How to Build Features using Domain Knowledge¶

The best algorithms in the world cannot discover business logic that isn't physically present in the training matrix.

What You Will Learn¶

Transform institutional business rules explicitly into code
Write vectorised operations to map continuous values conditionally

Step 1: Flagging High-Value Interactions¶

A Retail dataset might possess hundreds of columns, but the business structurally defines a "VIP Customer" purely via an arbitrary threshold: someone who spent $> \$100$ and registered $> 5$ years ago.

An algorithm will struggle blindly to discover this exact boundary linearly. You must explicitly inject this Domain Knowledge geometrically!

import pandas as pd
import numpy as np
import seaborn as sns

# We utilize Pandas DataFrame manipulation without external libraries 
df = pd.DataFrame({
    'total_spend': [50, 150, 200, 45],
    'years_registered': [1, 6, 2, 8]
})

# Condition: Spend > 100 AND Years > 5
condition = (df['total_spend'] > 100) & (df['years_registered'] > 5)

# np.where executes this vectorised across 10 million rows instantly
df['is_vip'] = np.where(condition, 1, 0)

print(df)

Expected Output

   total_spend  years_registered  is_vip
0           50                 1       0
1          150                 6       1
2          200                 2       0
3           45                 8       0

Step 2: Complex Business Logic (np.select)¶

Instead of a binary 1/0 VIP status dynamically, your company might have complex tiered pricing logically.

Instead of writing dangerously slow .apply(lambda x: ...) functions natively, np.select() structurally chains multiple conditions simultaneously.

df_flights = sns.load_dataset('flights')

# We know historically logically that July/August are "Peak", 
# Dec/Jan are "Holiday", and everything else is "Off-Peak"

conditions = [
    df_flights['month'].isin(['Jul', 'Aug']),
    df_flights['month'].isin(['Dec', 'Jan'])
]

choices = ['Peak', 'Holiday']

# The 'default' is what happens geographically if no conditions trigger 
df_flights['seasonality_tier'] = np.select(conditions, choices, default='Off-Peak')

print(df_flights.sample(5, random_state=1))

Expected Output

     year month  passengers seasonality_tier
94   1956   Nov         271         Off-Peak
73   1955   Feb         233         Off-Peak
55   1953   Aug         272             Peak
115  1958   Aug         505             Peak
67   1954   Aug         293             Peak

Workplace Tip

Whenever a domain expert logically tells you "Sales explicitly drop whenever it rains consecutively after a bank holiday", mechanically encode that EXACT sentence recursively into a mathematical feature is_rain_after_holiday = [0,1]. Domain experts understand causal reality natively better than XGBoost does experimentally!

KSB Mapping¶

KSB	Description	How This Addresses It
K4.2	Advanced analytics and ML techniques	Feature selection algorithms and dimensionality reduction
K5.2	Data formats and structures	Encoding categorical variables, handling mixed feature types
S2	Data engineering	Creating and transforming features from raw data
S4	Feature selection and ML	Applying feature selection methods and PCA
B1	Inquisitive approach	Exploring creative feature engineering strategies