OpenToonz and Canva Digital Art Animation in Silom, Bangkok

1. What Feature Engineering Is

Feature engineering builds new columns from existing ones using domain knowledge and simple math. Well-chosen features often improve accuracy more than swapping algorithms.

2. Five Practical Techniques

Date and temporal decomposition

Raw timestamps are hard for many estimators. Split them into year, month, and day of week to expose seasonality and promo cycles.

df["Sign-up Date"] = pd.to_datetime(df["Sign-up Date"])
df["Sign_up_Year"] = df["Sign-up Date"].dt.year
df["Sign_up_Month"] = df["Sign-up Date"].dt.month
df["Sign_up_DayOfWeek"] = df["Sign-up Date"].dt.day_name()

Ratio features

Absolute bills hide relative burden. Bill-to-income ratio = Monthly Bill / Monthly Income shows whether a charge is heavy for that customer.

df["Bill_to_Income_Ratio"] = df["Monthly Bill"] / df["Monthly Income"]

Interaction features

Multiply related risk signals. Support tickets × monthly bill spikes when high-paying customers also open many tickets; zero tickets keep the product at zero.

df["Support_Bill_Interaction"] = df["Support Tickets"] * df["Monthly Bill"]

Text metadata

Character length of free-text feedback is a cheap numeric signal: longer notes often mean more serious issues.

df["Feedback_Length"] = df["Customer Feedback"].astype(str).apply(len)

Rule-based keyword flags

A binary complaint flag from keywords (slow, poor, expensive, unreliable, drops) alerts the model without full NLP.

complaint_keywords = ["slow", "poor", "expensive", "unreliable", "drops"]
pattern = "|".join(complaint_keywords)
df["Complaint_Flag"] = df["Customer Feedback"].str.contains(
    pattern, case=False, na=False
).astype(int)

3. Check the New Columns

engineered_cols = [
    "Sign_up_Month", "Bill_to_Income_Ratio",
    "Support_Bill_Interaction", "Feedback_Length", "Complaint_Flag"
]
print(df[engineered_cols].head())

These columns give direct signals for seasonality, affordability, frustration, complaint detail, and service issues without changing the original raw data.