- 📌 Project Overview
- 🔍 Key Challenges Addressed
- 📊 Dataset
- 📊 Dataset Link
- 🎥 Final Presentation Link
- ⚙️ Methodology
- 📈 Evaluation Metrics
- 🚀 Results
- 🔬 Key Findings
- 🔮 Future Improvements
- 💻 Installation and Usage
- 📋 Project Structure
- 📌 Requirements
- 🚀 Setup and Execution
- 🔗 References
This project implements advanced data mining techniques to detect fraudulent credit card transactions.
With global fraud losses exceeding $28.65 billion annually, we aim to tackle key challenges such as imbalanced data, computational efficiency, and evolving fraud patterns.
- Severe Class Imbalance: Only 0.17% (492 out of 284,807) transactions are fraudulent
- Computational Complexity: 28 PCA-transformed features requiring efficient processing
- Evolving Fraud Patterns: Adapting to changing fraud tactics over time
We utilized the Credit Card Fraud Detection Dataset from Kaggle, which contains:
✔ 284,807 transactions (492 fraudulent cases)
✔ 28 principal components (from PCA transformation)
✔ Time elapsed features and transaction amounts
✔ Anonymized data to protect user privacy
- creditcard.csv - The original credit card transaction dataset from Kaggle
- smote_processed_data.csv - Dataset after applying SMOTE for class balancing
- train_preprocessed.csv - Preprocessed training dataset
- test_preprocessed.csv - Preprocessed testing dataset
Credit Card Fraud Detection Final Project presentation Download Link
✅ Feature scaling with RobustScaler
✅ Correlation analysis and feature selection
✅ Class Imbalance Handling: Implemented SMOTE (Synthetic Minority Over-sampling Technique)
✅ Outlier Detection: Used Isolation Forest to identify and handle anomalies
✅ Stratified train-test splitting
✔ Configuration: 100 trees, max depth of 10, balanced class weights
✔ ROC-AUC Score: 0.979
✔ Threshold Optimization: Improved F1 score from 0.5784 to 0.8021 (38.67% improvement)
✔ Feature importance analysis
✔ Configuration: For computational efficiency, we optimized our original Random Forest model (100 trees, max depth of 10) to a lighter version (50 trees, max depth of 8, increased min_samples_split, parallelized processing) ✔ ROC-AUC Score: 0.979 ✔ Threshold Optimization: Improved F1 score from 0.5784 to 0.8021 (38.67% improvement) ✔ SHAP Analysis Enhancements: Unlike traditional feature importance measures in the base Random Forest, SHAP analysis provides detailed per-prediction feature contribution values, revealing how specific features like V14, V10, and V4 individually influence model decisions, creating a more transparent and interpretable fraud detection system
✔ Architecture: Three hidden layers (64, 32, 16 neurons) with batch normalization
✔ Activation: ReLU for hidden layers, Sigmoid for output
✔ Optimization: Adam with learning rate of 0.001
✔ Dropout layers (0.3) to prevent overfitting
- Precision, Recall, and F1-Score
- ROC Curve and AUC
- Confusion Matrix Analysis
- Probability Distribution Analysis
| Model | Precision | Recall | F1-Score | AUC |
|---|---|---|---|---|
| Random Forest | 0.96 | 0.83 | 0.89 | 0.97 |
| Neural Network | 0.89 | 0.91 | 0.90 | 0.96 |
- Random Forest: Better interpretability with clearer feature importance rankings
- Neural Network: Superior at capturing complex fraud patterns with fewer false positives
- Threshold Optimization: Critical for improving model performance in imbalanced datasets
- Model Performance Enhancement: Neural Network achieved high recall (0.85) and precision (0.87)
- SMOTE's Role: Transformed dataset from 0.17% fraudulent transactions to a balanced distribution
- Comparative Insights: Random Forest provided better interpretability while Neural Network reduced false positives
- Implement real-time fraud detection capability
- Explore ensemble methods for improved performance
- Develop an adaptive learning approach for evolving fraud patterns
- Create a web-based dashboard for monitoring fraud detection metrics
- Implement XAI techniques (SHAP, LIME) for better model interpretability
credit-card-fraud-detection/
├── code/ # Source code files
│ ├── models/ # Model implementation
│ │ ├── part4-Random forest.ipynb # Random Forest model
│ │ ├── part4b-Random forest(SHAP).ipynb # Random Forest with SHAP analysis model
│ │ └── part5-Neural network.ipynb # Neural Network model
│ └── preprocessing/ # Data preprocessing pipeline
│ ├── part1_credit_card_fraud_preprocessing.ipynb # Feature scaling & initial preprocessing
│ ├── part2_credit_card_fraud_smote.ipynb # SMOTE implementation
│ └── part3-outlier_detection_and_models.ipynb # Outlier detection
├── image/ # Visualization images
│ ├── SMOTE/ # SMOTE-related visualizations
│ ├── neural network/ # Neural network results & visualizations
│ ├── outliter detection/ # Outlier detection visualizations
│ ├── preprocessed image/ # Data preprocessing visualizations
│ └── random foreast/ # Random Forest results & visualizations
├── tableau/ # Tableau files and visualizations
│ └── tableau dashboard.png # Dashboard visualization
├── README.md # Project documentation
├── banner.png # Project banner image
└── workflow.pdf # Workflow diagram
numpy==1.26.4
pandas==2.2.1
scikit-learn==1.4.2
tensorflow==2.15.0
matplotlib==3.8.3
seaborn==0.13.2
imbalanced-learn==0.12.0
# Clone the repository
git clone https://github.com/thisissophiawang/credit-card-fraud-detection.git
cd credit-card-fraud-detection
# Install dependencies
pip install -r requirements.txt
# Run the preprocessing pipeline
python src/preprocessing.py
# Train and evaluate models
python src/train_models.py- Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5-32.
- Chawla, N. V., et al. (2002). SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research, 16, 321-357.
- Dal Pozzolo, A., et al. (2015). Calibrating Probability with Undersampling for Unbalanced Classification. IEEE Symposium Series on Computational Intelligence and Data Mining.
- Fernández, A., et al. (2018). Learning from Imbalanced Data Sets. Springer.