Some classifier families are intrinsically multiclass while others — notably margin-based and discriminant-function methods — are formulated for two classes only, so multiclass problems must be reduced to collections of binary problems. The two canonical reductions are one-versus-all, which trains one classifier per class against the union of the rest, and one-versus-one, which trains a classifier for every unordered pair of classes; they differ in the number of models required, in the class balance of each induced subproblem, and in how their outputs are combined. The topic belongs to supervised classification in machine learning and connects to class imbalance, decision aggregation, and hierarchical decomposition of the label space.
Some classifier families are intrinsically multiclass while others — notably margin-based and discriminant-function methods — are formulated for two classes only, so multiclass problems must be reduc…