💳 Credit Card Fraud Detection System

A comprehensive Machine Learning project analyzing 284,807 credit card transactions to detect fraudulent activity in real-time. This project tackles the challenge of extreme class imbalance using SMOTE and delivers a robust detection system with high recall and low false alarms.

---

📖 Project Overview

Credit card fraud is a "needle in a haystack" problem where fraudulent transactions are rare (0.17%) but financially devastating. This project aims to:

1. Analyze transaction patterns to visualize how fraud differs from normal spending using PCA & t-SNE.
2. Handle Imbalance using SMOTE (Synthetic Minority Over-sampling Technique) to prevent model bias.
3. Train ensemble classifiers (Random Forest, XGBoost) to distinguish fraud from legitimate transactions.
4. Recommend risk thresholds for banking systems to minimize customer friction while stopping theft.

🔑 Key Insights

• The 0.17% Challenge: The dataset is highly skewed. Without intervention (SMOTE), models would predict "Safe" 100% of the time and miss every fraud.
• Hidden Patterns: Dimensionality reduction (t-SNE) revealed that while fraud looks random, it actually forms distinct clusters in high-dimensional space.
• Model Trade-off: Logistic Regression caught the most fraud but flagged too many innocent people. Random Forest offered the best "Real-World" performance by virtually eliminating false alarms.

---

🛠️ Tech Stack

• Data Engineering: Python, Pandas, NumPy
• Visualization: Matplotlib, Seaborn (Correlation Heatmaps), Scikit-learn (PCA/t-SNE)
• Machine Learning:
  • Imbalanced-learn: SMOTE (Synthetic Minority Over-sampling)
  • Ensemble Models: Random Forest, XGBoost (for high precision)
  • Base Model: Logistic Regression (for baseline comparison)

---

📂 Project Structure

CreditCardFraudDetection/
│
├── src/
│   ├── data_loader.py      # Pipeline: Loads data, scales features, applies SMOTE
│   ├── visualization.py    # Reports: Generates Correlation Heatmap, PCA & t-SNE plots
│   └── model.py            # ML Core: Trains Logistic Regression, RF, and XGBoost
│
├── main.py                 # Entry Point: Runs the full analysis pipeline
├── requirements.txt        # Project Dependencies
├── README.md               # Documentation & Risk Analysis Report
│
└── (Generated Output)
    ├── data/creditcard.csv # The Dataset (Kaggle)
    ├── Figure_1.png        # Class Distribution (Imbalance)
    ├── Figure_2.png        # Feature Correlation Matrix
    └── Figure_3.png        # t-SNE Fraud Clusters

---

🚀 Installation & Usage

1️⃣ Clone & Install Dependencies

git clone https://github.com/xx-devvv/Credit-Card-Fraud-Detection.git
cd CreditCardFraudDetection
pip install -r requirements.txt

2️⃣ Setup Data

• Download the dataset from Kaggle (https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud)
• Place creditcard.csv inside the data/ folder

3️⃣ Run Analysis & Training

This single command runs the entire pipeline (EDA → SMOTE → Training → Evaluation):

python main.py

---

📊 Model Evaluation Results

Model	Recall (Fraud Capture)	False Alarms (False Positives)	Strength
Logistic Regression	92% (High)	~1,458 (High)	Good at catching fraud, but annoys too many genuine customers.
Random Forest	90% (Balanced)	~15 (Very Low)	Champion Model. Excellent precision; only stops a card when it's truly suspicious.
XGBoost	91% (High)	Low	Powerful gradient boosting alternative with high accuracy.

Technical Note: We prioritized Random Forest for the final recommendation because in a banking environment, blocking 1,400+ innocent users (as Logistic Regression did) causes significant reputation damage.

---

📢 Risk Management & Banking Recommendations

Based on our predictive analysis and confusion matrices, we recommend the following deployment strategies:

🏛️ Banking Policy

✅ Tiered Response System

• Score > 90% (Red Zone): Immediate auto-block of the transaction (Powered by Random Forest)
• Score 50–89% (Yellow Zone): Trigger Step-Up Authentication (SMS OTP / App Verification) instead of blocking

✅ Dynamic Thresholding

• Adjust the fraud threshold during peak shopping seasons (e.g., Black Friday) to reduce false positives when transaction volume spikes.

---

🛡️ Security Operations

✅ Feature Monitoring

• Correlation analysis showed that V14, V17, and V12 are the strongest indicators of fraud.
• Security teams should prioritize monitoring these vectors in raw logs.

✅ Continuous Retraining

• The t-SNE clusters indicate evolving fraud patterns.
• The model should be retrained weekly with new fraud labels to detect novel attack vectors.

---

👨‍💻 Author

Dev Pandey
Role: Software Engineer

---

📝 License

This project is open-source and available for educational purposes.