Polynomial Features

Also known as · interaction terms · polynomial regression

Squared, higher-order and cross-product (interaction) columns added to the inputs so a linear model can fit curves and joint effects. The model is still linear in its coefficients, not in xx.

When to use

Reach for it when a straight line underfits curved data, using PolynomialFeatures in a pipeline with Linear Regression. The lecture says degree 2 to 3 is typical and degree 10 or more Overfitting badly. Lecture 8 notes that hand-guessing interactions explodes: 7 inputs give over 90 extra columns. Also in PS 02.

See DS Lec 4.

Appears in

No references yet.