vix.ing · top · new · best · stats · spec

Online Gradient Boosting

2015/06/16 by Alina Beygelzimer, Beygelzimer, Alina, Elad Hazan +5 · 4 citations
Computer Science · Decision Sciences · Engineering · #Advanced Bandit Algorithms Research #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning and Algorithms #Sparse and Compressive Sensing Techniques

paper · pdf · doi:10.48550/arxiv.1506.04820

openalex publication_date 2015/06/16 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

We extend the theory of boosting for regression problems to the online learning setting. Generalizing from the batch setting for boosting, the notion of a weak learning algorithm is modeled as an online learning algorithm with linear loss functions that competes with a base class of regression functions, while a strong learning algorithm is an online learning algorithm with convex loss functions that competes with a larger class of regression functions. Our main result is an online gradient boosting algorithm which converts a weak online learning algorithm into a strong one where the larger class of functions is the linear span of the base class. We also give a simpler boosting algorithm that converts a weak online learning algorithm into a strong one where the larger class of functions is the convex hull of the base class, and prove its optimality.

Citations

Cited by

Related