🚀 NovaPay Fraud Detection System

An end-to-end machine learning project designed to detect fraudulent transactions in a highly imbalanced financial dataset, with a strong focus on performance, interpretability, and real-world usability.

---

📌 Project Overview

Fraud detection is a critical challenge in fintech, where fraudulent transactions are rare but highly impactful. In this project, we built a robust and interpretable fraud detection system for NovaPay that:

• Accurately identifies fraudulent transactions
• Handles severe class imbalance (<10% fraud rate)
• Provides transparent, audit-ready explanations for each prediction

---

⚙️ Key Features

• ✅ Behavioural risk scoring system
• ✅ Advanced feature engineering (velocity, anomaly, interaction features)
• ✅ Time-based model validation (prevents data leakage)
• ✅ LightGBM & XGBoost model comparison
• ✅ Imbalance handling (SMOTE, class weights, undersampling)
• ✅ SHAP-based explainability
• ✅ Fraud review template with reason codes
• ✅ Deployment-ready (FastAPI + Docker)

---

📊 Model Performance

Metric	Score
Accuracy	~98%
ROC-AUC	~0.97–0.98
F1-score	~0.95

👉 Final model: LightGBM (Baseline)

• Best balance between precision and recall
• Stable without heavy rebalancing

---

🧠 Approach

1. Exploratory Data Analysis

Identified key fraud patterns:

• High transaction velocity
• Risky IP/device behaviour
• New accounts

---

2. Risk Scoring

• Built a behavioural risk score using domain-driven rules
• Normalized into interpretable risk levels

---

3. Feature Engineering

Customer behaviour:

• Transaction frequency, averages, totals

Anomaly detection:

• Velocity ratios, deviations

Interaction features:

• IP × device risk
• Velocity × amount

---

4. Model Training

Models used:

• LightGBM
• XGBoost

Techniques tested:

• SMOTE
• Class weighting
• Undersampling

---

5. Explainability (SHAP)

Implemented SHAP for:

• Global feature importance
• Transaction-level explanations

Each prediction includes:

• Prediction (Fraud / Legitimate)
• Confidence score
• Risk level (Low / Medium / High)
• Top risk drivers
• Protective factors
• Business-friendly reason codes

---

🔍 Example Output

Prediction: FRAUD
Confidence: 100%

Top Risk Drivers:

• High behavioural risk score
• New account
• High transaction velocity
• Suspicious IP/device interaction

---

🖥️ Interactive Dashboard (Streamlit)

A user-friendly interface allows stakeholders to:

• Input transaction details
• View fraud prediction instantly
• See:
  • Risk level (colour-coded)
  • Reason codes
  • Full fraud review summary

👉 Transforms model output into actionable insights

---

🏗️ Deployment

The model is production-ready and deployed using:

FastAPI → REST API for predictions Docker → Containerized deployment

---

Project Structure

nova-fraud-api/ │ ├── app/ │ ├── main.py │ ├── utils.py │ ├── artifacts/ │ ├── lgb_model.pkl │ ├── scaler.pkl │ ├── encoders.pkl │ ├── features.pkl │ ├── shap_explainer.pkl ├── notebooks/ ├── requirements.txt ├── Dockerfile

---

💡 Key Insights

• Feature engineering had the biggest impact on performance
• LightGBM handled imbalance effectively without heavy resampling
• Fraud detection requires balancing:
  • Recall (catch fraud)
  • Precision (avoid false alarms)
• Explainability is critical for:
  • Trust
  • Regulatory Compliance
  • Analyst decision-making

---

🚀 Future Improvements

• Real-time streaming fraud detection
• Threshold optimisation for business strategies
• Hybrid system (rules + ML)
• Model monitoring & drift detection
• Integration with transaction systems

---

👩‍💻 Author

Priscillia Ejiro Data Scientist | Machine Learning | Fraud Analytics