nkaodoom/E-XGBoost-FairSHAP-MoMo-Ghana

Engineered XGBoost with Focal Loss, SHAP Explainability, and Fairness Auditing for Mobile Money Fraud Detection in Ghana # E-XGBoost-FairSHAP-MoMo-Ghana Engineered XGBoost with Focal Loss, SHAP Explainability, and Fairness Auditing for Mobile Money Fraud Detection in Ghana # E-XGBoost-FairSHAP: Mobile Money Fraud Detection in Ghana **Engineered XGBoost with Focal Loss, SHAP Explainability, and Fairness Auditing** Department of Computer Science, KNUST | June 2026 ## Overview This repository contains the code and resources for E-XGBoost-FairSHAP, a focal-loss-engineered XGBoost model with integrated SHAP explainability and fairness auditing, applied to mobile money fraud detection in the Ghanaian MTN MoMo / Telecel ecosystem. ## Repository Structure - `notebooks/01_baselines_locked.ipynb` — locked baseline experiments - `notebooks/02_e_xgboost_fairshap.ipynb` — engineered model - `results/locked_baseline_metrics.csv` — baseline results (never modified) - `results/shap_plots/` — SHAP visualisations - `results/fairness_audit.csv` — group-wise FPR results- `survey/instrument_v1.pdf` — Ghanaian MoMo user survey instrument - `data/README.md` — PaySim dataset instructions and survey protocol ## Dataset PaySim: kaggle.com Primary survey: collected under KNUST CHRPE approval (Ref: pending) ## Reproducibility All experiments use SEED = 42. Install dependencies: `pip install -r requirements.txt` ## Status - [x] Repository created - [ ] CHRPE ethics approval received - [ ] Baseline experiments complete - [ ] E-XGBoost-FairSHAP experiments complete