Multi-class Classification

Multi-classification, sometimes called multi-classification, refers to the classification of more than two categories in a classification task.

Existing multi-class classification techniques can be divided into (i) conversion to binary (ii) expansion from binary and (iii) hierarchical classification.

Common strategies

1) The one-vs.-all strategy requires a unique classifier to be built for each class, and all samples belonging to this class are positive examples, and all others are negative examples. This strategy requires the base classifier to generate a real-valued confidence for decision-making, rather than just a class label; the class label generated alone may lead to ambiguous classification, so that a sample will be predicted to belong to multiple classes.

2) In the one-vs-one (OvO) strategy, for a K-class multivariate problem, K (K − 1) / 2 binary classifiers are trained; each receives a pair of class examples from the initial training set and must learn to distinguish between the two classes. During prediction time, there is a vote: all K (K − 1) / 2 explainers are applied to an unknown example, and the class that gets the most "+1" predictions becomes the prediction of the combined classifier.