Struct MultiClassLogLoss 

pub struct MultiClassLogLoss {
    pub num_classes: usize,
    /* private fields */
}

Multi-class LogLoss (Softmax Cross-Entropy)

Used for multi-class classification where targets are class indices (0, 1, …, K-1).

Note

This loss function works differently from regression/binary losses:

• gradient() and hessian() compute for a single class given all raw predictions
• Training builds K trees per round (one per class)
• compute_gradient_for_class() is the primary method for multi-class training

Fields

num_classes: usize

Number of classes

Implementations

impl MultiClassLogLoss

pub fn new(num_classes: usize) -> Self

Create a new multi-class log loss

pub fn gradient_hessian_for_class( &self, target_class: usize, class_idx: usize, raw_predictions: &[f32], ) -> (f32, f32)

Compute gradient and hessian for a specific class

Arguments

• target_class - The true class label (0 to K-1)
• class_idx - The class we’re computing gradient for
• raw_predictions - Raw predictions for all K classes

Returns

(gradient, hessian) for the specified class

pub fn gradient_hessian_all_classes( &self, target_class: usize, raw_predictions: &[f32], ) -> (Vec<f32>, Vec<f32>)

Compute gradients and hessians for all classes at once

More efficient when you need all K gradients for a sample.

pub fn compute_gradients_batch( &self, class_idx: usize, targets: &[f32], predictions: &[f32], sample_indices: &[usize], gradients: &mut [f32], hessians: &mut [f32], )

Compute gradients and hessians for a batch of samples for a specific class

This is optimized for the training loop where we compute gradients for all training samples for one class at a time.

Arguments

• class_idx - The class we’re computing gradients for
• targets - Target class indices for all samples
• predictions - Flattened predictions matrix: predictions[sample * num_classes + class]
• sample_indices - Indices of samples to process
• gradients - Output buffer for gradients (indexed by sample)
• hessians - Output buffer for hessians (indexed by sample)

pub fn initial_predictions(&self, targets: &[f32]) -> Vec<f32>

Compute initial predictions for all classes (log of class priors)

Uses the class frequencies as initial probabilities. Sets f_k = log(p_k) so that softmax(f) = class frequencies.

Math

If f_k = log(p_k), then:

• exp(f_k) = p_k
• softmax_k = p_k / sum(p_j) = p_k / 1 = p_k

This ensures the model starts with predictions matching the class distribution.

pub fn num_classes(&self) -> usize

Get number of classes

Trait Implementations

impl Clone for MultiClassLogLoss

fn clone(&self) -> MultiClassLogLoss

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MultiClassLogLoss

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

Formats the value using the given formatter.