You have a feature with 50,000 unique category values. How do you encode it?

medium

Answer

$One-hot is out — 50k sparse columns kill most models.$
$Options: (1) target encoding (with K-fold + smoothing) for gradient boosting or linear models; (2) frequency encoding to distinguish common vs rare; (3) feature hashing to a fixed-size vector; (4) entity embeddings — a learned dense vector per category, trained with a neural net or via factorization machines; (5) group rare categories into 'Other' below a min-count threshold.$
$Often combine target + frequency encoding.$

Check yourself — multiple choice

One-hot encode with a sparse matrix
Target encoding, frequency, hashing, entity embeddings, or grouping — pick the mix per model
Drop the feature
One-hot is the only correct answer

Target / frequency / hashing / embeddings / rare-bucket — no single answer, combine as needed.

#encoding#features#feature-engineering

215 interview questions in this topic.