Export Reviews, Discussions, Author Feedback and Meta-Reviews

The authors study the online setting for boosting in the context of regression problems. Specifically, they describe and analyze two algorithms for online boosting for regression: (1) a boosting algorithm that uses a linear span of the base learning functions as the prediction function (i.e., the standard case) and (2) a boosting algorithm that uses a convex hull (CH) of the base functions as the prediction function. Algorithm (1) more closely aligns with existing gradient boosting approaches and provides the most practical insight. Algorithm (2) has some nice theoretical properties with respect to being optimal for the specified setting (and may give some insight to optimality in the span(F) case of algorithm (1). Experiments are also performed on 14 standard datasets and show that the proposed approaches outperform the base learners on average (and nearly universally when looking at the supplementary material).