Extend a binary classifier to multiclass data
A multiclass dataset where the chosen base estimator is binary-classification oriented.
Scikit-Learn · Class
OneVsRestClassifier
OneVsRestClassifier wraps a binary classifier to solve multiclass problems by fitting one model per class against all other classes. It also supports multilabel classification through independent binary models for each label.
How it works
OneVsRestClassifier wraps a binary estimator to solve multiclass problems by training a separate model for each class against all remaining classes. It can also handle multilabel targets when fit receives a two-dimensional binary indicator matrix, training one independent binary model per label.
Syntax
class sklearn.multiclass.OneVsRestClassifier(estimator, *, n_jobs=None, verbose=0)Parameters
| Name | Type | Description |
|---|---|---|
estimator | estimator object | The binary classifier to clone and fit for each class or label. |
n_jobs | int or None, default=None | Number of jobs to run in parallel; None means one job unless a backend context changes it. |
verbose | int, default=0 | Verbosity level for fitting and prediction operations. |
Returns
A fitted OneVsRestClassifier instance when fit is called; its prediction methods produce class or multilabel outputs.
Key points
- This is the scikit-learn meta-estimator for the one-versus-rest, also called one-versus-all, strategy.
- For a multiclass target, it creates one binary classifier associated with each class.
- The strategy uses one model per class, so its classifier count grows linearly with the number of classes.
- Each class has a corresponding model that can be inspected separately, which makes class-specific behavior easier to analyze.
- The class supports both ordinary multiclass classification and multilabel prediction.
- Multilabel targets must use a two-dimensional matrix whose entries indicate whether each sample has each label.
- Its multilabel implementation is binary relevance: one independent binary learning problem is created for every label.
Common mistakes
- For multilabel classification, pass a two-dimensional binary indicator matrix as the target to fit rather than a one-dimensional class-label array.
Example: Multiclass one-versus-rest classification
from sklearn.datasets import load_iris
from sklearn.multiclass import OneVsRestClassifier
from sklearn.svm import LinearSVC
X, y = load_iris(return_X_y=True)
model = OneVsRestClassifier(LinearSVC(random_state=0))
model.fit(X, y)
predictions = model.predict(X[:3])
print(predictions)This trains one linear binary classifier per iris class and predicts the classes of three samples.
When to use it
Train independent multilabel predictors
A target represented by a sample-by-label 0/1 matrix.
Analyze one model per class
A need to inspect class-specific decision boundaries or coefficients.
Interview questions
QHow does OneVsRestClassifier implement multiclass classification?
It trains one binary classifier for every class. Each classifier separates its own class from all other classes, producing an OvR multiclass model.
Follow-up: How does this extend to multilabel classification? Pass a binary indicator matrix; the wrapper independently trains one binary classifier for each label.
QWhat determines whether probability predictions are available?
The wrapped estimator must be a classifier compatible with the estimator interface. The wrapper delegates fitting and prediction to one cloned binary estimator per class or label.
Follow-up: What must the base estimator provide for probability output? It must implement predict_proba; otherwise the wrapper cannot expose probability predictions.
QHow must multilabel targets be represented?
For multilabel data, each row of y represents a sample and each column represents a label. A 1 in position [i, j] means label j is present in sample i.
Follow-up: Why can OvR be useful for multilabel tasks? Each label is modeled independently, so the approach directly implements binary relevance.
Related
OneVsOneClassifierOutputCodeClassifierClassifierChainsklearn.multiclass