Feature Engineering

Overview

Feature engineering is where most model performance is won or lost. The aim is to express the signal in a form the model can use, while never letting information from the target or the test set leak into a feature.

When to use

• After cleaning, before/iterating with modeling.
• A model plateaus and you suspect under-expressed signal.
• You have raw datetime, text, or relational data to turn into columns.

Encoding categoricals