Omar-Elhadidi/tax-fraud-detection-tunisia

Machine learning pipeline to detect tax fraud using Tunisia Ministry of Finance data # 🧾 Tunisia Tax Fraud Detection > Detecting tax fraud in Tunisia using supervised machine learning on real Ministry of Finance data — achieving **6th place out of 286 teams** worldwide on the Zindi leaderboard. --- ## 🏆 Achievement Ranked **#6 globally out of 286 participants** in the Tunisian Fraud Detection Challenge on Zindi. --- ## 📌 Objective - Improve tax fraud detection accuracy using ML - Handle missing data and outliers effectively - Create advanced engineered features for better performance - Reach the lowest possible **Root Mean Squared Error (RMSE)** --- ## 🔗 Data Source This project is based on the Tunisian Fraud Detection Challenge on Zindi which provides the data and problem definition used here. --- ## 📁 Project Structure | File | Description | |------|-------------| | `tunisia_tax_fraud_model.ipynb` | Complete pipeline from loading data to model evaluation | | `Report.pdf` | Final project summary and presentation | | `Submission_Enhanced.csv` | Final prediction submission file | | `README.md` | Project documentation | --- ## 🧪 Model & Techniques - **Model**: LightGBM (fast gradient boosting) - **Validation**: 7-Fold Cross-Validation - **Encoding**: CatBoostEncoder + Target Encoding - **Feature Engineering**: - Ratio and log-ratio features - Categorical interactions - Missing value indicators - Aggregated statistics --- ## 📊 Dataset Overview | File | Description | |------|-------------| | `Train.csv` | 15,000 rows with features and a `target` column | | `Test.csv` | 5,000 rows with features only | | `submission.csv` | Sample format for submission | --- ## 🔍 Evaluation Metric - **Root Mean Squared Error (RMSE)** - Lower RMSE = better model performance --- ## ✅ Results | Metric | Value | |--------|-------| | Baseline RMSE | 7.0856 | | Final RMSE (OOF) | **5.377** | | Improvement | ↓ ~24% | | **Leaderboard Rank** | **🥇 6th / 286** | --- ## 📉 Visualizations The notebook includes: - Target distribution plots - Missing val …