Label Encoding

What is label encoding?

Label encoding converts categories into integers:

• high → 2

When it’s OK

• For ordinal categories (where order matters):

  • low < medium < high

• For tree models sometimes (but still be careful)

When to avoid it

For nominal categories (no order):

• city, color, country

Label encoding may trick models into believing an order exists.

Example: ordinal encoding (manual)

Manual ordinal encoding

import pandas as pd
 
df = pd.DataFrame({"priority": ["low", "high", "medium", "low"]})
 
mapping = {"low": 0, "medium": 1, "high": 2}
df["priority_code"] = df["priority"].map(mapping)
 
print(df)

Using scikit-learn LabelEncoder

LabelEncoder is mainly designed for encoding target labels, not features.

LabelEncoder

from sklearn.preprocessing import LabelEncoder
 
le = LabelEncoder()
y = ["spam", "ham", "ham", "spam"]
print(le.fit_transform(y))
print(le.classes_)

Tip

Use:

• One-hot encoding for nominal categories
• Ordinal mapping for ordered categories

Visualize it

diagram Label encoding flow mermaid

  flowchart LR
  A["priority column
(low, medium, high)"] --> B["mapping: low->0, medium->1, high->2"]
  B --> C["priority_code column
(0, 1, 2)"]

Each distinct category is assigned an integer code; order is preserved only when the categories are truly ordinal.

From the book

McKinney describes this exact idea for pandas’ Categorical type: the distinct values are called categories, and the integers that reference them are called codes. Label encoding is that same dictionary-encoded representation — df["col"].astype("category").cat.codes gives you integer codes for free, without needing scikit-learn at all.

🧪 Try It Yourself

Exercise 1 – Encode with a manual mapping

pch.exerciseTag pch.exerciseTitle

Exercise 2 – Encode with scikit-learn’s LabelEncoder

pch.exerciseTag pch.exerciseTitle

Exercise 3 – Get codes from a category dtype

pch.exerciseTag pch.exerciseTitle

Next

For categories with no natural order, prefer One-Hot Encoding so the model doesn’t assume a ranking that isn’t real.