Feature Engineering: Encoding & Scaling

Clean data is not yet model-ready

Cleaning fixes wrong data — blanks, duplicates, bad types. Feature engineering is the next step: reshaping correct data into the form a model can learn from. A feature is just an input column the model uses to predict. Most accuracy gains in real projects come from better features, not fancier models.

Models do maths, so they hit two walls with raw data: (1) they cannot read text categories like “Pune” or “Delhi”, and (2) they get confused when one column is in thousands and another in single digits. We fix the first with encoding and the second with scaling.

Problem 1: a model cannot read text

Take a city column with values like “Pune” and “Delhi”. A model needs numbers. The naive fix — Pune=0, Delhi=1, Mumbai=2 — is a trap: it tells the model Mumbai (2) is “twice” Pune (0), which is nonsense. The right fix is one-hot encoding: make one yes/no (1/0) column per category.

import pandas as pd

df = pd.DataFrame({
    'city':  ['Pune', 'Delhi', 'Pune', 'Mumbai'],
    'price': [60, 120, 65, 150]
})

encoded = pd.get_dummies(df, columns=['city'])
print(encoded)

Problem 2: columns on wildly different scales

Imagine predicting from age (around 20–60) and income (around 20,000–90,000). Many models judge “distance” between rows, and income’s big numbers would drown out age completely — not because age matters less, but because its numbers are smaller. Scaling rescales every column to a comparable range so each gets a fair say.

The most common method is standardisation with StandardScaler: it shifts each column to a mean of 0 and rescales it by its spread, so a value becomes “how many standard deviations above or below average”.

from sklearn.preprocessing import StandardScaler
import pandas as pd

df = pd.DataFrame({
    'age':    [20, 40, 60],
    'income': [20000, 50000, 90000]
})

scaler = StandardScaler()
scaled = scaler.fit_transform(df)
print(scaled.round(2))

A taste of creating new features

Feature engineering also means inventing useful columns from ones you have — especially from dates. A raw date is hard to model; the month or day-of-week hidden inside it often is not.

import pandas as pd

df = pd.DataFrame({'order_date': ['2026-01-15', '2026-07-04', '2026-12-25']})
df['order_date'] = pd.to_datetime(df['order_date'])

df['month']   = df['order_date'].dt.month
df['weekday'] = df['order_date'].dt.day_name()
print(df)