Skip to content

Working with Data

Feature Engineering & Crosses

7 min+70 XPStep 1/10

When 1 + 1 is greater than 2

Two features can be nearly useless alone and powerful together. Latitude alone tells you the average price of a horizontal stripe of a city; longitude alone, a vertical stripe. Neither pins down a neighborhood — but the pair locates a street corner.

Feature engineering means turning domain knowledge into new columns: ratios, time-since-event, day-of-week. The sharpest tool in the kit is a feature cross — a synthetic feature combining two or more inputs so the model can see their interaction.

analogy

Latitude and longitude are like a Battleship guess: a row or a column means little. The hit only happens where they cross.