SandraBirikorang/TB-TARL-Ghana-Fraud-Detection

Full implementation of the TB-TARL anomaly detection model, dataset simulation, baselines, SHAP, fairness analysis, and ablation studies for my PhD dissertation. # TB-TARL: Transformer-Based Anomaly Detection for Ghana’s Financial Transactions This repository contains the full code, simulation pipeline, models, and research outputs for PHD Dissertation project by Sandra Birikorang: "Designing a Novel Transformer-Based Algorithm for Anomaly Detection in Ghana’s Financial Transactions." Data Availability & Reproducibility This project uses external dataset, which is already included in the repository: API_GHA_DS2_en_csv_v2_43964.zip World Bank Ghana development indicators used to anchor the simulation. All other datasets (transactions, features, windows, labels, anomalies) are generated automatically when you run the notebook. How Data Is Generated in the Workflow The notebook builds the full synthetic Ghana mobile-money dataset through the following steps: 1️⃣Big Simulation Pipeline (first cell) reads the World Bank ZIP file, Generates 5,000 synthetic users and demographic profiles and Creates Ghana-style distributions for channels, devices, amounts, merchants, regions 2️⃣ Cell X — Cleaning Setup prepares cleaned structures and ensures stable feature computation 3️⃣ Cell 1 — Transaction + Feature Generation simulates all transactions, computes timing gaps, switching behavior, geo distance, merchant drift, builds 30-step sliding windows, injects anomalies realistically and saves raw and cleaned datasets into /content/data/ 4️⃣ Cell 2 — TB-TARL Model trains the reconstruction + attention hybrid model and computes anomaly scores for each window 5️⃣ Cell 3 — Baselines (Isolation Forest, One-Class SVM ,Autoencoder) aggregates scores to window-level 6️⃣ Cell 4 — SHAP (RQ3) trains surrogate Random Forest and computes and ranks SHAP values 7️⃣ Cell 5 — Stratified Evaluation computes fairness metrics across gender, region, income band, and urbanicity 8️⃣ Cell 6 — Ablation Study tests the importance of time encoding, sequence order, attention, and reconstruction loss Why the Large Data Files Are Not Uploaded The CSVs generated by …