Class Binarization

Synonyms

Error-correcting output codes (ECOC); One- against-all training; One-against-one training; Pairwise classification

Abstract

Many learning algorithms are only designed to separate two classes from each other. For example, concept-learning algorithms assume positive examples and negative examples (counterexamples) for the concept to learn, and many statistical learning techniques, such as neural networks or support vector machines, can only find a single separating decision surface. One way to apply these algorithms to multi-class problem is to transform the original multi-class problem into multiple binary problems.

Methods

The best-known techniques are:

• One against all: one concept-learning problem is defined for each class, i.e., each class is in turn used as the positive class, and all other classes form the negative class.

• Pairwise (One against one): one concept is learned for each pair of classes (Fürnkranz 2002). This may be viewed as a special case of preference learning....