derrick4411/-Tunisian-Fraud-Detection-Challenge

# Tax Fraud Detection Project **Detecting fraudulent tax declarations using Machine Learning** A complete data science project for identifying potential VAT/Tax fraud based on Tunisian tax authority data. --- ## Project Overview This project builds a robust fraud detection model using tax declaration data. It includes: - Exploratory Data Analysis - Feature Engineering focused on tax compliance patterns - Handling class imbalance (SMOTE) - Machine Learning model (Random Forest) - Model explainability - Tableau visualization recommendations - Production-ready prediction pipeline ## Project Structure ```bash tax-fraud-detection/ ├── notebook/ │ └── Tax_Fraud_Detection.ipynb # Main Jupyter Notebook ├── data/ │ └── raw_data.csv ├── models/ │ └── tax_fraud_detection_model.pkl ├── src/ │ ├── feature_engineering.py │ └── utils.py ├── README.md ├── requirements.txt └── Tableau_Dashboard_Guide.pdf ``` ## Key Features & Insights ### Engineered Features - `TVA_COMPLIANCE_RATIO` - `DEDUCTION_RATIO` - `ACTIVITY_DECLARATION_DIFF` - `SUSPICIOUS_HIGH_DEDUCTION` - High-risk category flags ### Fraud Indicators - Abnormally high deduction-to-turnover ratio - Large refund requests with low declared activity - Mismatches in different TVA declarations - Companies in high-risk categories with suspicious patterns --- ## Model Performance - **Algorithm**: xgboost + SMOTE - **Key Metrics**: Precision, Recall, F1-Score, ROC-AUC - **Best Use Case**: High Recall to catch most fraud cases --- ## How to Use ### Run Prediction on New Data ```python import joblib import pandas as pd model = joblib.load('models/tax_fraud_detection_model.pkl') new_data = pd.DataFrame(...) # Must have same columns prediction = model.predict(new_data) probability = model.predict_proba(new_data)[:, 1] ``` --- ## Resources & Next Steps - Improve with LightGBM + Optuna tunin - Add SHAP explainability - Deploy as API (FastAPI) - Real-time monitoring system