ML//feature selection

Cutting a candidate feature set down to the few that carry signal, before trusting an evaluation.

Cutting a candidate feature set down to the few that carry signal, before trusting an evaluation.

In rare-event problems it is debt control: every extra feature is a loan against evidence you probably do not have. It fights the curse of dimensionality.

Redundant features (two ways of saying the same thing) get dropped by correlation and by meaning: keep one reading, drop synonyms.

Fewer features means a model small enough that a human can still argue with it, and less room for overfitting.