🛡️ Fraud Detection Using Machine Learning

End-to-end project for detecting fraudulent online transactions — from data cleaning and exploratory analysis, through feature engineering and model training, to a deployable Gradio app for inference. Clean, reproducible, and structured for both academic evaluation and portfolio presentation.

📋 Overview

Domain: Transaction fraud detection

Approach: Feature engineering + XGBoost (with careful imbalance handling)

Deliverables: Jupyter notebook, trained model artifacts, Gradio app, sample inputs, and presentation slides

Why this project: demonstrates handling of extreme class imbalance, preventing data leakage (SMOTE timing), model tuning (scale_pos_weight), and production-ready inference.

🎯 Key Features

• High-Performance Model: XGBoost-based classifier optimized for fraud detection
• Advanced Feature Engineering: Temporal, cyclical, and behavioral pattern analysis
• Interactive Web Interface: User-friendly Gradio application for real-time predictions
• Comprehensive Analysis: Detailed exploratory data analysis and feature importance insights
• Production-Ready: Scalable architecture with saved model artifacts for deployment

🏗️ Architecture

┌─────────────────────────────────────────┐
│         Input Data (CSV)                │
└────────────┬────────────────────────────┘
             ▼
┌─────────────────────────────────────────┐
│      Feature Engineering Pipeline       │
│  • Temporal Features                    │
│  • Cyclical Encoding                    │
│  • Device Frequency Mapping             │
│  • Geographic Categorization            │
└────────────┬────────────────────────────┘
             ▼
┌─────────────────────────────────────────┐
│       Data Preprocessing                │
│  • Label Encoding                       │
│  • Min-Max Scaling                      │
│  • Feature Selection                    │
└────────────┬────────────────────────────┘
             ▼
┌─────────────────────────────────────────┐
│      XGBoost Classifier                 │
│  • Optimized Hyperparameters            │
│  • Probability Threshold: 0.45          │
└────────────┬────────────────────────────┘
             ▼
┌─────────────────────────────────────────┐
│    Prediction & Risk Assessment         │
└─────────────────────────────────────────┘

🚀 Quick Start

Prerequisites

Python 3.10.0 or higher
pip package manager

Installation

1. Install dependencies

pip install -r requirements.txt

2. Download model artifacts Ensure the following pre-trained model files are in your project directory:

• best_xgb_model.joblib
• min_max_scaler.joblib
• device_counts_map.joblib
• top_18_countries.joblib
• time_bins.joblib
• time_labels.joblib
• Label encoder files (le_*.joblib)

3. Run the application

python app.py

4. Access the web interface Open your browser and navigate to http://localhost:7860

📊 Features & Methodology

Feature Engineering

Feature Category	Description	Techniques Used
Temporal Features	Time-based patterns between signup and purchase	Time difference categorization, Hour/Day extraction
Cyclical Encoding	Circular representation of time variables	Sine/Cosine transformations for hours, days, months
Device Analytics	Device usage patterns and frequency	User count mapping per device ID
Geographic Features	Country-based risk assessment	Top 18 countries + "Other" categorization
Behavioral Patterns	User interaction characteristics	Source, Browser, Demographics analysis

Model Performance

• Algorithm: XGBoost (Extreme Gradient Boosting)
• Threshold Optimization: Custom threshold of 0.45 for optimal precision-recall balance
• Cross-validation: Stratified k-fold validation ensuring robust performance

📁 Project Structure

Fraud_Detection/
├─ app.py                         # Gradio app — model + preprocessing pipeline for inference
├─ requirements.txt               # Python dependencies
├─ fraud_detection.ipynb          # Full notebook: EDA → FE → training → evaluation
├─ fraud detection.pptx           # Project presentation (slides)
├─ final_data.csv                 # Cleaned dataset used for experiments
├─ encoders/                      # Label encoders & mapping artifacts (joblib)
│   ├─ le_country.joblib
│   ├─ le_browser.joblib
│   └─ ...
├─ models/                        # Saved model + scalers + feature maps (joblib)
│   ├─ best_xgb_model.joblib
│   ├─ min_max_scaler.joblib
│   ├─ device_counts_map.joblib
│   └─ ...
└─ user_input/                    # Example CSVs for demo / bulk prediction
  ├─ input1.csv
  ├─ input2.csv
  ├─ input3.csv
  └─ input4.csv

🔧 Usage

Web Interface

1. Launch the Gradio application
2. Upload a CSV file with transaction data
3. Click "Submit" to get fraud predictions
4. View probability scores for fraud/valid classification

CSV Input Format

Your input CSV should contain the following columns:

Column	Type	Description
user_id	int	Unique user identifier
signup_time	datetime	User registration timestamp
purchase_time	datetime	Transaction timestamp
purchase_value	float	Transaction amount
device_id	string	Device identifier
source	string	Traffic source (SEO/Ads/Direct)
browser	string	Browser type
sex	string	User gender (M/F)
age	int	User age
ip_address	float	IP address (numeric)
country	string	Transaction country

📈 Model Details

Training Pipeline

1. Data Preprocessing

   • Handle missing values
   • Parse datetime columns
   • Encode categorical variables

2. Feature Engineering

   • Create time-based features
   • Apply cyclical transformations
   • Generate device frequency maps

3. Model Training

   • XGBoost with hyperparameter tuning
   • Class weight balancing
   • Cross-validation for stability

4. Evaluation Metrics

   • Precision, Recall, F1-Score
   • ROC-AUC analysis
   • Confusion matrix visualization

📌 Notes on evaluation & best practices

PR-AUC over ROC-AUC: for highly imbalanced classification PR-AUC reflects precision/recall tradeoffs better than ROC.

SMOTE timing matters: creating synthetic minority samples before splitting causes target leakage — always apply synthetic oversampling after the split only on training data.

Model explainability (future): add SHAP plots or feature contribution breakdown to help stakeholders understand high-risk flags.

👨‍💻 Author

Suhaas D
Masters in Computer Science Student
Manipal School of Information Sciences

• GitHub: @suhaasd
• LinkedIn: suhaasd
• Email: [email protected]

🙏 Acknowledgments

• Manipal School of Information Sciences for academic support
• XGBoost and Gradio communities for excellent documentation
• Dataset providers for enabling research in fraud detection

📊 Results & Insights

The fraud detection system demonstrates:

• High Precision: Minimizes false positives in fraud alerts
• Scalable Architecture: Handles large-scale transaction volumes
• Real-time Processing: Sub-second prediction times
• Interpretable Results: Clear probability scores for decision-making

🔮 Future Enhancements

Add SHAP-based explanations for flagged transactions. Compare LightGBM / CatBoost to XGBoost with same pipeline. Set up CI to run notebook tests and ensure artifacts are up to date. Deploy to Hugging Face Spaces, Streamlit Cloud, or a small cloud VM for public demo.