Credit Card Fraud Detection

Detect fraudulent transactions from 28 PCA-transformed features using Logistic Regression, Random Forest, and XGBoost.

Problem

Credit card fraud accounts for billions in annual losses. With only 0.17% of transactions being fraudulent, this is an extreme class imbalance problem. The dataset (Kaggle) contains 284,807 European cardholder transactions with PCA-transformed features for confidentiality.

Approach

Step	What
EDA	Class distribution, feature correlations, amount analysis
Feature Engineering	Cyclical time encoding (hour sin/cos), scaled Amount
Modeling	3 models with default params + class imbalance handling
Evaluation	PR-AUC (precision-recall), confusion matrices, feature importance

Results

Model	PR-AUC	Precision	Recall	F1
XGBoost	0.819	0.913	0.768	0.834
Random Forest	0.815	0.959	0.737	0.833
Logistic Regression	0.676	0.056	0.874	0.105

Best for production: Random Forest — near-identical PR-AUC to XGBoost but with higher precision (0.959), meaning fewer false alarms for fraud teams.

Key insight: V14 is the strongest fraud predictor, followed by V17 and V10. Tree-based models outperform logistic regression, confirming non-linear patterns in the data.

How to Run

1. Download from Kaggle and save as data.csv
2. pip install -r requirements.txt
3. jupyter notebook fraud_detection.ipynb

Built With

pandas, scikit-learn, matplotlib, seaborn, XGBoost