💳 Credit Card Fraud Detection – Recall-Focused Pipeline (In Progress)

This project detects fraudulent transactions using the Kaggle credit card fraud dataset, which contains highly imbalanced data — with fraud accounting for only ~0.17% of all records.

---

🎯 Objective

To build a machine learning pipeline that maximizes recall, catching as many fraud cases as possible.

---

⚙️ Approach Summary

Rather than oversampling or generating synthetic data, this project uses:

• class_weight='balanced' in Logistic Regression
• scale_pos_weight in XGBoost

These approaches adjust model focus without introducing noise or overfitting risks.

---

🛠️ Features Engineered

Feature	Description
Hour	Hour of transaction (0–23)
TimeBucket	6-hour time segments (e.g., 0–6 AM)
IsNight	Binary flag for night hours (22:00–06:00)
TimeSinceLastTx	Time delta since the previous transaction
TxInPastHour	Count of transactions in the past hour (rolling)
DayPart	Categorical: Morning, Afternoon, Evening, Night

PCA features V1–V28, Time, and Amount were also used as-is.

---

🧪 Modeling Pipeline

1. EDA & Feature Engineering

   • Visual patterns in fraud by time
   • Bimodal distribution in transaction frequency
   • Nighttime fraud concentration observed

2. Train/Test Split

   • Stratified 70/30 split
   • Scaled with StandardScaler

3. Model Training

   • ✅ Logistic Regression (class_weight='balanced')
   • ✅ XGBoost (scale_pos_weight)
   • ✅ Isolation Forest (unsupervised)
   • ✅ Ensemble: Logistic Regression + XGBoost

4. Evaluation Metrics

   • Recall (priority)
   • Precision, F1 Score, ROC AUC
   • Confusion Matrices
   • ROC & Precision-Recall curves

---

📈 Key Results

Model	Precision	Recall	F1 Score	ROC AUC
Logistic Regression	6.3%	87.8%	0.12	0.9666
XGBoost	87.8%	77.7%	0.83	0.9654
Isolation Forest	29.7%	29.7%	0.30	0.6480

---

✅ Project Milestones Completed

• Data loading and exploration
• Handling class imbalance (SMOTE, class weights)
• Feature engineering (time-based and PCA features)
• Correlation analysis and visualization
• Model training (Logistic Regression, XGBoost, Isolation Forest)
• Model evaluation (Confusion Matrix, ROC AUC, F1, Recall)
• Threshold tuning using Precision-Recall curve
• Ensemble model with soft voting (LR + XGBoost)
• Interpretability with XGBoost feature importances

---

📁 Project Structure

credit-card-fraud-detection/ ├── data/ # Input dataset (e.g., creditcard.csv) ├── notebooks/ # Jupyter notebooks (EDA, modeling) ├── outputs/ # Visualizations, metrics, exports ├── credit-card-fraud-detection.ipynb ├── README.md # Project overview └── LICENSE # License file

---

📚 Dataset

• Source: Kaggle - Credit Card Fraud Detection
• Type: PCA-transformed numeric features + engineered time-based features
• Imbalance: 492 frauds out of 284,807 transactions (~0.17%)

---

✍️ Author

Gleidy R.
LinkedIn

---

Feel free to fork, contribute, or share feedback!