Struct LinearSVC 

#[non_exhaustive]
pub struct LinearSVC { /* private fields */ }

Linear Support Vector Classifier.

Uses the Pegasos SGD algorithm to minimize hinge loss with L2 regularisation. Binary problems use a single weight vector; multiclass problems use one-vs-rest (one weight vector per class, prediction = argmax of decision function scores).

Example

use scry_learn::dataset::Dataset;
use scry_learn::svm::LinearSVC;

let features = vec![
    vec![0.0, 0.0, 10.0, 10.0],
    vec![0.0, 0.0, 10.0, 10.0],
];
let target = vec![0.0, 0.0, 1.0, 1.0];
let data = Dataset::new(features, target, vec!["x".into(), "y".into()], "class");

let mut svc = LinearSVC::new().c(1.0).max_iter(500);
svc.fit(&data).unwrap();

let preds = svc.predict(&[vec![1.0, 1.0]]).unwrap();
assert_eq!(preds[0] as usize, 0);

Implementations

impl LinearSVC

pub fn new() -> Self

Create a new LinearSVC with default parameters.

Defaults: C = 1.0, max_iter = 1000, tol = 1e-4.

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

Set the regularisation parameter C.

Larger values penalise misclassification more (tighter margin).

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

Set the maximum number of SGD epochs.

pub fn tol(self, t: f64) -> Self

Set convergence tolerance on the max weight change per epoch.

pub fn class_weight(self, cw: ClassWeight) -> Self

Set class weighting strategy for imbalanced datasets.

pub fn probability(self, enable: bool) -> Self

Enable Platt scaling for probability estimates.

When true, predict_proba returns calibrated class probabilities after fitting.

pub fn fit(&mut self, data: &Dataset) -> Result<()>

Train the SVM on the given dataset.

Uses Pegasos-style SGD with one-vs-rest decomposition for multiclass problems (≥ 3 classes). Auto-dispatches to sparse kernels when the dataset uses sparse storage.

pub fn predict_sparse(&self, csr: &CsrMatrix) -> Result<Vec<f64>>

Predict class labels from sparse input (CSR format).

pub fn predict(&self, features: &[Vec<f64>]) -> Result<Vec<f64>>

Predict class labels for the given row-major feature matrix.

Returns the class whose OVR decision function is largest.

pub fn decision_function(&self, features: &[Vec<f64>]) -> Result<Vec<Vec<f64>>>

Compute the raw decision function score for each class.

Returns scores[sample][class] = w · x + b for each OVR sub-problem.

Example

use scry_learn::dataset::Dataset;
use scry_learn::svm::LinearSVC;

let features = vec![
    vec![0.0, 0.0, 10.0, 10.0],
    vec![0.0, 0.0, 10.0, 10.0],
];
let target = vec![0.0, 0.0, 1.0, 1.0];
let data = Dataset::new(features, target, vec!["x".into(), "y".into()], "class");

let mut svc = LinearSVC::new();
svc.fit(&data).unwrap();

let scores = svc.decision_function(&[vec![1.0, 1.0]]).unwrap();
assert_eq!(scores[0].len(), 2); // two classes

pub fn predict_proba(&self, features: &[Vec<f64>]) -> Result<Vec<Vec<f64>>>

Predict class probabilities using Platt scaling.

Requires .probability(true) to have been set before fitting. Returns probabilities[sample][class] normalised to sum to 1.

Trait Implementations

impl CalibrableClassifier for LinearSVC

fn fit(&mut self, data: &Dataset) -> Result<()>

Train on a dataset.

fn predict(&self, features: &[Vec<f64>]) -> Result<Vec<f64>>

Predict class labels.

fn predict_proba(&self, features: &[Vec<f64>]) -> Result<Vec<Vec<f64>>>

Predict class probabilities. Returns [n_samples][n_classes].

fn clone_box(&self) -> Box<dyn CalibrableClassifier>

Clone into a boxed trait object.

impl Clone for LinearSVC

fn clone(&self) -> LinearSVC

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Default for LinearSVC

fn default() -> Self

Returns the “default value” for a type.

impl PipelineModel for LinearSVC

fn fit(&mut self, data: &Dataset) -> Result<()>

Train the model on a dataset.

fn predict(&self, features: &[Vec<f64>]) -> Result<Vec<f64>>

Predict on row-major feature matrix.

impl Tunable for LinearSVC

fn set_param(&mut self, name: &str, _value: ParamValue) -> Result<()>

Apply a named hyperparameter.

fn clone_box(&self) -> Box<dyn Tunable>

Clone this model into a boxed trait object.

fn fit(&mut self, data: &Dataset) -> Result<()>

Train on a dataset.

fn predict(&self, features: &[Vec<f64>]) -> Result<Vec<f64>>

Predict on row-major features.