt(self, alpha: Option<f64>) -> Self

Optionally set the per-sample decay factor for leaf statistics.

pub fn split_reeval_interval(self, interval: usize) -> Self

Set the re-evaluation interval for max-depth leaves.

pub fn split_reeval_interval_opt(self, interval: Option<usize>) -> Self

Optionally set the re-evaluation interval for max-depth leaves.

pub fn feature_types(self, types: Vec<FeatureType>) -> Self

Set the per-feature type declarations.

pub fn feature_types_opt(self, types: Option<Vec<FeatureType>>) -> Self

Optionally set the per-feature type declarations.

pub fn gradient_clip_sigma(self, sigma: f64) -> Self

Set the gradient clipping threshold in standard deviations.

pub fn gradient_clip_sigma_opt(self, sigma: Option<f64>) -> Self

Optionally set the gradient clipping threshold.

pub fn monotone_constraints(self, constraints: Vec<i8>) -> Self

Set per-feature monotonic constraints.

pub fn monotone_constraints_opt(self, constraints: Option<Vec<i8>>) -> Self

Optionally set per-feature monotonic constraints.

pub fn leaf_model_type(self, lmt: LeafModelType) -> Self

Set the leaf prediction model type.

LeafModelType::Linear is recommended for low-depth configurations (depth 2–4) where per-leaf linear models significantly reduce approximation error compared to constant leaves.

LeafModelType::Adaptive automatically selects between closed-form and a trainable model per leaf, using the Hoeffding bound for promotion.

Trait Implementations

impl Clone for TreeConfig

fn clone(&self) -> TreeConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TreeConfig

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

Formats the value using the given formatter.

impl Default for TreeConfig

fn default() -> Self

Returns the “default value” for a type.