Scikit-Learn · Class

LinearSVC

LinearSVC is a linear support-vector classifier implemented with liblinear. It supports dense and sparse inputs, flexible penalties and losses, and one-vs-the-rest multiclass classification.

How it works

LinearSVC learns a linear decision boundary for classification using liblinear. It supports dense or sparse data and handles multiple classes with one-vs-the-rest training.

Syntax

class sklearn.svm.LinearSVC(*, penalty='l2', loss='squared_hinge', dual='auto', tol=0.0001, C=1.0, multi_class='ovr', fit_intercept=True, intercept_scaling=1, class_weight=None, verbose=0, random_state=None, max_iter=1000)

Parameters

Name	Type	Description
`penalty`	`str, default='l2'`	Selects the penalty type: 'l1' or 'l2'.
`loss`	`str, default='squared_hinge'`	Chooses 'hinge' or 'squared_hinge' loss.
`dual`	`bool or 'auto', default='auto'`	Whether to solve the dual optimization problem; 'auto' chooses automatically.
`tol`	`float, default=1e-4`	Tolerance for stopping criterion.
`C`	`float, default=1.0`	Regularization parameter; smaller values mean stronger regularization.
`multi_class`	`str, default='ovr'`	Multiclass strategy: 'ovr' or 'crammer_singer'.
`fit_intercept`	`bool, default=True`	Whether to fit the intercept.
`intercept_scaling`	`float, default=1`	Scaling applied to the synthetic intercept feature.
`class_weight`	`dict, 'balanced', or None, default=None`	Class weights, or 'balanced' to derive them from class frequencies.
`verbose`	`int, default=0`	Verbosity level.
`random_state`	`int, RandomState instance, or None, default=None`	Seed controlling solver randomness.
`max_iter`	`int, default=1000`	Maximum number of solver iterations.

Returns

A fitted LinearSVC estimator with learned coefficients and intercept; prediction methods return class labels or decision scores.

Key points

It provides a linear SVM alternative to SVC(kernel='linear') while using the liblinear implementation.
Training accepts both dense feature matrices and sparse representations.
For multiple classes, it fits one binary model per class against all remaining classes.
L1 regularization can produce sparse feature-weight vectors, which may help with feature selection.
With dual set to auto, the solver selects the primal or dual formulation based on the problem dimensions and related settings.
C is the main regularization trade-off parameter and defaults to 1.0.

Common mistakes

Not every penalty and loss combination is valid; in particular, L1 regularization cannot be paired with hinge loss.

Example: Fit and predict with LinearSVC

from sklearn.svm import LinearSVC

X = [[0, 0], [1, 1], [1, 0], [0, 1]]
y = [0, 0, 1, 1]

model = LinearSVC(random_state=0, max_iter=10_000)
model.fit(X, y)
print(model.predict([[0.9, 0.1]]))

Creates a linear classifier, fits it to labeled samples, and predicts a class for a new sample.

When to use it

Large sparse classification

Classify text or other high-dimensional sparse feature vectors.

Sparse feature selection

Obtain a compact model with fewer active feature weights.

Multiclass prediction

Train a linear classifier across several target categories.

Interview questions

QWhat is LinearSVC and when would you choose it over SVC with a linear kernel?

LinearSVC is a linear support-vector classifier implemented with liblinear. Compared with SVC using a linear kernel, it offers different penalty and loss choices and is intended to scale better to many samples.

Follow-up: How are multiclass problems handled? The estimator trains one classifier per class in a one-vs-the-rest scheme.

QHow does C affect LinearSVC?

C controls the regularization strength inversely: smaller values apply stronger regularization, while larger values allow the model to fit the training data more closely. The default is 1.0.

Follow-up: What does class_weight do? It changes the weight of classes in the objective; balanced weighting derives weights from class frequencies.

QWhich learned attributes and prediction methods are useful after fitting?

The estimator exposes coef_ for the learned feature weights and intercept_ for the intercept. decision_function returns confidence scores, while predict maps samples to class labels.

Follow-up: What is a useful sparse-model configuration? Select penalty='l1' with a compatible loss to obtain sparse coefficient vectors.

sklearn.svm.SVC
sklearn.linear_model.LogisticRegression
sklearn.svm.NuSVC

Source: LinearSVC — scikit-learn 1.8.dev0 documentation. Checked 2026-10-15.