Week 7 lecture notes
Lecture Notes: Ensemble Methods
1. Motivation for Ensembles
Goal: Improve predictive performance by combining multiple models rather than
relying on a single classifier/regressor.
Statistical benefit: Primarily reduces variance, leading to more stable estimates and
often better empirical accuracy.
Typical effect: Unstable (high-variance) base learners become robust when
ensembled.
Implementation notes:
Embarrassingly parallel: Each can be trained independently, ideal for distributed
systems.
Flexibility: Can introduce extra randomness—e.g., different random initializations,
feature subsets per bag—to increase diversity.
Overfitting immunity: Since each base model is weak, adding more bags rarely leads
to overfitting.