anwarica134/tunisian-fraud-detection

# Tunisian Fraud Detection — Zindi Competition > **Leaderboard rank: #203** | Stacking ensemble of 5 models | Final RMSE: 6.73 ## Overview This project was built for the **Zindi Africa Tunisian Fraud Detection Challenge** — a regression competition to predict fraudulent transaction values from anonymised financial data. The solution uses a **stacking ensemble** of five models combined via a Ridge meta-learner, with log-transformation applied to the target variable to handle severe skewness. --- ## Results | Model | RMSE | |---|---| | LightGBM | 6.7819 | | XGBoost | 6.7863 | | CatBoost | 6.8481 | | ExtraTrees | 8.1691 | | **Stacked Ensemble (final)** | **6.7318** | **Zindi leaderboard rank: #203** --- ## Approach ### Target Engineering The fraud value target was heavily right-skewed. Applied `np.log1p` transformation before training and `np.expm1` to reverse predictions — this stabilised training and improved generalisation. ### Feature Engineering - Row-level statistical features: sum, mean, standard deviation, and zero-count across all numerical columns - Variance threshold filtering to remove low-information features - Median imputation for missing values ### Modelling - Trained 5 base models: LightGBM, XGBoost, CatBoost, ExtraTrees, Ridge - Used **5-fold cross-validation** with early stopping on tree-based models - Out-of-fold predictions from base models fed into a **Ridge meta-learner** as second-level features - Final prediction is the meta-learner output inverse-transformed back to original scale ### Visualisations - Target distribution before and after log-transformation - Correlation heatmap of engineered features - Feature importance ranking from LightGBM --- ## Tech Stack | Category | Libraries | |---|---| | Data processing | pandas, NumPy, scikit-learn | | Models | LightGBM, XGBoost, CatBoost, scikit-learn (ExtraTrees, Ridge) | | Visualisation | Matplotlib, Seaborn | | Environment | Google Colab | --- ## How to Run 1. Open the `.py` f …