k-Nearest Neighbors

The only major supervised family that never fits parameters. k-NN keeps the training set and defers every decision to prediction time, which makes it the honest baseline you run before anything else, and a fast lesson in why feature scaling, distance metrics, and the curse of dimensionality matter.

Learning Objectives

• →Compute pairwise Euclidean distances with broadcasting instead of loops
• →Classify by majority vote over the k nearest neighbours with a deterministic tie-break
• →Explain why unscaled features silently destroy a distance-based model
• →Read the bias-variance tradeoff off a sweep over k

Practice