Credit Card Fraud Detection

This project builds a robust system to detect fraudulent credit card transactions using classical machine learning models and a deep learning model with Focal Loss and using a dataset from Kaggle, which includes 284,807 credit card transactions with only 492 frauds, is highly unbalanced. It includes complete workflows for data processing, model training, evaluation, and reporting. achieving an F1-score of 0.85 and 0.86 PR_AUC for the positive class (fraud).

---

Project Structure

CREDIT-CARD-FRAUD-DETECTION/
├── config/                     # Configuration YAMLs
│   ├── data.yml
│   └── trainer.yml
├── data/                       # Input CSV files
│   ├── train.csv
│   ├── val.csv
│   ├── test.csv
│   └── trainval.csv
├── models/                     # Saved models and evaluation results
│   ├── <timestamped folders>/
│   └── focal_loss_checkpointsfocal_last/
├── credit_fraud_train.py       # Training for classical ML models
├── Focal_loss.py               # Training PyTorch model with Focal Loss
├── test_script.py              # Evaluation on test set
├── credit_fraud_utils_data.py  # Data loading, scaling, balancing
├── credit_fraud_utils_eval.py  # Evaluation metrics and plots
├── help.py                     # Helper functions (config/model IO)
├── EDA.ipynb                   # Exploratory Data Analysis notebook
└── README.md                   # 📄 You are here

---

Models Used

• Logistic Regression
• Random Forest
• K-Nearest Neighbors
• Neural Network (MLPClassifier)
• Voting Classifier (Ensemble)
• Custom Neural Network with Focal Loss (PyTorch)

---

Configuration Files

• config/data.yml – Paths, balancing, scaling
• config/trainer.yml – Training settings for each model, evaluation config

---

How to Run

1. Install Dependencies

pip install -r requirements.txt

Includes scikit-learn, pandas, matplotlib, seaborn, torch, mlxtend

---

2. Train Models (Classical ML)

python credit_fraud_train.py --data data.yml --trainer trainer.yml

---

3. Train Deep Model (Focal Loss)

python Focal_loss.py

---

4. Evaluate on Test Set

python test_script.py

---

Evaluation Outputs

Each trained model is saved in models/<timestamp>/ with:

• Classification Reports
• Confusion Matrix Heatmaps
• Precision-Recall Curves
• Threshold-based tuning results
• Voting classifier ensemble (if enabled)
• PR-AUC metrics

Results

Note:

• The optimal threshold calculated by finding highest F1-score and it's threshold in training data only.

---

Focal Loss Model (PyTorch)

Trained via custom architecture with:

BatchNorm → Tanh → Dropout → Focal Loss

Designed to handle extreme class imbalance with tunable alpha and gamma.

Focal-loss results

• I tried Server combination of Alpha (0.80-0.99, +0.5) and gamma (0-4, +1).

• The best result archieve by Alpha 0.75 and gamma 2.

  

• Notes:

  • Alpth and gamma sometimes unstables train using batchnorm make this effect less occur and switching from Adam to SGD also.
  • High gamma (5~7) gives very noisey loss curve.

---

Class Imbalance Handling

Supports multiple techniques:

• Class Weights (for LR, RF)
• Oversampling (do_balance)
• Focal Loss (for PyTorch NN)

---

Smote and undersampling technique

• SMOTE (Synthetic Minority Over-sampling Technique) is an oversampling method used to generate synthetic samples for the minority class. Despite experimenting with SMOTE, random over-sampling, and under-sampling techniques, the results on the validation data were poor.

• Smote (0.05-ratio) results:

• RandomUnderSampler (0.05-ratio) results:

---