Scikit-Learn · Class

SGDClassifier

SGDClassifier trains regularized linear classifiers with stochastic gradient descent. It supports configurable losses, sparse or dense floating-point features, and incremental learning through partial_fit.

How it works

SGDClassifier fits a regularized linear classification model by estimating each sample's loss gradient and updating parameters using a decreasing learning-rate schedule. The selected loss determines the model behavior; partial_fit supports incremental or out-of-core training, while suitable feature scaling can improve results with the default schedule.

Syntax

SGDClassifier(loss='hinge', *, penalty='l2', alpha=0.0001, l1_ratio=0.15, fit_intercept=True, max_iter=1000, tol=0.001, shuffle=True, verbose=0, epsilon=0.1, n_jobs=None, random_state=None, learning_rate='optimal', eta0=0.0, power_t=0.5, early_stopping=False, validation_fraction=0.1, n_iter_no_change=5, class_weight=None, warm_start=False, average=False)

Parameters

Name	Type	Description
`loss`	`str, default='hinge'`	The loss function to use; the default is `hinge`.
`penalty`	`str, default='l2'`	The penalty specification used for regularization.
`max_iter`	`int, default=1000`	The maximum number of passes over the training data for `fit`.
`learning_rate`	`str, default='optimal'`	The learning-rate schedule.
`alpha`	`float, default=0.0001`	The coefficient used to multiply the regularization term.
`average`	`bool or int, default=False`	Whether to use averaged SGD.
`fit_intercept`	`bool, default=True`	Whether to fit the intercept.
`random_state`	`int, RandomState instance or None, default=None`	The random seed used for shuffling and reproducibility.

Returns

An SGDClassifier estimator object.

Key points

Use it for linear classification tasks, including SVM-like and logistic-regression-like models.
Training estimates the loss gradient per sample rather than computing a full-batch gradient.
Parameters are updated progressively while the learning-rate strength decreases.
Use partial_fit to process data incrementally or in minibatches.
Feature matrices may be dense or sparse, provided feature values are floating point.
Choose the classification model behavior through the loss setting.
Enable averaged stochastic gradient descent with average=True when appropriate.

Common mistakes

Standardize features when using the default learning-rate schedule to improve optimization.

Example: Train and predict with SGDClassifier

from sklearn.linear_model import SGDClassifier

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

clf = SGDClassifier(loss="hinge", penalty="l2", max_iter=1000, random_state=0)
clf.fit(X, y)
print(clf.predict([[0.8, 0.8]]))

Creates a regularized linear classifier, trains it on labeled samples, and predicts a class for a new sample.

When to use it

Minibatch or out-of-core classification

Large labeled datasets that can be processed in chunks.

Sparse linear classification

Sparse feature representations such as high-dimensional text features.

Loss-configurable linear models

Tasks needing different linear classification objectives.

Averaged SGD

Models where a lower-variance online estimate is useful.

Interview questions

QHow does SGDClassifier train a model?

It is a linear classifier trained with regularized stochastic gradient descent. The gradient is estimated one sample at a time, and the parameters are updated under a decreasing learning-rate schedule.

Follow-up: How would you train when the full dataset cannot fit in memory? Use partial_fit for minibatch or out-of-core updates.

QWhat does the loss parameter control?

The loss parameter selects the classification objective; with the default loss, the estimator fits a linear support vector machine. Other supported losses can produce different linear-model behavior.

Follow-up: What is one practical choice for probability-like outputs? Select a probabilistic loss such as log_loss, then use the estimator's probability methods when supported.

QWhat is the difference between fit and partial_fit?

fit trains for up to max_iter passes over the training data. partial_fit is intended for incremental updates and is not governed by max_iter.

Follow-up: What preprocessing is especially useful with the default schedule? Scale features to approximately zero mean and unit variance.

SGDRegressor
LogisticRegression
LinearSVC
sklearn.linear_model

Source: SGDClassifier — scikit-learn 1.9.1 documentation. Checked 2026-10-15.