Introducing Gradient Boosting Machines | Generated by AI

Home 2025.11

Overview

The 2001 paper “Greedy Function Approximation: A Gradient Boosting Machine” by Jerome H. Friedman introduces Gradient Boosting Machines (GBMs), a powerful ensemble learning method for supervised tasks like regression and classification. It frames boosting as a form of functional gradient descent, where simple “weak” learners (often decision trees) are sequentially added to an additive model to minimize a specified loss function. This approach generalizes earlier boosting algorithms (e.g., AdaBoost) and emphasizes greedy optimization in function space, leading to highly accurate, robust, and interpretable models.

Abstract (Paraphrased)

GBMs build flexible predictive models by combining weak learners in a sequential, additive fashion to approximate the minimizer of a differentiable loss function. Using regression trees as base learners yields competitive, robust procedures for regression and classification. The method outperforms alternatives like multivariate adaptive regression splines (MARS) in empirical tests, with low error rates across diverse datasets.

Key Methods

The core idea is to iteratively fit new learners to the negative gradient (pseudo-residuals) of the loss with respect to the current model’s predictions, mimicking gradient descent in function space.

Supported losses include:

Variants like LogitBoost adapt this for specific losses (e.g., binomial deviance).

Contributions

Main Findings

This paper laid the foundation for modern implementations like XGBoost and LightGBM.

References


Back

x-ai/grok-4-fast

Donate