Gradient Boosting Machines¶
Authors: Jerome H. Friedman Year: 2001 Citations: 20,000+
Summary¶
Weak learners sequentially trained to predict residuals via gradient descent on function space. Powerful ensemble technique winning Kaggle competitions.
Key Concepts¶
- Residual Learning: Each learner fits previous errors
- Functional Gradients: Gradients on function space
- Shrinkage: Learning rate regularization
- Tree Boosting: Gradient boosting with decision trees
- Loss Functions: Works with differentiable objectives
Impact¶
- 20,000+ citations
- Foundation for XGBoost, LightGBM, CatBoost
- Dominant in ML competitions
- State-of-the-art on tabular data
- Better than Random Forests on many tasks
- Widely used in industry
Related Papers¶
- AdaBoost (Freund & Schapire, 1997)
- Bagging (Breiman, 1996)
- XGBoost (Chen & Guestrin, 2016)