jacques-twizeyimana/momo-fraud-ml

Detecting Fraud in Mobile Money Transactions — A Rwandan Perspective # Mobile Money Fraud Detection — Rwanda Comparing traditional machine learning (scikit-learn) against deep learning (TensorFlow) for detecting fraudulent mobile-money transactions, motivated by the rising threat of mobile-money fraud in Rwanda. **Summative project — Introduction to Machine Learning** Author: **Jacques Twizeyimana** ### Quick links - **Report:** `report.pdf` - **Demo video:** bugufi.link - **Dataset (PaySim):** kaggle.com --- ## Problem Mobile money underpins Rwanda's digital economy, but its growth has been matched by a rise in fraud that the Rwanda Investigation Bureau (RIB) has publicly warned about. This project builds a reproducible pipeline that detects fraudulent transfers and critically compares five models, reproducing and contextualising the findings of the Carnegie Mellon University Africa study _"Mitigating Mobile Money Services Frauds in Rwanda."_ ## Dataset **PaySim** — a synthetic mobile-money transaction log (6,362,620 rows) by Lopez-Rojas _et al._ It is **not committed** to this repo (471 MB). To reproduce: 1. Download from Kaggle (link above). 2. Place `PS_20174392719_1491204439457_log.csv` in the project root. Genuine mobile-money logs are confidential, so PaySim is the standard public benchmark. Its transaction-type and balance fields let us engineer the same fraud signals the CMU-Africa team used. ## Results | Model | Precision | Recall | F1 | ROC-AUC | | ------------------- | --------- | ------ | ---------- | ---------- | | Random Forest | 1.000 | 0.999 | **0.9995** | 0.9994 | | XGBoost | 0.947 | 0.996 | 0.971 | **0.9997** | | DL Functional (TF) | 0.806 | 0.992 | 0.889 | 0.9994 | | DL Sequential (TF) | 0.603 | 0.996 | 0.751 | 0.9988 | | Logistic Regression | 0.278 | 0.889 | 0.423 | 0.9772 | Tree-based ensembles dominate this tabular problem; the neural netw …