方法对比

	梯度提升（Gradient Boosting）	在线梯度提升	在线学习
领域	机器学习	机器学习	机器学习
方法族	Machine learning	Machine learning	Machine learning
起源年份	2001	2011–2015	1958–2000s
提出者	Friedman, J. H.	Grubb, A. & Bagnell, J. A.; Beygelzimer, A. et al.	Rosenblatt, F.; Littlestone, N.; Shalev-Shwartz, S. (key contributors)
类型	Ensemble (sequential boosting of decision trees)	Online ensemble (sequential boosting on streaming data)	Learning paradigm (sequential model update)
开创性文献	Friedman, J. H. (2001). Greedy Function Approximation: A Gradient Boosting Machine. Annals of Statistics, 29(5), 1189–1232.	Grubb, A. & Bagnell, J. A. (2011). Generalized Boosting Algorithms for Convex Optimization. Proceedings of the 28th International Conference on Machine Learning (ICML 2011), 1209–1216.	Shalev-Shwartz, S. (2011). Online Learning and Online Convex Optimization. Foundations and Trends in Machine Learning, 4(2), 107–194.
别名	Gradient Boosting (GBM), GBM, gradient boosted trees, gradient boosting machine	OGB, streaming gradient boosting, incremental gradient boosting, online boosting with gradient descent	incremental learning, sequential learning, streaming learning, online machine learning
相关	5	6	6
摘要	Gradient Boosting is an ensemble learning method, formalised by Jerome H. Friedman in 2001, that combines a sequence of weak learners — typically shallow decision trees — so that each new tree is fitted to minimise the residual errors of the trees before it. It is the core algorithm behind popular implementations such as XGBoost, LightGBM and CatBoost.	Online Gradient Boosting adapts the gradient boosting framework for streaming settings where data arrives one sample at a time rather than as a fixed batch. At each step the model computes a pseudo-residual for the incoming observation and updates a weak learner in place, growing an additive ensemble without storing or revisiting past data. This makes it suitable for real-time prediction and large-scale streaming pipelines where retraining from scratch is infeasible.	Online learning is a machine learning paradigm in which a model is updated incrementally as each new data point arrives, rather than being trained once on a fixed dataset. It is essential when data streams continuously, storage is limited, or the underlying distribution shifts over time. Theoretical performance is measured by cumulative regret relative to the best fixed predictor in hindsight.