06 - Feature Engineering¶

Get the Video Course

Feature engineering creates new variables that help the model detect patterns more effectively. This section introduces amount_ratio, a feature that captures how much a transaction deviates from the customer's normal spending pattern. Fraudsters typically spend far more than the victim's historical average, making this ratio a potential fraud indicator.

What is Feature Engineering?¶

We go from 8 input features (baseline) to 9 input features by adding amount_ratio. The diagram above shows the before/after comparison and how amount_ratio is calculated: transaction_amount / avg_transaction_amount. A legitimate transaction gives a ratio near 1.0x (normal spending), while a fraud transaction gives 25x (suspicious).

ML and CCFD Workflow¶

Architecture¶

Table of Contents¶

Pre-requisite: Python Environment Setup¶

# Create conda environment
conda create -n mlops-env1 python=3.14 -c conda-forge -y

# Activate environment
conda activate mlops-env1

# Install dependencies (locked versions)
cd ccfd-project
pip install -r requirements.txt

Note: All sections in this course use the same mlops-env1 environment. You only need to create it once. After that, just activate it with conda activate mlops-env1 before running any scripts.

Step-01: Understand the amount_ratio Feature¶

Formula¶

amount_ratio = transaction_amount / avg_transaction_amount

Why This Matters for Fraud Detection¶

• A customer with avg spending of $100 making a $50 purchase → ratio = 0.5 (normal)
• Same customer making a $2,500 purchase → ratio = 25.0 (suspicious!)
• Fraudsters don't know the victim's spending patterns, so they tend to spend much more than normal

Example Calculations¶

Step-02: Create the Feature¶

Prerequisite: Each section has its own ccfd-project/ folder. You must generate data first before running feature engineering.

cd 06_Feature_Engineering/ccfd-project
python p1_01_generate_initial_dataset.py    # Generate data (required first)
python p4_01_feature_engineering_fe1.py     # Create amount_ratio feature

Alternative: If you're using the shared folder, run all commands from ccfd-project-main/ instead. See Shared Project Folder in the root README.

Expected Output:

======================================================================
FEATURE ENGINEERING: Creating amount_ratio Feature
======================================================================

----------------------------------------------------------------------
STEP 1: Loading Raw Data
----------------------------------------------------------------------
Loaded 10,000 transactions (9 columns)

----------------------------------------------------------------------
STEP 2: Applying Feature Engineering
----------------------------------------------------------------------
Added: amount_ratio (9 total features)
Saved: data/credit_card_transactions_fe1_amount_ratio.csv

----------------------------------------------------------------------
STEP 3: Verifying Results
----------------------------------------------------------------------
Raw columns: 9 → FE columns: 10
New feature added: amount_ratio

======================================================================
FEATURE ENGINEERING COMPLETE!
======================================================================

   Output: data/credit_card_transactions_fe1_amount_ratio.csv

Step-03: Analyze the Feature¶

python p4_02_eda_amount_ratio_fe1.py

Expected Output:

======================================================================
EDA FOR FEATURE ENGINEERING 1: AMOUNT RATIO
======================================================================

======================================================================
SECTION 2: DISTRIBUTION ANALYSIS
======================================================================

   Legitimate mean: 2.81
   Fraud mean:      12.10
   Difference:      9.29   GOOD separation!
   Saved: eda_plots/07_fe1_amount_ratio_distribution.png

======================================================================
SECTION 3: CORRELATION COMPARISON
======================================================================

   transaction_amount correlation: 0.3111
   amount_ratio correlation:       0.2415
   Saved: eda_plots/08_fe1_amount_ratio_fraud_comparison.png
   Saved: eda_plots/09_fe1_amount_ratio_correlation.png

======================================================================
FE1 EDA COMPLETE!
======================================================================

Feature Analysis Results¶

Generated Plots¶

amount_ratio Distribution (Legit vs Fraud)¶

Fraud Rate by amount_ratio Range¶

Updated Correlation Heatmap (with amount_ratio)¶

Step-04: Train Model with amount_ratio¶

python p4_03_train_model_with_fe1.py

Expected Output:

======================================================================
Logistic Regression (FE1: amount_ratio) - PERFORMANCE METRICS
======================================================================

TECHNICAL METRICS:
   • F1 Score:     0.2347
   • Precision:    0.1386
   • Recall:       0.7667
   • ROC AUC:      0.8837

BUSINESS METRICS:
   • Net Benefit:    $39,700
   • Fraud Caught:   46
   • Fraud Missed:   14
   • False Alarms:   286

======================================================================
TRAINING SUMMARY
======================================================================

   Model: Logistic Regression (FE1)
   Features: 9 (includes amount_ratio)

   F1 Score: 0.2347
   Net Benefit: $39,700

======================================================================
FE1 TRAINING COMPLETE!
======================================================================

Step-05: Compare Results - Baseline vs Feature Engineered¶

Why Minimal Improvement?¶

The improvement is minimal (+$75) because:

1. Logistic Regression already captures amount patterns - The raw transaction_amount feature already has strong predictive power (+0.31 correlation)
2. amount_ratio provides redundant information - Since it's derived from transaction_amount and avg_transaction_amount, Logistic Regression can learn similar patterns from the raw features
3. More value comes from hyperparameter tuning - Section 07 shows that threshold optimization provides much larger improvements than feature engineering alone
4. Small dataset (10,000 synthetic transactions) - Real production fraud systems process millions of transactions per day. With only 60 fraud cases in our test set, small improvements from a single engineered feature get lost in statistical noise. At production scale, even a 0.5% lift translates to substantial business value (millions of transactions x ~$1,500 average fraud loss prevented).
5. Limited feature space (8 base features, 1 engineered) - Real-world fraud systems use 100+ input features (device fingerprint, IP velocity, merchant risk scores, transaction sequences, user behavior history) and typically ship 10-20+ engineered features over months of iteration. Our amount_ratio is illustrative of the technique, not representative of what production feature engineering looks like end-to-end.

Next Steps¶

05 - EDA and Preprocessing Next: 07 - Hyperparameter Tuning