Feature Engineering Basics

What is feature engineering?

Feature engineering is the process of converting raw data into useful inputs (features).

Examples:

• Date → day of week
• Amount → log(amount)
• Text → length, presence of keywords
• Customer transactions → total spend, avg order value

Example: derive features from datetime

Datetime features

import pandas as pd
 
df = pd.DataFrame({
    "order_time": pd.to_datetime([
        "2025-01-01 10:15:00",
        "2025-01-02 18:30:00",
        "2025-01-03 09:10:00",
    ]),
    "amount": [250, 180, 90],
})
 
df["hour"] = df["order_time"].dt.hour
df["weekday"] = df["order_time"].dt.day_name()
df["is_weekend"] = df["order_time"].dt.weekday >= 5
 
print(df)

Example: ratios and flags

Ratios

import pandas as pd
 
df = pd.DataFrame({"revenue": [1000, 500], "users": [100, 20]})
 
df["revenue_per_user"] = df["revenue"] / df["users"]
df["is_high_value"] = df["revenue_per_user"] > 20
 
print(df)

Example: aggregated features (groupby)

Aggregations

import pandas as pd
 
orders = pd.DataFrame({
    "customer_id": [1, 1, 2, 2, 2],
    "amount": [100, 250, 90, 180, 300],
})
 
cust = orders.groupby("customer_id").agg(
    total_spend=("amount", "sum"),
    avg_order=("amount", "mean"),
    orders=("amount", "count"),
).reset_index()
 
print(cust)

Guiding principles

• Don’t leak target/future information.
• Prefer simple features first.
• Validate with charts and summary stats.
• Document each feature and its meaning.

Visualize it

diagram Feature engineering flow mermaid

  flowchart LR
  A["order_time (raw)"] --> B["Extract hour, weekday, is_weekend"]
  C["revenue, users (raw)"] --> D["Derive revenue_per_user ratio"]
  E["customer transactions (raw)"] --> F["Aggregate: total_spend, avg_order"]
  B --> G["Engineered features"]
  D --> G
  F --> G

Raw columns are transformed and combined into new, more useful features for modeling.

From the book

McKinney’s data["food"].map(meat_to_animal) example is a textbook feature-engineering move: turn one raw column into a new, more useful signal using a lookup dictionary. .map() on a Series and .groupby().agg() on a DataFrame are the two workhorses behind most features you’ll ever engineer.

🧪 Try It Yourself

Exercise 1 – Derive a feature with map

pch.exerciseTag pch.exerciseTitle

Exercise 2 – Extract datetime parts

pch.exerciseTag pch.exerciseTitle

Exercise 3 – Aggregate a feature per group

pch.exerciseTag pch.exerciseTitle

Next

You’ve now covered the whole phase — wrap it up by making sure raw columns have the right Data Type Conversion and Validation before any of these features are built.