💳 Credit Card Fraud Detection with Imbalanced Data Handling

📌 Problem Statement

Fraud detection is a classic example of an extremely imbalanced classification problem, where fraudulent transactions represent only ~1–2% of total data.

In such cases, accuracy becomes misleading. A model predicting "no fraud" for every transaction would still achieve 98%+ accuracy.

This project focuses on building and evaluating models using appropriate metrics and threshold tuning strategies.

---

📊 Dataset

Synthetic dataset generated using make_classification with:

• 10,000 transactions
• 10 features
• 1.6% fraud rate
• Severe class imbalance (≈ 98% normal, 2% fraud)

---

⚙️ Approach

1. Generated highly imbalanced dataset
2. Applied feature scaling using StandardScaler
3. Compared two models:
   • Logistic Regression
   • Random Forest
4. Evaluated using:
   • ROC-AUC
   • Precision-Recall Curve
   • F1-score
5. Tuned classification threshold to optimize fraud detection performance

---

📈 Model Performance

Logistic Regression

• ROC-AUC: 0.69
• Poor precision on fraud class
• High false positives

Random Forest (Default Threshold = 0.50)

• ROC-AUC: 0.81
• Fraud F1-score: 0.31
• Low recall on fraud

Random Forest (Tuned Threshold = 0.13)

• Fraud F1-score improved to 0.56
• Fraud Precision: 0.78
• Fraud Recall: 0.44

Threshold tuning significantly improved fraud detection performance without drastically increasing false positives.

---

📉 Key Insights

• Accuracy is not reliable for imbalanced datasets
• Precision-Recall curve is more informative than ROC in rare-event detection
• Default threshold (0.5) is not always optimal
• Threshold tuning can dramatically improve real-world fraud detection systems

---

🛠 Tech Stack

• Python
• Pandas
• NumPy
• Scikit-learn
• Matplotlib
• Seaborn

---

▶️ How to Run

1️⃣ Install dependencies

pip install -r requirements.txt