Struct MultinomialFit 

pub struct MultinomialFit { /* private fields */ }

A fitted baseline-category (multinomial) logistic model for an unordered categorical response with K ≥ 2 classes.

Class 0 is the reference. For each non-reference class k = 1 … K−1 there is a coefficient vector βₖ and linear predictor ηᵢₖ = xᵢᵀβₖ (with ηᵢ₀ ≡ 0), giving the softmax probabilities

P(yᵢ = k) = exp(ηᵢₖ) / (1 + Σⱼ exp(ηᵢⱼ)).

Fit by Newton–Raphson on the full (K−1)·p parameter vector: at convergence the per-class score equations Xᵀ(yₖ − pₖ) = 0 hold, and the inverse of the block information matrix gives the coefficient covariance the Wald statistics use.

With K = 2 this reduces exactly to binary LogisticFit (the βₖ for the single non-reference class equal the logistic coefficients).

Implementations

impl MultinomialFit

pub fn new(x: Array2<f64>, y: Array1<f64>) -> Result<Self>

Fit multinomial logistic regression of class-labelled y on X (default: up to 100 Newton iterations, tolerance 1e-10 on the step).

The response must hold integer class labels 0 … K−1 with every class present; K is inferred as max(y) + 1. The caller owns the design matrix, intercept column included.

Errors

• RegressionError::EmptyInput / RegressionError::ShapeMismatch.
• RegressionError::InvalidResponse if labels are not consecutive integers from 0, some class is empty, or there are fewer than two classes.
• RegressionError::RankDeficient if the information matrix is singular.
• RegressionError::NotConverged if Newton’s method fails to converge.

pub fn with_options( x: Array2<f64>, y: Array1<f64>, max_iter: usize, tol: f64, ) -> Result<Self>

Like MultinomialFit::new with an explicit iteration cap and tolerance.

pub fn n_observations(&self) -> usize

Number of observations.

pub fn n_features(&self) -> usize

Number of design columns p (per class).

pub fn n_classes(&self) -> usize

Number of response classes K.

pub fn n_parameters(&self) -> usize

Total number of free coefficients, (K − 1)·p.

pub fn has_intercept(&self) -> bool

Whether a constant (intercept) column was detected.

pub fn iterations(&self) -> usize

Newton iterations taken to converge.

pub fn design_matrix(&self) -> ArrayView2<'_, f64>

The design matrix as fitted.

pub fn response(&self) -> ArrayView1<'_, f64>

The integer class labels.

pub fn coefficients(&self) -> ArrayView2<'_, f64>

Coefficients, shape (K−1) × p; row k−1 is βₖ for class k relative to the reference class 0.

pub fn fitted_probabilities(&self) -> ArrayView2<'_, f64>

Fitted class probabilities, shape n × K.

pub fn covariance(&self) -> ArrayView2<'_, f64>

Covariance of the stacked (K−1)·p coefficient vector (class-major).

pub fn log_likelihood(&self) -> f64

Maximized log-likelihood.

pub fn coefficient_standard_errors(&self) -> Array2<f64>

Coefficient standard errors, shape (K−1) × p, aligned with coefficients.

pub fn z_values(&self) -> Array2<f64>

Wald z-statistics βₖⱼ / seₖⱼ, shape (K−1) × p.

pub fn p_values(&self) -> Array2<f64>

Two-sided Wald p-values from the standard normal, shape (K−1) × p.

pub fn residual_deviance(&self) -> f64

Residual deviance −2ℓ.

pub fn null_deviance(&self) -> f64

Deviance of the intercept-only model (class marginals nₖ/n).

pub fn mcfadden_r2(&self) -> f64

McFadden’s pseudo-R², 1 − ℓ/ℓ₀.

pub fn aic(&self) -> f64

Akaike information criterion, −2ℓ + 2·(K−1)·p.

pub fn bic(&self) -> f64

Bayesian information criterion, −2ℓ + ln(n)·(K−1)·p.

pub fn predict_proba(&self, x: ArrayView2<'_, f64>) -> Array2<f64>

Predicted class probabilities for a new design matrix x (same column layout as training), shape rows × K.

Trait Implementations

impl Clone for MultinomialFit

fn clone(&self) -> MultinomialFit

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MultinomialFit

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

Formats the value using the given formatter.