Multiclass learnability and the ERM principle

Multiclass learning is an area of growing practical relevance, for which the currently available theory is still far from providing satisfactory understanding. We study the learnability of multiclass prediction, and derive upper and lower bounds on the sample complexity of multiclass hypothesis classes in different learning models: batch/online, realizable/unrealizable, full information/bandit feedback. Our analysis reveals a surprising phenomenon: In the multiclass setting, in sharp contrast to binary classification, not all Empirical Risk Minimization (ERM) algorithms are equally successful. We show that there exist hypotheses classes for which some ERM learners have lower sample complexity than others. Furthermore, there are classes that are learnable by some ERM learners, while other ERM learner will fail to learn them. We propose a principle for designing good ERM learners, and use this principle to prove tight bounds on the sample complexity of learning symmetric multiclass hypothesis classes (that is, classes that are invariant under any permutation of label names). We demonstrate the relevance of the theory by analyzing the sample complexity of two widely used hypothesis classes: generalized linear multiclass models and reduction trees. We also obtain some practically relevant conclusions.