Title:Robust LogitBoost and Adaptive Base Class (ABC) LogitBoost

Abstract:Logitboost is an influential boosting algorithm for classification. In this paper, we develop robust logitboost to provide an explicit formulation of tree-split criterion for building weak learners (regression trees) for logitboost. This formulation leads to a numerically stable implementation of logitboost. We then propose abc-logitboost for multi-class classification, by combining robust logitboost with the prior work of abc-boost. Previously, abc-boost was implemented as abc-mart using the mart algorithm. Our extensive experiments on multi-class classification compare four algorithms: mart, abcmart, (robust) logitboost, and abc-logitboost, and demonstrate the superiority of abc-logitboost. Comparisons with other learning methods including SVM and deep learning are also available through prior publications.