Struct LinearSVC 

pub struct LinearSVC<F> {
    pub c: F,
    pub max_iter: usize,
    pub tol: F,
    pub loss: LinearSVCLoss,
    pub penalty: LinearSVCPenalty,
    pub dual: DualMode,
    pub fit_intercept: bool,
    pub intercept_scaling: F,
    pub class_weight: ClassWeight<F>,
    pub multi_class: MultiClass,
}

Linear Support Vector Classifier (liblinear dual CD).

Solves the L2-regularized hinge or squared-hinge objective 0.5*||w||^2 + C * sum_i L(y_i, w.x_i) via liblinear’s dual coordinate descent. Supports binary and multiclass (one-vs-rest) classification. Mirrors sklearn.svm.LinearSVC.

Type Parameters

• F: The floating-point type (f32 or f64).

Fields

c: F

Inverse regularization strength. Larger values allow more misclassification. Must be strictly positive.

max_iter: usize

Maximum number of dual coordinate descent iterations.

tol: F

Convergence tolerance on the projected-gradient span.

loss: LinearSVCLoss

Loss function to use.

penalty: LinearSVCPenalty

Regularizer norm (l1 / l2). Default l2. l1 is only supported with squared_hinge + dual=false (liblinear solver type 5, _get_liblinear_solver_type, _base.py:1014).

dual: DualMode

Dual / primal selector. Default Auto, resolved via _validate_dual_parameter (_classes.py:13-29). Observably immaterial for penalty=l2 (R-DEV-7 dual-invariance), load-bearing for the unsupported-combination rejects and the l1 primal solver.

fit_intercept: bool

Whether to fit an intercept. When true, the design matrix is augmented with a synthetic constant column equal to intercept_scaling; the augmented weight is penalized like any feature (liblinear convention).

intercept_scaling: F

Value of the synthetic intercept feature when fit_intercept is true. Must be strictly positive. intercept_ = intercept_scaling * w_last.

class_weight: ClassWeight<F>

Per-class scaling of C. Default ClassWeight::None (all classes weighted 1.0). The effective penalty for class i is class_weight[i]·C (compute_class_weight, _base.py:1179; weighted_C[i] = C·class_weight[i], linear.cpp:2496-2507).

multi_class: MultiClass

Multiclass strategy. Default MultiClass::Ovr (one-vs-rest). MultiClass::CrammerSinger runs the joint Crammer-Singer solver, ignoring penalty/loss/dual (_base.py:1017, _classes.py:239).

Implementations

impl<F: Float> LinearSVC<F>

pub fn new() -> Self

Create a new LinearSVC with scikit-learn’s default settings.

Defaults (matching sklearn.svm.LinearSVC, _classes.py): C = 1.0, max_iter = 1000, tol = 1e-4, loss = SquaredHinge, penalty = L2, dual = Auto, fit_intercept = true, intercept_scaling = 1.0, class_weight = None, multi_class = Ovr.

pub fn with_c(self, c: F) -> Self

Set the regularization parameter C.

pub fn with_max_iter(self, max_iter: usize) -> Self

Set the maximum number of iterations.

pub fn with_tol(self, tol: F) -> Self

Set the convergence tolerance.

pub fn with_loss(self, loss: LinearSVCLoss) -> Self

Set the loss function.

pub fn with_penalty(self, penalty: LinearSVCPenalty) -> Self

Set the penalty (regularizer) norm (sklearn penalty). l1 requires squared_hinge + dual=false (_base.py:1014).

pub fn with_dual(self, dual: DualMode) -> Self

Set the dual / primal selector (sklearn dual). Auto (default) resolves via _validate_dual_parameter (_classes.py:13-29).

pub fn with_fit_intercept(self, fit_intercept: bool) -> Self

Set whether to fit an intercept (sklearn fit_intercept).

pub fn with_intercept_scaling(self, intercept_scaling: F) -> Self

Set the intercept scaling (sklearn intercept_scaling). Must be strictly positive when fit_intercept is true.

pub fn with_class_weight(self, class_weight: ClassWeight<F>) -> Self

Set the per-class C scaling (sklearn class_weight, _classes.py:118-124). ClassWeight::None (default) leaves every class at 1.0; ClassWeight::Balanced uses n_samples / (n_classes · count_c); ClassWeight::Explicit takes a (label, weight) map (unlisted classes default to 1.0).

pub fn with_multi_class(self, multi_class: MultiClass) -> Self

Set the multiclass strategy (sklearn multi_class, _classes.py:239). MultiClass::Ovr (default) trains one-vs-rest binary classifiers; MultiClass::CrammerSinger runs the joint Crammer-Singer solver (ignoring penalty/loss/dual, _base.py:1017).

Trait Implementations

impl<F: Clone> Clone for LinearSVC<F>

fn clone(&self) -> LinearSVC<F>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<F: Debug> Debug for LinearSVC<F>

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

Formats the value using the given formatter.

impl<F: Float> Default for LinearSVC<F>

fn default() -> Self

Returns the “default value” for a type.

impl<F: Float + Send + Sync + ScalarOperand + 'static> Fit<ArrayBase<OwnedRepr<F>, Dim<[usize; 2]>>, ArrayBase<OwnedRepr<usize>, Dim<[usize; 1]>>> for LinearSVC<F>

fn fit( &self, x: &Array2<F>, y: &Array1<usize>, ) -> Result<FittedLinearSVC<F>, FerroError>

Fit the linear SVC model using liblinear’s dual coordinate descent.

Minimizes 0.5*||w||^2 + C * sum_i L(y_i, w.x_i) (no 1/n). When fit_intercept is set, the design matrix is augmented with a constant column equal to intercept_scaling, the augmented weight is penalized like any feature, and intercept_ = intercept_scaling * w_last (_base.py:1188-1198,:1240-1245).

Errors

• FerroError::ShapeMismatch — sample count mismatch.
• FerroError::InvalidParameter — C not positive, or intercept_scaling not positive when fitting an intercept.
• FerroError::InsufficientSamples — fewer than 2 distinct classes.

type Fitted = FittedLinearSVC<F>

The fitted model type returned by fit.

type Error = FerroError

The error type returned by fit.