onevsrest

One-vs-rest (OvR), also known as one-vs-all (OvA), is a strategy for multiclass classification that reduces the problem to multiple binary classifications. For a problem with k classes, OvR trains k independent binary classifiers. Each classifier learns to distinguish one class from the rest. At prediction time, every classifier provides a score or probability for its class, and the final predicted class is the one with the highest score.

Most commonly, the base binary classifier is logistic regression or a linear support vector machine, but any binary classifier can be used. In practice, many machine learning libraries provide a wrapper that implements OvR by training one model per class, often with shared data and features.

Advantages of OvR include simplicity and scalability to large numbers of classes, as well as the ability to reuse well-understood binary learners and to calibrate probabilities per class. It can also be robust when some classes are easy to separate.

Limitations include the possibility that the resulting decision boundaries are not globally optimal because they ignore interactions among classes during training. Performance can suffer if classes are highly imbalanced or if probability estimates are poorly calibrated. Train time grows linearly with the number of classes, since k binary models are built. In contrast to one-vs-one, OvR typically requires fewer models but can require careful thresholding.

Related approaches include one-vs-one (OvO), which trains a binary classifier for every pair of classes, and error-correcting output codes (ECOC), which use a designed coding scheme to combine multiple binary classifiers. OvR is commonly used as a baseline in multiclass tasks such as text classification and image recognition, and is implemented in many libraries under the OneVsRestClassifier or OvR wrappers.