Selorm20/mobile-money-fraud-detection

Machine learning project for detecting mobile money fraud in Ghana # ghana-mobile-money-fraud-detection Machine learning project for detecting mobile money fraud in Ghana ## Data Access Due to file size and data management best practices, the dataset is not included in this repository. To reproduce the results: 1. Download the PaySim simulated mobile money dataset from Kaggle 2. Rename the file to `mobile_money_ghana.csv` 3. Place it in the directory: `data/raw/` The dataset simulates mobile money transactions similar to those used in Ghana and other emerging markets. ### Data Preprocessing - Engineered features combined with original transaction attributes - Categorical variables encoded using one-hot encoding - Stratified train-test split to preserve fraud class imbalance #### Logistic Regression Model concerns The baseline Logistic Regression achieves very high recall (94.5%), meaning it successfully detects most fraudulent transactions. However, due to extreme class imbalance, the model produces many false positives, resulting in low precision. The ROC-AUC score of 0.99 indicates strong discriminative power, suggesting that better decision thresholds or more advanced models can significantly improve performance.