Time-Based and Domain-Driven Features

Concept

Dates and timestamps contain structure that raw strings hide. Domain knowledge helps decide which parts of that structure represent useful, prediction-time signals.

Date-Derived Features

events["timestamp"] = pd.to_datetime(events["timestamp"])
events["hour"] = events["timestamp"].dt.hour
events["day_of_week"] = events["timestamp"].dt.dayofweek
events["month"] = events["timestamp"].dt.month
events["is_weekend"] = events["day_of_week"].isin([5, 6]).astype(int)

Signup dates can become account age; last-purchase dates can become days since last purchase. For e-commerce, repeat-purchase rate, recency, and basket size can represent customer behavior. Finance may use debt-to-income or utilization; housing may use property age or rooms per area.

Time Is Cyclical

Hour 23 and hour 0 are adjacent in reality but far apart as raw numbers. Sine/cosine encodings can represent this cycle when time-of-day patterns matter. The goal is not trigonometry; it is avoiding a representation that treats midnight as distant from 11 PM.

Validate Domain Intuition

Domain-inspired does not mean automatically valuable. days_until_churn is invalid for churn prediction because it knows the outcome. Compare a feature using the same validation setup as the baseline and keep it only when it improves useful held-out behavior.

Practice Questions

1. Create two features from a signup date.
2. Why can hour require cyclic representation?
3. Is days until churn valid for churn prediction?
4. Name an e-commerce recency feature.
5. Why must domain intuition be validated?

Next Lesson

Next, test whether engineered features are valid, stable, and worth maintaining.