Daniel-Muruthi/mpesa_fraud_detection

FraudWatch Africa is an end-to-end unsupervised fraud detection project built around Isolation Forest to detect anomalous mobile money transactions without labeled fraud data. # FraudWatch Africa — Unsupervised Mobile-Money Fraud Detection **Short:** Isolation-Forest based anomaly detection pipeline and Streamlit app for flagging suspicious mobile-money transactions in Sub-Saharan Africa. Contains a training notebook, Streamlit app documentation, presentation slides, and an example Kenya dataset. --- ## Table of Contents 1. Project Overview 2. What's included 3. Dataset (example) 4. Quickstart 5. Install & environment 6. Usage - Open the notebook - Run the Streamlit app - Use the saved model (FastAPI / script examples) 7. Model details & notes 8. Evaluation & interpretation 9. Project structure (recommended) 10. Troubleshooting 11. Contributing 12. License & contact --- ## Project overview FraudWatch Africa is an end-to-end unsupervised fraud detection project built around **Isolation Forest** to detect anomalous mobile money transactions without labeled fraud data. It includes data preprocessing, feature engineering, model training & tuning, visualizations, a Streamlit front-end, and API ideas for real-time scoring. Primary goals: - Detect anomalous transactions with minimal false positives - Provide quick, interpretable anomaly scores for analysts - Offer a deployable pipeline and a Streamlit dashboard for exploration Notebook (main analysis & training): `github.com` App documentation (uploaded): `FraudWatch Africa App Documentation.pdf` Presentation slides (uploaded): `FraudWatch_Africa_Presentation.pptx` Sample dataset (uploaded): `kenya_fraud_detection.xlsx` --- ## What's included - `index.ipynb` — main notebook: data exploration, preprocessing, model training, evaluation, visualizations. - Streamlit app docs (`FraudWatch Africa App Documentation.pdf`) and slides (`FraudWatch_Africa_Presentation.pptx`). - Example dataset: `kenya_fraud_detection.xlsx` (10,000 sample transactions). - Expected model artifacts after training (not committed unless safe): - `iso …