Encoding categories and dates

Feature engineering

Feature engineering means creating the inputs a model learns from. A raw order timestamp like "2025-10-25 21:40" is nearly useless to most models. "Saturday", "9 p.m." and "5 days before Diwali" are very useful. Good features often matter more than the choice of algorithm.

One-hot encoding

For a category without order, such as payment method (UPI, Card, COD), make one column per category with 1 if the row has it and 0 otherwise. Why not just number them UPI = 1, Card = 2, COD = 3? Because a model would think COD is "more" than UPI and that Card is halfway between, which is meaningless.

import pandas as pd
from sklearn.preprocessing import OneHotEncoder

pd.get_dummies(df["payment"], prefix="pay")                  # quick, in pandas
enc = OneHotEncoder(handle_unknown="ignore").fit(X_train[["payment"]])   # safe in pipelines

Ordinal encoding

When categories have a real order, such as education level (10th < 12th < Graduate < Postgraduate) or T-shirt size (S < M < L), map them to ordered numbers. Tree models can also handle plain ordinal codes for unordered categories reasonably well, but linear models and k-NN cannot.

Many categories

A pin-code column has thousands of values; one-hot would create thousands of mostly-zero columns. Options: group rare categories into "Other", use frequency encoding (replace each category by how often it appears), or target encoding (replace it by the average target for that category). Target encoding is powerful but leaks the answer unless it is computed on training folds only; libraries do this with cross-fitting.

Dates and times

• Parts: year, month, day of month, day of week (Monday = 0 in Python's weekday()), hour.
• Flags: is_weekend, is_month_end (salary days change spending), is_holiday or "days to Diwali".
• Differences: account age in days, days since last order.
• Cyclical features: hour 23 is next to hour 0; some models benefit from sine and cosine encodings of hour and month.

from datetime import date

d = date.fromisoformat("2025-10-25")
print(d.year, d.month, d.weekday(), d.weekday() >= 5)

Text and other raw columns

Simple numeric features can be pulled from text: length of a review, number of exclamation marks, whether an address contains a flat number. Module 10 goes much further with text.

Common mistakes

• Label-encoding unordered categories for a linear model. Fix: one-hot.
• Building the category list from all data, then crashing on new categories in production. Fix: fit on training data with handle_unknown="ignore".
• Target encoding computed on the whole dataset. Fix: compute it inside training folds.
• Keeping the raw date string as a category. Fix: extract parts and differences.