Stop 05 · Exploratory Data Analysis & Feature EngineeringTopics · 5/9

Encoding Categorical Variables

In short

Models only understand numbers, but your data says 'pune', 'delhi' and 'premium'. How you turn words into numbers can quietly help or wreck a model.

Converting text categories into numbers a model can use — one-hot encoding for nominal categories, label/ordinal encoding when there's a genuine order, and more advanced methods for high-cardinality features.

The mascot ticks one of three boxes on a slip next to a cup of chai, a coffee mug and a green tea cup.

Think of it like this

The office pantry order slip

The pantry slip has three tick boxes: chai, coffee, green tea. You tick one. Nobody writes 'chai = 1, coffee = 2', because coffee isn't 'more' than chai. One-hot encoding is that tick-box slip, one box per option.

Watch it happen

Turning a city column into numbers

pune

delhi

pune

goa

1/4A text column. The model can't do maths on words.

The big ideas

What you need to know

One-hot for no order

Cities, colours, payment methods. One 0/1 column per value, no fake ranking.

Ordinal for real order

Sizes S < M < L, or education levels. Map them to 0, 1, 2 so the order is kept.

Many values need care

A column with 5,000 pin codes makes 5,000 columns. Group rare values or try target or frequency encoding.

Plan for new values

Test data may contain a city training never saw. Set handle_unknown='ignore' so it doesn't crash.

See it in code

Walk through the code

python

1import pandas as pd2from sklearn.preprocessing import OneHotEncoder3 4df = pd.DataFrame({"city": ["pune", "delhi", "pune", "goa"],5                   "size": ["S", "L", "M", "S"]})6 7df["size_code"] = df["size"].map({"S": 0, "M": 1, "L": 2})   # real order8 9enc = OneHotEncoder(handle_unknown="ignore", sparse_output=False)10city = enc.fit_transform(df[["city"]])11print(enc.get_feature_names_out())   # ['city_delhi' 'city_goa' 'city_pune']12print(city)

1/4Two text columns: city (no order) and size (has an order).

Ordinal mapping for size, which has a real order, and one-hot encoding for city, which doesn't.

Save yourself the pain

Common mistakes

Oops

Label-encoding cities and feeding them to linear regression.

Do this instead

Use one-hot for unordered categories in linear models. Tree models cope better with label codes.

Oops

Using pd.get_dummies separately on train and test.

Do this instead

Columns may not match. Fit one OneHotEncoder on train and reuse it on test.

Your mission

Encode a real column

~20 min

Quick check

A column 'payment_method' has UPI, card and cash. Which encoding fits best for logistic regression?

Go deeper

1One-hot encode a categorical column and inspect the resulting feature columns.
2Learn when label encoding is appropriate (ordinal data) versus misleading (nominal data).
3Research target encoding as an option for high-cardinality categorical features.