Online Agnostic Multiclass Boosting
This work addresses the need for boosting algorithms in online multiclass classification, particularly in agnostic settings where labels may be noisy, but it is incremental as it builds directly on prior binary classification work.
The authors tackled the problem of extending online agnostic boosting from binary to multiclass classification, resulting in the first boosting algorithm for online agnostic multiclass classification, which also enables algorithms for statistical agnostic, online realizable, and statistical realizable settings.
Boosting is a fundamental approach in machine learning that enjoys both strong theoretical and practical guarantees. At a high-level, boosting algorithms cleverly aggregate weak learners to generate predictions with arbitrarily high accuracy. In this way, boosting algorithms convert weak learners into strong ones. Recently, Brukhim et al. extended boosting to the online agnostic binary classification setting. A key ingredient in their approach is a clean and simple reduction to online convex optimization, one that efficiently converts an arbitrary online convex optimizer to an agnostic online booster. In this work, we extend this reduction to multiclass problems and give the first boosting algorithm for online agnostic mutliclass classification. Our reduction also enables the construction of algorithms for statistical agnostic, online realizable, and statistical realizable multiclass boosting.